diff options
Diffstat (limited to 'src/gallium/drivers/llvmpipe')
88 files changed, 20920 insertions, 0 deletions
diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile new file mode 100644 index 00000000000..5ac09de79ea --- /dev/null +++ b/src/gallium/drivers/llvmpipe/Makefile @@ -0,0 +1,54 @@ +TOP = ../../../.. +include $(TOP)/configs/current + +LIBNAME = llvmpipe + +CFLAGS += -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS + +C_SOURCES = \ + lp_bld_alpha.c \ + lp_bld_arit.c \ + lp_bld_blend_aos.c \ + lp_bld_blend_logicop.c \ + lp_bld_blend_soa.c \ + lp_bld_const.c \ + lp_bld_conv.c \ + lp_bld_debug.c \ + lp_bld_depth.c \ + lp_bld_flow.c \ + lp_bld_format_aos.c \ + lp_bld_interp.c \ + lp_bld_intr.c \ + lp_bld_logic.c \ + lp_bld_swizzle.c \ + lp_bld_struct.c \ + lp_bld_tgsi_soa.c \ + lp_bld_type.c \ + lp_buffer.c \ + lp_clear.c \ + lp_context.c \ + lp_draw_arrays.c \ + lp_flush.c \ + lp_jit.c \ + lp_prim_setup.c \ + lp_prim_vbuf.c \ + lp_setup.c \ + lp_query.c \ + lp_screen.c \ + lp_state_blend.c \ + lp_state_clip.c \ + lp_state_derived.c \ + lp_state_fs.c \ + lp_state_rasterizer.c \ + lp_state_sampler.c \ + lp_state_surface.c \ + lp_state_vertex.c \ + lp_state_vs.c \ + lp_surface.c \ + lp_tex_cache.c \ + lp_tex_sample.c \ + lp_texture.c \ + lp_tile_cache.c \ + lp_tile_soa.c + +include ../../Makefile.template diff --git a/src/gallium/drivers/llvmpipe/README b/src/gallium/drivers/llvmpipe/README new file mode 100644 index 00000000000..498d21dea6c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/README @@ -0,0 +1,142 @@ +LLVMPIPE -- a fork of softpipe that employs LLVM for code generation. + + +Status +====== + +Done so far is: + + - the whole fragment pipeline is code generated in a single function + + - depth testing + + - fragment shader TGSI translation + - same level of support as the TGSI SSE2 exec machine, with the exception + we don't fallback to TGSI interpretation when an unsupported opcode is + found, but just ignore it + - texture sampling via an intrinsic call + - done in SoA layout + - input interpolation also code generated + + - alpha testing + + - blend (including logic ops) + - both in SoA and AoS layouts, but only the former used for now + + - code is generic + - intermediates can be vectors of floats, ubytes, fixed point, etc, and of + any width and length + - not all operations are implemented for these types yet though + +Most mesa/progs/demos/* work. Speed is on par with Keith's softpipe-opt branch, +which includes hand written fast implementations for common cases. + +To do (probably by this order): + + - code generate stipple and stencil testing + + - code generate texture sampling + + - translate TGSI control flow instructions, and all other remaining opcodes + + - code generate the triangle setup and rasterization + + +Requirements +============ + + - Linux + + - udis86, http://udis86.sourceforge.net/ . Use my repository, which decodes + opcodes not yet supported by upstream. + + git clone git://people.freedesktop.org/~jrfonseca/udis86 + cd udis86 + ./configure --with-pic + make + sudo make install + + - LLVM 2.5. On Debian based distributions do: + + aptitude install llvm-dev + + There is a typo in one of the llvm-dev 2.5 headers, that causes compilation + errors in the debug build: + + --- /usr/include/llvm-c/Core.h.orig 2009-08-10 15:38:54.000000000 +0100 + +++ /usr/include/llvm-c/Core.h 2009-08-10 15:38:25.000000000 +0100 + @@ -831,7 +831,7 @@ + template<typename T> + inline T **unwrap(LLVMValueRef *Vals, unsigned Length) { + #if DEBUG + - for (LLVMValueRef *I = Vals, E = Vals + Length; I != E; ++I) + + for (LLVMValueRef *I = Vals, *E = Vals + Length; I != E; ++I) + cast<T>(*I); + #endif + return reinterpret_cast<T**>(Vals); + + - A x86 or amd64 processor with support for sse2, sse3, and sse4.1 SIMD + instructions. This is necessary because we emit several SSE intrinsics for + convenience. See /proc/cpuinfo to know what your CPU supports. + + - scons + + +Building +======== + +To build everything invoke scons as: + + scons debug=yes statetrackers=mesa drivers=llvmpipe winsys=xlib dri=false -k + +Alternatively, you can build it with GNU make, if you prefer, by invoking it as + + make linux-llvm + +but the rest of these instructions assume scons is used. + + +Using +===== + +Building will create a drop-in alternative for libGL.so. To use it set the +environment variables: + + export LD_LIBRARY_PATH=$PWD/build/linux-x86_64-debug/lib:$LD_LIBRARY_PATH + +or + + export LD_LIBRARY_PATH=$PWD/build/linux-x86-debug/lib:$LD_LIBRARY_PATH + + +Unit testing +============ + +Building will also create several unit tests in +build/linux-???-debug/gallium/drivers/llvmpipe: + + - lp_test_blend: blending + - lp_test_conv: SIMD vector conversion + - lp_test_format: pixel unpacking/packing + +Some of this tests can output results and benchmarks to a tab-seperated-file +for posterior analysis, e.g.: + + build/linux-x86_64-debug/gallium/drivers/llvmpipe/lp_test_blend -o blend.tsv + + +Development Notes +================= + +- When looking to this code by the first time start in lp_state_fs.c, and + then skim through the lp_bld_* functions called in there, and the comments + at the top of the lp_bld_*.c functions. + +- All lp_bld_*.[ch] are isolated from the rest of the driver, and could/may be + put in a standalone Gallium state -> LLVM IR translation module. + +- We use LLVM-C bindings for now. They are not documented, but follow the C++ + interfaces very closely, and appear to be complete enough for code + generation. See + http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html + for a standalone example. diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript new file mode 100644 index 00000000000..5c29bdac56e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/SConscript @@ -0,0 +1,81 @@ +Import('*') + +env = env.Clone() + +env.Tool('llvm') +if 'LLVM_VERSION' not in env: + print 'warning: LLVM not found: not building llvmpipe' + Return() + +env.Tool('udis86') + +llvmpipe = env.ConvenienceLibrary( + target = 'llvmpipe', + source = [ + 'lp_bld_alpha.c', + 'lp_bld_arit.c', + 'lp_bld_blend_aos.c', + 'lp_bld_blend_logicop.c', + 'lp_bld_blend_soa.c', + 'lp_bld_const.c', + 'lp_bld_conv.c', + 'lp_bld_debug.c', + 'lp_bld_depth.c', + 'lp_bld_flow.c', + 'lp_bld_format_aos.c', + 'lp_bld_interp.c', + 'lp_bld_intr.c', + 'lp_bld_struct.c', + 'lp_bld_logic.c', + 'lp_bld_swizzle.c', + 'lp_bld_tgsi_soa.c', + 'lp_bld_type.c', + 'lp_buffer.c', + 'lp_clear.c', + 'lp_context.c', + 'lp_draw_arrays.c', + 'lp_flush.c', + 'lp_jit.c', + 'lp_prim_setup.c', + 'lp_prim_vbuf.c', + 'lp_setup.c', + 'lp_query.c', + 'lp_screen.c', + 'lp_state_blend.c', + 'lp_state_clip.c', + 'lp_state_derived.c', + 'lp_state_fs.c', + 'lp_state_rasterizer.c', + 'lp_state_sampler.c', + 'lp_state_surface.c', + 'lp_state_vertex.c', + 'lp_state_vs.c', + 'lp_surface.c', + 'lp_tex_cache.c', + 'lp_tex_sample.c', + 'lp_texture.c', + 'lp_tile_cache.c', + 'lp_tile_soa.c', + ]) + + +env = env.Clone() + +env.Prepend(LIBS = [llvmpipe] + auxiliaries) + +env.Program( + target = 'lp_test_format', + source = ['lp_test_format.c'], +) + +env.Program( + target = 'lp_test_blend', + source = ['lp_test_blend.c', 'lp_test_main.c'], +) + +env.Program( + target = 'lp_test_conv', + source = ['lp_test_conv.c', 'lp_test_main.c'], +) + +Export('llvmpipe') diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c new file mode 100644 index 00000000000..49c2f911af7 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c @@ -0,0 +1,64 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Alpha testing to LLVM IR translation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_state.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_arit.h" +#include "lp_bld_logic.h" +#include "lp_bld_flow.h" +#include "lp_bld_debug.h" +#include "lp_bld_alpha.h" + + +void +lp_build_alpha_test(LLVMBuilderRef builder, + const struct pipe_alpha_state *state, + union lp_type type, + struct lp_build_mask_context *mask, + LLVMValueRef alpha, + LLVMValueRef ref) +{ + struct lp_build_context bld; + + lp_build_context_init(&bld, builder, type); + + if(state->enabled) { + LLVMValueRef test = lp_build_cmp(&bld, state->func, alpha, ref); + + lp_build_name(test, "alpha_mask"); + + lp_build_mask_update(mask, test); + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h new file mode 100644 index 00000000000..9dbcdb4daab --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h @@ -0,0 +1,54 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Alpha testing to LLVM IR translation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_ALPHA_H +#define LP_BLD_ALPHA_H + + +#include <llvm-c/Core.h> + +struct pipe_alpha_state; +union lp_type; +struct lp_build_mask_context; + + +void +lp_build_alpha_test(LLVMBuilderRef builder, + const struct pipe_alpha_state *state, + union lp_type type, + struct lp_build_mask_context *mask, + LLVMValueRef alpha, + LLVMValueRef ref); + + +#endif /* !LP_BLD_ALPHA_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c new file mode 100644 index 00000000000..09a57ff33d5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c @@ -0,0 +1,963 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helper + * + * LLVM IR doesn't support all basic arithmetic operations we care about (most + * notably min/max and saturated operations), and it is often necessary to + * resort machine-specific intrinsics directly. The functions here hide all + * these implementation details from the other modules. + * + * We also do simple expressions simplification here. Reasons are: + * - it is very easy given we have all necessary information readily available + * - LLVM optimization passes fail to simplify several vector expressions + * - We often know value constraints which the optimization passes have no way + * of knowing, such as when source arguments are known to be in [0, 1] range. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_memory.h" +#include "util/u_debug.h" +#include "util/u_string.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_intr.h" +#include "lp_bld_logic.h" +#include "lp_bld_arit.h" + + +/** + * Generate min(a, b) + * No checks for special case values of a or b = 1 or 0 are done. + */ +static LLVMValueRef +lp_build_min_simple(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + const char *intrinsic = NULL; + LLVMValueRef cond; + + /* TODO: optimize the constant case */ + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width * type.length == 128) { + if(type.floating) { + if(type.width == 32) + intrinsic = "llvm.x86.sse.min.ps"; + if(type.width == 64) + intrinsic = "llvm.x86.sse2.min.pd"; + } + else { + if(type.width == 8 && !type.sign) + intrinsic = "llvm.x86.sse2.pminu.b"; + if(type.width == 8 && type.sign) + intrinsic = "llvm.x86.sse41.pminsb"; + if(type.width == 16 && !type.sign) + intrinsic = "llvm.x86.sse41.pminuw"; + if(type.width == 16 && type.sign) + intrinsic = "llvm.x86.sse2.pmins.w"; + if(type.width == 32 && !type.sign) + intrinsic = "llvm.x86.sse41.pminud"; + if(type.width == 32 && type.sign) + intrinsic = "llvm.x86.sse41.pminsd"; + } + } +#endif + + if(intrinsic) + return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); + + cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b); + return lp_build_select(bld, cond, a, b); +} + + +/** + * Generate max(a, b) + * No checks for special case values of a or b = 1 or 0 are done. + */ +static LLVMValueRef +lp_build_max_simple(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + const char *intrinsic = NULL; + LLVMValueRef cond; + + /* TODO: optimize the constant case */ + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width * type.length == 128) { + if(type.floating) { + if(type.width == 32) + intrinsic = "llvm.x86.sse.max.ps"; + if(type.width == 64) + intrinsic = "llvm.x86.sse2.max.pd"; + } + else { + if(type.width == 8 && !type.sign) + intrinsic = "llvm.x86.sse2.pmaxu.b"; + if(type.width == 8 && type.sign) + intrinsic = "llvm.x86.sse41.pmaxsb"; + if(type.width == 16 && !type.sign) + intrinsic = "llvm.x86.sse41.pmaxuw"; + if(type.width == 16 && type.sign) + intrinsic = "llvm.x86.sse2.pmaxs.w"; + if(type.width == 32 && !type.sign) + intrinsic = "llvm.x86.sse41.pmaxud"; + if(type.width == 32 && type.sign) + intrinsic = "llvm.x86.sse41.pmaxsd"; + } + } +#endif + + if(intrinsic) + return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); + + cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b); + return lp_build_select(bld, cond, a, b); +} + + +/** + * Generate 1 - a, or ~a depending on bld->type. + */ +LLVMValueRef +lp_build_comp(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + + if(a == bld->one) + return bld->zero; + if(a == bld->zero) + return bld->one; + + if(type.norm && !type.floating && !type.fixed && !type.sign) { + if(LLVMIsConstant(a)) + return LLVMConstNot(a); + else + return LLVMBuildNot(bld->builder, a, ""); + } + + if(LLVMIsConstant(a)) + return LLVMConstSub(bld->one, a); + else + return LLVMBuildSub(bld->builder, bld->one, a, ""); +} + + +/** + * Generate a + b + */ +LLVMValueRef +lp_build_add(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + LLVMValueRef res; + + if(a == bld->zero) + return b; + if(b == bld->zero) + return a; + if(a == bld->undef || b == bld->undef) + return bld->undef; + + if(bld->type.norm) { + const char *intrinsic = NULL; + + if(a == bld->one || b == bld->one) + return bld->one; + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width * type.length == 128 && + !type.floating && !type.fixed) { + if(type.width == 8) + intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b"; + if(type.width == 16) + intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w"; + } +#endif + + if(intrinsic) + return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); + } + + if(LLVMIsConstant(a) && LLVMIsConstant(b)) + res = LLVMConstAdd(a, b); + else + res = LLVMBuildAdd(bld->builder, a, b, ""); + + /* clamp to ceiling of 1.0 */ + if(bld->type.norm && (bld->type.floating || bld->type.fixed)) + res = lp_build_min_simple(bld, res, bld->one); + + /* XXX clamp to floor of -1 or 0??? */ + + return res; +} + + +/** + * Generate a - b + */ +LLVMValueRef +lp_build_sub(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + LLVMValueRef res; + + if(b == bld->zero) + return a; + if(a == bld->undef || b == bld->undef) + return bld->undef; + if(a == b) + return bld->zero; + + if(bld->type.norm) { + const char *intrinsic = NULL; + + if(b == bld->one) + return bld->zero; + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width * type.length == 128 && + !type.floating && !type.fixed) { + if(type.width == 8) + intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b"; + if(type.width == 16) + intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w"; + } +#endif + + if(intrinsic) + return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b); + } + + if(LLVMIsConstant(a) && LLVMIsConstant(b)) + res = LLVMConstSub(a, b); + else + res = LLVMBuildSub(bld->builder, a, b, ""); + + if(bld->type.norm && (bld->type.floating || bld->type.fixed)) + res = lp_build_max_simple(bld, res, bld->zero); + + return res; +} + + +/** + * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions. + */ +static LLVMValueRef +lp_build_unpack_shuffle(unsigned n, unsigned lo_hi) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i, j; + + assert(n <= LP_MAX_VECTOR_LENGTH); + assert(lo_hi < 2); + + for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) { + elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0); + elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0); + } + + return LLVMConstVector(elems, n); +} + + +/** + * Build constant int vector of width 'n' and value 'c'. + */ +static LLVMValueRef +lp_build_const_vec(LLVMTypeRef type, unsigned n, long long c) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(n <= LP_MAX_VECTOR_LENGTH); + + for(i = 0; i < n; ++i) + elems[i] = LLVMConstInt(type, c, 0); + + return LLVMConstVector(elems, n); +} + + +/** + * Normalized 8bit multiplication. + * + * - alpha plus one + * + * makes the following approximation to the division (Sree) + * + * a*b/255 ~= (a*(b + 1)) >> 256 + * + * which is the fastest method that satisfies the following OpenGL criteria + * + * 0*0 = 0 and 255*255 = 255 + * + * - geometric series + * + * takes the geometric series approximation to the division + * + * t/255 = (t >> 8) + (t >> 16) + (t >> 24) .. + * + * in this case just the first two terms to fit in 16bit arithmetic + * + * t/255 ~= (t + (t >> 8)) >> 8 + * + * note that just by itself it doesn't satisfies the OpenGL criteria, as + * 255*255 = 254, so the special case b = 255 must be accounted or roundoff + * must be used + * + * - geometric series plus rounding + * + * when using a geometric series division instead of truncating the result + * use roundoff in the approximation (Jim Blinn) + * + * t/255 ~= (t + (t >> 8) + 0x80) >> 8 + * + * achieving the exact results + * + * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995, + * ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf + * @sa Michael Herf, The "double blend trick", May 2000, + * http://www.stereopsis.com/doubleblend.html + */ +static LLVMValueRef +lp_build_mul_u8n(LLVMBuilderRef builder, + LLVMValueRef a, LLVMValueRef b) +{ + static LLVMValueRef c01 = NULL; + static LLVMValueRef c08 = NULL; + static LLVMValueRef c80 = NULL; + LLVMValueRef ab; + + if(!c01) c01 = lp_build_const_vec(LLVMInt16Type(), 8, 0x01); + if(!c08) c08 = lp_build_const_vec(LLVMInt16Type(), 8, 0x08); + if(!c80) c80 = lp_build_const_vec(LLVMInt16Type(), 8, 0x80); + +#if 0 + + /* a*b/255 ~= (a*(b + 1)) >> 256 */ + b = LLVMBuildAdd(builder, b, c01, ""); + ab = LLVMBuildMul(builder, a, b, ""); + +#else + + /* t/255 ~= (t + (t >> 8) + 0x80) >> 8 */ + ab = LLVMBuildMul(builder, a, b, ""); + ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c08, ""), ""); + ab = LLVMBuildAdd(builder, ab, c80, ""); + +#endif + + ab = LLVMBuildLShr(builder, ab, c08, ""); + + return ab; +} + + +/** + * Generate a * b + */ +LLVMValueRef +lp_build_mul(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + + if(a == bld->zero) + return bld->zero; + if(a == bld->one) + return b; + if(b == bld->zero) + return bld->zero; + if(b == bld->one) + return a; + if(a == bld->undef || b == bld->undef) + return bld->undef; + + if(!type.floating && !type.fixed && type.norm) { +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width == 8 && type.length == 16) { + LLVMTypeRef i16x8 = LLVMVectorType(LLVMInt16Type(), 8); + LLVMTypeRef i8x16 = LLVMVectorType(LLVMInt8Type(), 16); + static LLVMValueRef ml = NULL; + static LLVMValueRef mh = NULL; + LLVMValueRef al, ah, bl, bh; + LLVMValueRef abl, abh; + LLVMValueRef ab; + + if(!ml) ml = lp_build_unpack_shuffle(16, 0); + if(!mh) mh = lp_build_unpack_shuffle(16, 1); + + /* PUNPCKLBW, PUNPCKHBW */ + al = LLVMBuildShuffleVector(bld->builder, a, bld->zero, ml, ""); + bl = LLVMBuildShuffleVector(bld->builder, b, bld->zero, ml, ""); + ah = LLVMBuildShuffleVector(bld->builder, a, bld->zero, mh, ""); + bh = LLVMBuildShuffleVector(bld->builder, b, bld->zero, mh, ""); + + /* NOP */ + al = LLVMBuildBitCast(bld->builder, al, i16x8, ""); + bl = LLVMBuildBitCast(bld->builder, bl, i16x8, ""); + ah = LLVMBuildBitCast(bld->builder, ah, i16x8, ""); + bh = LLVMBuildBitCast(bld->builder, bh, i16x8, ""); + + /* PMULLW, PSRLW, PADDW */ + abl = lp_build_mul_u8n(bld->builder, al, bl); + abh = lp_build_mul_u8n(bld->builder, ah, bh); + + /* PACKUSWB */ + ab = lp_build_intrinsic_binary(bld->builder, "llvm.x86.sse2.packuswb.128" , i16x8, abl, abh); + + /* NOP */ + ab = LLVMBuildBitCast(bld->builder, ab, i8x16, ""); + + return ab; + } +#endif + + /* FIXME */ + assert(0); + } + + if(LLVMIsConstant(a) && LLVMIsConstant(b)) + return LLVMConstMul(a, b); + + return LLVMBuildMul(bld->builder, a, b, ""); +} + + +/** + * Generate a / b + */ +LLVMValueRef +lp_build_div(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + + if(a == bld->zero) + return bld->zero; + if(a == bld->one) + return lp_build_rcp(bld, b); + if(b == bld->zero) + return bld->undef; + if(b == bld->one) + return a; + if(a == bld->undef || b == bld->undef) + return bld->undef; + + if(LLVMIsConstant(a) && LLVMIsConstant(b)) + return LLVMConstFDiv(a, b); + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width == 32 && type.length == 4) + return lp_build_mul(bld, a, lp_build_rcp(bld, b)); +#endif + + return LLVMBuildFDiv(bld->builder, a, b, ""); +} + + +/** + * Generate min(a, b) + * Do checks for special cases. + */ +LLVMValueRef +lp_build_min(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + if(a == bld->undef || b == bld->undef) + return bld->undef; + + if(a == b) + return a; + + if(bld->type.norm) { + if(a == bld->zero || b == bld->zero) + return bld->zero; + if(a == bld->one) + return b; + if(b == bld->one) + return a; + } + + return lp_build_min_simple(bld, a, b); +} + + +/** + * Generate max(a, b) + * Do checks for special cases. + */ +LLVMValueRef +lp_build_max(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b) +{ + if(a == bld->undef || b == bld->undef) + return bld->undef; + + if(a == b) + return a; + + if(bld->type.norm) { + if(a == bld->one || b == bld->one) + return bld->one; + if(a == bld->zero) + return b; + if(b == bld->zero) + return a; + } + + return lp_build_max_simple(bld, a, b); +} + + +/** + * Generate abs(a) + */ +LLVMValueRef +lp_build_abs(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + + if(!type.sign) + return a; + + /* XXX: is this really necessary? */ +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(!type.floating && type.width*type.length == 128) { + LLVMTypeRef vec_type = lp_build_vec_type(type); + if(type.width == 8) + return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a); + if(type.width == 16) + return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a); + if(type.width == 32) + return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a); + } +#endif + + return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, "")); +} + + +LLVMValueRef +lp_build_sqrt(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + char intrinsic[32]; + + /* TODO: optimize the constant case */ + /* TODO: optimize the constant case */ + + assert(type.floating); + util_snprintf(intrinsic, sizeof intrinsic, "llvm.sqrt.v%uf%u", type.length, type.width); + + return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a); +} + + +LLVMValueRef +lp_build_rcp(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + + if(a == bld->zero) + return bld->undef; + if(a == bld->one) + return bld->one; + if(a == bld->undef) + return bld->undef; + + assert(type.floating); + + if(LLVMIsConstant(a)) + return LLVMConstFDiv(bld->one, a); + + /* XXX: is this really necessary? */ +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width == 32 && type.length == 4) + return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a); +#endif + + return LLVMBuildFDiv(bld->builder, bld->one, a, ""); +} + + +/** + * Generate 1/sqrt(a) + */ +LLVMValueRef +lp_build_rsqrt(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + + assert(type.floating); + + /* XXX: is this really necessary? */ +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width == 32 && type.length == 4) + return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type), a); +#endif + + return lp_build_rcp(bld, lp_build_sqrt(bld, a)); +} + + +/** + * Generate cos(a) + */ +LLVMValueRef +lp_build_cos(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + char intrinsic[32]; + + /* TODO: optimize the constant case */ + + assert(type.floating); + util_snprintf(intrinsic, sizeof intrinsic, "llvm.cos.v%uf%u", type.length, type.width); + + return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a); +} + + +/** + * Generate sin(a) + */ +LLVMValueRef +lp_build_sin(struct lp_build_context *bld, + LLVMValueRef a) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + char intrinsic[32]; + + /* TODO: optimize the constant case */ + + assert(type.floating); + util_snprintf(intrinsic, sizeof intrinsic, "llvm.sin.v%uf%u", type.length, type.width); + + return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a); +} + + +/** + * Generate pow(x, y) + */ +LLVMValueRef +lp_build_pow(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef y) +{ + /* TODO: optimize the constant case */ + if(LLVMIsConstant(x) && LLVMIsConstant(y)) + debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + + return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y)); +} + + +/** + * Generate exp(x) + */ +LLVMValueRef +lp_build_exp(struct lp_build_context *bld, + LLVMValueRef x) +{ + /* log2(e) = 1/log(2) */ + LLVMValueRef log2e = lp_build_const_scalar(bld->type, 1.4426950408889634); + + return lp_build_mul(bld, log2e, lp_build_exp2(bld, x)); +} + + +/** + * Generate log(x) + */ +LLVMValueRef +lp_build_log(struct lp_build_context *bld, + LLVMValueRef x) +{ + /* log(2) */ + LLVMValueRef log2 = lp_build_const_scalar(bld->type, 1.4426950408889634); + + return lp_build_mul(bld, log2, lp_build_exp2(bld, x)); +} + + +#define EXP_POLY_DEGREE 3 +#define LOG_POLY_DEGREE 5 + + +/** + * Generate polynomial. + * Ex: x^2 * coeffs[0] + x * coeffs[1] + coeffs[2]. + */ +static LLVMValueRef +lp_build_polynomial(struct lp_build_context *bld, + LLVMValueRef x, + const double *coeffs, + unsigned num_coeffs) +{ + const union lp_type type = bld->type; + LLVMValueRef res = NULL; + unsigned i; + + /* TODO: optimize the constant case */ + if(LLVMIsConstant(x)) + debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + + for (i = num_coeffs; i--; ) { + LLVMValueRef coeff = lp_build_const_scalar(type, coeffs[i]); + if(res) + res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res)); + else + res = coeff; + } + + if(res) + return res; + else + return bld->undef; +} + + +/** + * Minimax polynomial fit of 2**x, in range [-0.5, 0.5[ + */ +const double lp_build_exp2_polynomial[] = { +#if EXP_POLY_DEGREE == 5 + 9.9999994e-1, 6.9315308e-1, 2.4015361e-1, 5.5826318e-2, 8.9893397e-3, 1.8775767e-3 +#elif EXP_POLY_DEGREE == 4 + 1.0000026, 6.9300383e-1, 2.4144275e-1, 5.2011464e-2, 1.3534167e-2 +#elif EXP_POLY_DEGREE == 3 + 9.9992520e-1, 6.9583356e-1, 2.2606716e-1, 7.8024521e-2 +#elif EXP_POLY_DEGREE == 2 + 1.0017247, 6.5763628e-1, 3.3718944e-1 +#else +#error +#endif +}; + + +void +lp_build_exp2_approx(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef *p_exp2_int_part, + LLVMValueRef *p_frac_part, + LLVMValueRef *p_exp2) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef ipart = NULL; + LLVMValueRef fpart = NULL; + LLVMValueRef expipart = NULL; + LLVMValueRef expfpart = NULL; + LLVMValueRef res = NULL; + + if(p_exp2_int_part || p_frac_part || p_exp2) { + /* TODO: optimize the constant case */ + if(LLVMIsConstant(x)) + debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + + assert(type.floating && type.width == 32); + + x = lp_build_min(bld, x, lp_build_const_scalar(type, 129.0)); + x = lp_build_max(bld, x, lp_build_const_scalar(type, -126.99999)); + + /* ipart = int(x - 0.5) */ + ipart = LLVMBuildSub(bld->builder, x, lp_build_const_scalar(type, 0.5f), ""); + ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, ""); + + /* fpart = x - ipart */ + fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, ""); + fpart = LLVMBuildSub(bld->builder, x, fpart, ""); + } + + if(p_exp2_int_part || p_exp2) { + /* expipart = (float) (1 << ipart) */ + expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_int_const_scalar(type, 127), ""); + expipart = LLVMBuildShl(bld->builder, expipart, lp_build_int_const_scalar(type, 23), ""); + expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, ""); + } + + if(p_exp2) { + expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial, + Elements(lp_build_exp2_polynomial)); + + res = LLVMBuildMul(bld->builder, expipart, expfpart, ""); + } + + if(p_exp2_int_part) + *p_exp2_int_part = expipart; + + if(p_frac_part) + *p_frac_part = fpart; + + if(p_exp2) + *p_exp2 = res; +} + + +LLVMValueRef +lp_build_exp2(struct lp_build_context *bld, + LLVMValueRef x) +{ + LLVMValueRef res; + lp_build_exp2_approx(bld, x, NULL, NULL, &res); + return res; +} + + +/** + * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[ + * These coefficients can be generate with + * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html + */ +const double lp_build_log2_polynomial[] = { +#if LOG_POLY_DEGREE == 6 + 3.11578814719469302614, -3.32419399085241980044, 2.59883907202499966007, -1.23152682416275988241, 0.318212422185251071475, -0.0344359067839062357313 +#elif LOG_POLY_DEGREE == 5 + 2.8882704548164776201, -2.52074962577807006663, 1.48116647521213171641, -0.465725644288844778798, 0.0596515482674574969533 +#elif LOG_POLY_DEGREE == 4 + 2.61761038894603480148, -1.75647175389045657003, 0.688243882994381274313, -0.107254423828329604454 +#elif LOG_POLY_DEGREE == 3 + 2.28330284476918490682, -1.04913055217340124191, 0.204446009836232697516 +#else +#error +#endif +}; + + +/** + * See http://www.devmaster.net/forums/showthread.php?p=43580 + */ +void +lp_build_log2_approx(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef *p_exp, + LLVMValueRef *p_floor_log2, + LLVMValueRef *p_log2) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + + LLVMValueRef expmask = lp_build_int_const_scalar(type, 0x7f800000); + LLVMValueRef mantmask = lp_build_int_const_scalar(type, 0x007fffff); + LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type); + + LLVMValueRef i = NULL; + LLVMValueRef exp = NULL; + LLVMValueRef mant = NULL; + LLVMValueRef logexp = NULL; + LLVMValueRef logmant = NULL; + LLVMValueRef res = NULL; + + if(p_exp || p_floor_log2 || p_log2) { + /* TODO: optimize the constant case */ + if(LLVMIsConstant(x)) + debug_printf("%s: inefficient/imprecise constant arithmetic\n"); + + assert(type.floating && type.width == 32); + + i = LLVMBuildBitCast(bld->builder, x, int_vec_type, ""); + + /* exp = (float) exponent(x) */ + exp = LLVMBuildAnd(bld->builder, i, expmask, ""); + } + + if(p_floor_log2 || p_log2) { + logexp = LLVMBuildLShr(bld->builder, exp, lp_build_int_const_scalar(type, 23), ""); + logexp = LLVMBuildSub(bld->builder, logexp, lp_build_int_const_scalar(type, 127), ""); + logexp = LLVMBuildSIToFP(bld->builder, logexp, vec_type, ""); + } + + if(p_log2) { + /* mant = (float) mantissa(x) */ + mant = LLVMBuildAnd(bld->builder, i, mantmask, ""); + mant = LLVMBuildOr(bld->builder, mant, one, ""); + mant = LLVMBuildSIToFP(bld->builder, mant, vec_type, ""); + + logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial, + Elements(lp_build_log2_polynomial)); + + /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/ + logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildMul(bld->builder, mant, bld->one, ""), ""); + + res = LLVMBuildAdd(bld->builder, logmant, logexp, ""); + } + + if(p_exp) + *p_exp = exp; + + if(p_floor_log2) + *p_floor_log2 = logexp; + + if(p_log2) + *p_log2 = res; +} + + +LLVMValueRef +lp_build_log2(struct lp_build_context *bld, + LLVMValueRef x) +{ + LLVMValueRef res; + lp_build_log2_approx(bld, x, NULL, NULL, &res); + return res; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h new file mode 100644 index 00000000000..fc8cb25966e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h @@ -0,0 +1,143 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper arithmetic functions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_ARIT_H +#define LP_BLD_ARIT_H + + +#include <llvm-c/Core.h> + + +union lp_type type; +struct lp_build_context; + + +/** + * Complement, i.e., 1 - a. + */ +LLVMValueRef +lp_build_comp(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_add(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_sub(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_mul(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_div(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_min(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_max(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_abs(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_sqrt(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_rcp(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_rsqrt(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_cos(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_sin(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_pow(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_exp(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_log(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_exp2(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_log2(struct lp_build_context *bld, + LLVMValueRef a); + +void +lp_build_exp2_approx(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef *p_exp2_int_part, + LLVMValueRef *p_frac_part, + LLVMValueRef *p_exp2); + +void +lp_build_log2_approx(struct lp_build_context *bld, + LLVMValueRef x, + LLVMValueRef *p_exp, + LLVMValueRef *p_floor_log2, + LLVMValueRef *p_log2); + +#endif /* !LP_BLD_ARIT_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend.h b/src/gallium/drivers/llvmpipe/lp_bld_blend.h new file mode 100644 index 00000000000..d19e18846c2 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend.h @@ -0,0 +1,107 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_BLD_BLEND_H +#define LP_BLD_BLEND_H + + +/** + * @file + * LLVM IR building helpers interfaces. + * + * We use LLVM-C bindings for now. They are not documented, but follow the C++ + * interfaces very closely, and appear to be complete enough for code + * genration. See + * http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html + * for a standalone example. + */ + +#include <llvm-c/Core.h> + +#include "pipe/p_format.h" + + +struct pipe_blend_state; +union lp_type; +struct lp_build_context; + + +/** + * Whether the blending function is commutative or not. + */ +boolean +lp_build_blend_func_commutative(unsigned func); + + +/** + * Whether the blending functions are the reverse of each other. + */ +boolean +lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func); + + +LLVMValueRef +lp_build_blend_func(struct lp_build_context *bld, + unsigned func, + LLVMValueRef term1, + LLVMValueRef term2); + + +LLVMValueRef +lp_build_blend_aos(LLVMBuilderRef builder, + const struct pipe_blend_state *blend, + union lp_type type, + LLVMValueRef src, + LLVMValueRef dst, + LLVMValueRef const_, + unsigned alpha_swizzle); + + +void +lp_build_blend_soa(LLVMBuilderRef builder, + const struct pipe_blend_state *blend, + union lp_type type, + LLVMValueRef src[4], + LLVMValueRef dst[4], + LLVMValueRef const_[4], + LLVMValueRef res[4]); + + +/** + * Apply a logic op. + * + * src/dst parameters are packed values. It should work regardless the inputs + * are scalars, or a vector. + */ +LLVMValueRef +lp_build_logicop(LLVMBuilderRef builder, + unsigned logicop_func, + LLVMValueRef src, + LLVMValueRef dst); + + +#endif /* !LP_BLD_BLEND_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c new file mode 100644 index 00000000000..c11a9398f87 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c @@ -0,0 +1,356 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Blend LLVM IR generation -- AoS layout. + * + * AoS blending is in general much slower than SoA, but there are some cases + * where it might be faster. In particular, if a pixel is rendered only once + * then the overhead of tiling and untiling will dominate over the speedup that + * SoA gives. So we might want to detect such cases and fallback to AoS in the + * future, but for now this function is here for historical/benchmarking + * purposes. + * + * Run lp_blend_test after any change to this file. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "pipe/p_state.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_arit.h" +#include "lp_bld_logic.h" +#include "lp_bld_swizzle.h" +#include "lp_bld_blend.h" +#include "lp_bld_debug.h" + + +/** + * We may the same values several times, so we keep them here to avoid + * recomputing them. Also reusing the values allows us to do simplifications + * that LLVM optimization passes wouldn't normally be able to do. + */ +struct lp_build_blend_aos_context +{ + struct lp_build_context base; + + LLVMValueRef src; + LLVMValueRef dst; + LLVMValueRef const_; + + LLVMValueRef inv_src; + LLVMValueRef inv_dst; + LLVMValueRef inv_const; + LLVMValueRef saturate; + + LLVMValueRef rgb_src_factor; + LLVMValueRef alpha_src_factor; + LLVMValueRef rgb_dst_factor; + LLVMValueRef alpha_dst_factor; +}; + + +static LLVMValueRef +lp_build_blend_factor_unswizzled(struct lp_build_blend_aos_context *bld, + unsigned factor, + boolean alpha) +{ + switch (factor) { + case PIPE_BLENDFACTOR_ZERO: + return bld->base.zero; + case PIPE_BLENDFACTOR_ONE: + return bld->base.one; + case PIPE_BLENDFACTOR_SRC_COLOR: + case PIPE_BLENDFACTOR_SRC_ALPHA: + return bld->src; + case PIPE_BLENDFACTOR_DST_COLOR: + case PIPE_BLENDFACTOR_DST_ALPHA: + return bld->dst; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + if(alpha) + return bld->base.one; + else { + if(!bld->inv_dst) + bld->inv_dst = lp_build_comp(&bld->base, bld->dst); + if(!bld->saturate) + bld->saturate = lp_build_min(&bld->base, bld->src, bld->inv_dst); + return bld->saturate; + } + case PIPE_BLENDFACTOR_CONST_COLOR: + case PIPE_BLENDFACTOR_CONST_ALPHA: + return bld->const_; + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + /* TODO */ + assert(0); + return bld->base.zero; + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + if(!bld->inv_src) + bld->inv_src = lp_build_comp(&bld->base, bld->src); + return bld->inv_src; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + if(!bld->inv_dst) + bld->inv_dst = lp_build_comp(&bld->base, bld->dst); + return bld->inv_dst; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + if(!bld->inv_const) + bld->inv_const = lp_build_comp(&bld->base, bld->const_); + return bld->inv_const; + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + /* TODO */ + assert(0); + return bld->base.zero; + default: + assert(0); + return bld->base.zero; + } +} + + +enum lp_build_blend_swizzle { + LP_BUILD_BLEND_SWIZZLE_RGBA = 0, + LP_BUILD_BLEND_SWIZZLE_AAAA = 1, +}; + + +/** + * How should we shuffle the base factor. + */ +static enum lp_build_blend_swizzle +lp_build_blend_factor_swizzle(unsigned factor) +{ + switch (factor) { + case PIPE_BLENDFACTOR_ONE: + case PIPE_BLENDFACTOR_ZERO: + case PIPE_BLENDFACTOR_SRC_COLOR: + case PIPE_BLENDFACTOR_DST_COLOR: + case PIPE_BLENDFACTOR_CONST_COLOR: + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + case PIPE_BLENDFACTOR_INV_DST_COLOR: + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + return LP_BUILD_BLEND_SWIZZLE_RGBA; + case PIPE_BLENDFACTOR_SRC_ALPHA: + case PIPE_BLENDFACTOR_DST_ALPHA: + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + case PIPE_BLENDFACTOR_CONST_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + return LP_BUILD_BLEND_SWIZZLE_AAAA; + default: + assert(0); + return LP_BUILD_BLEND_SWIZZLE_RGBA; + } +} + + +static LLVMValueRef +lp_build_blend_swizzle(struct lp_build_blend_aos_context *bld, + LLVMValueRef rgb, + LLVMValueRef alpha, + enum lp_build_blend_swizzle rgb_swizzle, + unsigned alpha_swizzle) +{ + if(rgb == alpha) { + if(rgb_swizzle == LP_BUILD_BLEND_SWIZZLE_RGBA) + return rgb; + if(rgb_swizzle == LP_BUILD_BLEND_SWIZZLE_AAAA) + return lp_build_broadcast_aos(&bld->base, rgb, alpha_swizzle); + } + else { + if(rgb_swizzle == LP_BUILD_BLEND_SWIZZLE_RGBA) { + boolean cond[4] = {0, 0, 0, 0}; + cond[alpha_swizzle] = 1; + return lp_build_select_aos(&bld->base, alpha, rgb, cond); + } + if(rgb_swizzle == LP_BUILD_BLEND_SWIZZLE_AAAA) { + unsigned char swizzle[4]; + swizzle[0] = alpha_swizzle; + swizzle[1] = alpha_swizzle; + swizzle[2] = alpha_swizzle; + swizzle[3] = alpha_swizzle; + swizzle[alpha_swizzle] += 4; + return lp_build_swizzle2_aos(&bld->base, rgb, alpha, swizzle); + } + } + assert(0); + return bld->base.undef; +} + + +/** + * @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendFuncSeparate.xml + */ +static LLVMValueRef +lp_build_blend_factor(struct lp_build_blend_aos_context *bld, + LLVMValueRef factor1, + unsigned rgb_factor, + unsigned alpha_factor, + unsigned alpha_swizzle) +{ + LLVMValueRef rgb_factor_; + LLVMValueRef alpha_factor_; + LLVMValueRef factor2; + enum lp_build_blend_swizzle rgb_swizzle; + + rgb_factor_ = lp_build_blend_factor_unswizzled(bld, rgb_factor, FALSE); + alpha_factor_ = lp_build_blend_factor_unswizzled(bld, alpha_factor, TRUE); + + rgb_swizzle = lp_build_blend_factor_swizzle(rgb_factor); + + factor2 = lp_build_blend_swizzle(bld, rgb_factor_, alpha_factor_, rgb_swizzle, alpha_swizzle); + + return lp_build_mul(&bld->base, factor1, factor2); +} + + +boolean +lp_build_blend_func_commutative(unsigned func) +{ + switch (func) { + case PIPE_BLEND_ADD: + case PIPE_BLEND_MIN: + case PIPE_BLEND_MAX: + return TRUE; + case PIPE_BLEND_SUBTRACT: + case PIPE_BLEND_REVERSE_SUBTRACT: + return FALSE; + default: + assert(0); + return TRUE; + } +} + + +boolean +lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func) +{ + if(rgb_func == alpha_func) + return FALSE; + if(rgb_func == PIPE_BLEND_SUBTRACT && alpha_func == PIPE_BLEND_REVERSE_SUBTRACT) + return TRUE; + if(rgb_func == PIPE_BLEND_REVERSE_SUBTRACT && alpha_func == PIPE_BLEND_SUBTRACT) + return TRUE; + return FALSE; +} + + +/** + * @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendEquationSeparate.xml + */ +LLVMValueRef +lp_build_blend_func(struct lp_build_context *bld, + unsigned func, + LLVMValueRef term1, + LLVMValueRef term2) +{ + switch (func) { + case PIPE_BLEND_ADD: + return lp_build_add(bld, term1, term2); + break; + case PIPE_BLEND_SUBTRACT: + return lp_build_sub(bld, term1, term2); + case PIPE_BLEND_REVERSE_SUBTRACT: + return lp_build_sub(bld, term2, term1); + case PIPE_BLEND_MIN: + return lp_build_min(bld, term1, term2); + case PIPE_BLEND_MAX: + return lp_build_max(bld, term1, term2); + default: + assert(0); + return bld->zero; + } +} + + +LLVMValueRef +lp_build_blend_aos(LLVMBuilderRef builder, + const struct pipe_blend_state *blend, + union lp_type type, + LLVMValueRef src, + LLVMValueRef dst, + LLVMValueRef const_, + unsigned alpha_swizzle) +{ + struct lp_build_blend_aos_context bld; + LLVMValueRef src_term; + LLVMValueRef dst_term; + + /* FIXME */ + assert(blend->colormask == 0xf); + + if(!blend->blend_enable) + return src; + + /* It makes no sense to blend unless values are normalized */ + assert(type.norm); + + /* Setup build context */ + memset(&bld, 0, sizeof bld); + lp_build_context_init(&bld.base, builder, type); + bld.src = src; + bld.dst = dst; + bld.const_ = const_; + + /* TODO: There are still a few optimization opportunities here. For certain + * combinations it is possible to reorder the operations and therefore saving + * some instructions. */ + + src_term = lp_build_blend_factor(&bld, src, blend->rgb_src_factor, blend->alpha_src_factor, alpha_swizzle); + dst_term = lp_build_blend_factor(&bld, dst, blend->rgb_dst_factor, blend->alpha_dst_factor, alpha_swizzle); + + lp_build_name(src_term, "src_term"); + lp_build_name(dst_term, "dst_term"); + + if(blend->rgb_func == blend->alpha_func) { + return lp_build_blend_func(&bld.base, blend->rgb_func, src_term, dst_term); + } + else { + /* Seperate RGB / A functions */ + + LLVMValueRef rgb; + LLVMValueRef alpha; + + rgb = lp_build_blend_func(&bld.base, blend->rgb_func, src_term, dst_term); + alpha = lp_build_blend_func(&bld.base, blend->alpha_func, src_term, dst_term); + + return lp_build_blend_swizzle(&bld, rgb, alpha, LP_BUILD_BLEND_SWIZZLE_RGBA, alpha_swizzle); + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_logicop.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_logicop.c new file mode 100644 index 00000000000..88321f62a2c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_logicop.c @@ -0,0 +1,108 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Blend LLVM IR generation -- logic ops. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "pipe/p_state.h" + +#include "lp_bld_blend.h" + + +LLVMValueRef +lp_build_logicop(LLVMBuilderRef builder, + unsigned logicop_func, + LLVMValueRef src, + LLVMValueRef dst) +{ + LLVMTypeRef type; + LLVMValueRef res; + + type = LLVMTypeOf(src); + + switch (logicop_func) { + case PIPE_LOGICOP_CLEAR: + res = LLVMConstNull(type); + break; + case PIPE_LOGICOP_NOR: + res = LLVMBuildNot(builder, LLVMBuildOr(builder, src, dst, ""), ""); + break; + case PIPE_LOGICOP_AND_INVERTED: + res = LLVMBuildAnd(builder, LLVMBuildNot(builder, src, ""), dst, ""); + break; + case PIPE_LOGICOP_COPY_INVERTED: + res = LLVMBuildNot(builder, src, ""); + break; + case PIPE_LOGICOP_AND_REVERSE: + res = LLVMBuildAnd(builder, src, LLVMBuildNot(builder, dst, ""), ""); + break; + case PIPE_LOGICOP_INVERT: + res = LLVMBuildNot(builder, dst, ""); + break; + case PIPE_LOGICOP_XOR: + res = LLVMBuildXor(builder, src, dst, ""); + break; + case PIPE_LOGICOP_NAND: + res = LLVMBuildNot(builder, LLVMBuildAnd(builder, src, dst, ""), ""); + break; + case PIPE_LOGICOP_AND: + res = LLVMBuildAnd(builder, src, dst, ""); + break; + case PIPE_LOGICOP_EQUIV: + res = LLVMBuildNot(builder, LLVMBuildXor(builder, src, dst, ""), ""); + break; + case PIPE_LOGICOP_NOOP: + res = dst; + break; + case PIPE_LOGICOP_OR_INVERTED: + res = LLVMBuildOr(builder, LLVMBuildNot(builder, src, ""), dst, ""); + break; + case PIPE_LOGICOP_COPY: + res = src; + break; + case PIPE_LOGICOP_OR_REVERSE: + res = LLVMBuildOr(builder, src, LLVMBuildNot(builder, dst, ""), ""); + break; + case PIPE_LOGICOP_OR: + res = LLVMBuildOr(builder, src, dst, ""); + break; + case PIPE_LOGICOP_SET: + res = LLVMConstAllOnes(type); + break; + default: + assert(0); + res = src; + } + + return res; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c new file mode 100644 index 00000000000..b92254a7d6f --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c @@ -0,0 +1,298 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Blend LLVM IR generation -- SoA layout. + * + * Blending in SoA is much faster than AoS, especially when separate rgb/alpha + * factors/functions are used, since no channel masking/shuffling is necessary + * and we can achieve the full throughput of the SIMD operations. Furthermore + * the fragment shader output is also in SoA, so it fits nicely with the rest of + * the fragment pipeline. + * + * The drawback is that to be displayed the color buffer needs to be in AoS + * layout, so we need to tile/untile the color buffer before/after rendering. + * A color buffer like + * + * R11 G11 B11 A11 R12 G12 B12 A12 R13 G13 B13 A13 R14 G14 B14 A14 ... + * R21 G21 B21 A21 R22 G22 B22 A22 R23 G23 B23 A23 R24 G24 B24 A24 ... + * + * R31 G31 B31 A31 R32 G32 B32 A32 R33 G33 B33 A33 R34 G34 B34 A34 ... + * R41 G41 B41 A41 R42 G42 B42 A42 R43 G43 B43 A43 R44 G44 B44 A44 ... + * + * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... + * + * will actually be stored in memory as + * + * R11 R12 R21 R22 R13 R14 R23 R24 ... G11 G12 G21 G22 G13 G14 G23 G24 ... B11 B12 B21 B22 B13 B14 B23 B24 ... A11 A12 A21 A22 A13 A14 A23 A24 ... + * R31 R32 R41 R42 R33 R34 R43 R44 ... G31 G32 G41 G42 G33 G34 G43 G44 ... B31 B32 B41 B42 B33 B34 B43 B44 ... A31 A32 A41 A42 A33 A34 A43 A44 ... + * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... + * + * NOTE: Run lp_blend_test after any change to this file. + * + * You can also run lp_blend_test to obtain AoS vs SoA benchmarks. Invoking it + * as: + * + * lp_blend_test -o blend.tsv + * + * will generate a tab-seperated-file with the test results and performance + * measurements. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "pipe/p_state.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_arit.h" +#include "lp_bld_blend.h" + + +/** + * We may the same values several times, so we keep them here to avoid + * recomputing them. Also reusing the values allows us to do simplifications + * that LLVM optimization passes wouldn't normally be able to do. + */ +struct lp_build_blend_soa_context +{ + struct lp_build_context base; + + LLVMValueRef src[4]; + LLVMValueRef dst[4]; + LLVMValueRef con[4]; + + LLVMValueRef inv_src[4]; + LLVMValueRef inv_dst[4]; + LLVMValueRef inv_con[4]; + + LLVMValueRef src_alpha_saturate; + + /** + * We store all factors in a table in order to eliminate redundant + * multiplications later. + */ + LLVMValueRef factor[2][2][4]; + + /** + * Table with all terms. + */ + LLVMValueRef term[2][4]; +}; + + +static LLVMValueRef +lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld, + unsigned factor, unsigned i) +{ + /* + * Compute src/first term RGB + */ + switch (factor) { + case PIPE_BLENDFACTOR_ONE: + return bld->base.one; + case PIPE_BLENDFACTOR_SRC_COLOR: + return bld->src[i]; + case PIPE_BLENDFACTOR_SRC_ALPHA: + return bld->src[3]; + case PIPE_BLENDFACTOR_DST_COLOR: + return bld->dst[i]; + case PIPE_BLENDFACTOR_DST_ALPHA: + return bld->dst[3]; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + if(i == 3) + return bld->base.one; + else { + if(!bld->inv_dst[3]) + bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]); + if(!bld->src_alpha_saturate) + bld->src_alpha_saturate = lp_build_min(&bld->base, bld->src[3], bld->inv_dst[3]); + return bld->src_alpha_saturate; + } + case PIPE_BLENDFACTOR_CONST_COLOR: + return bld->con[i]; + case PIPE_BLENDFACTOR_CONST_ALPHA: + return bld->con[3]; + case PIPE_BLENDFACTOR_SRC1_COLOR: + /* TODO */ + assert(0); + return bld->base.zero; + case PIPE_BLENDFACTOR_SRC1_ALPHA: + /* TODO */ + assert(0); + return bld->base.zero; + case PIPE_BLENDFACTOR_ZERO: + return bld->base.zero; + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + if(!bld->inv_src[i]) + bld->inv_src[i] = lp_build_comp(&bld->base, bld->src[i]); + return bld->inv_src[i]; + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + if(!bld->inv_src[3]) + bld->inv_src[3] = lp_build_comp(&bld->base, bld->src[3]); + return bld->inv_src[3]; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + if(!bld->inv_dst[i]) + bld->inv_dst[i] = lp_build_comp(&bld->base, bld->dst[i]); + return bld->inv_dst[i]; + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + if(!bld->inv_dst[3]) + bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]); + return bld->inv_dst[3]; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + if(!bld->inv_con[i]) + bld->inv_con[i] = lp_build_comp(&bld->base, bld->con[i]); + return bld->inv_con[i]; + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + if(!bld->inv_con[3]) + bld->inv_con[3] = lp_build_comp(&bld->base, bld->con[3]); + return bld->inv_con[3]; + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + /* TODO */ + assert(0); + return bld->base.zero; + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + /* TODO */ + assert(0); + return bld->base.zero; + default: + assert(0); + return bld->base.zero; + } +} + + +/** + * Generate blend code in SOA mode. + * \param src src/fragment color + * \param dst dst/framebuffer color + * \param con constant blend color + * \param res the result/output + */ +void +lp_build_blend_soa(LLVMBuilderRef builder, + const struct pipe_blend_state *blend, + union lp_type type, + LLVMValueRef src[4], + LLVMValueRef dst[4], + LLVMValueRef con[4], + LLVMValueRef res[4]) +{ + struct lp_build_blend_soa_context bld; + unsigned i, j, k; + + /* Setup build context */ + memset(&bld, 0, sizeof bld); + lp_build_context_init(&bld.base, builder, type); + for (i = 0; i < 4; ++i) { + bld.src[i] = src[i]; + bld.dst[i] = dst[i]; + bld.con[i] = con[i]; + } + + for (i = 0; i < 4; ++i) { + if (blend->colormask & (1 << i)) { + if (blend->logicop_enable) { + if(!type.floating) { + res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]); + } + else + res[i] = dst[i]; + } + else if (blend->blend_enable) { + unsigned src_factor = i < 3 ? blend->rgb_src_factor : blend->alpha_src_factor; + unsigned dst_factor = i < 3 ? blend->rgb_dst_factor : blend->alpha_dst_factor; + unsigned func = i < 3 ? blend->rgb_func : blend->alpha_func; + boolean func_commutative = lp_build_blend_func_commutative(func); + + /* It makes no sense to blend unless values are normalized */ + assert(type.norm); + + /* + * Compute src/dst factors. + */ + + bld.factor[0][0][i] = src[i]; + bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i); + bld.factor[1][0][i] = dst[i]; + bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i); + + /* + * Compute src/dst terms + */ + + for(k = 0; k < 2; ++k) { + /* See if this multiplication has been previously computed */ + for(j = 0; j < i; ++j) { + if((bld.factor[k][0][j] == bld.factor[k][0][i] && + bld.factor[k][1][j] == bld.factor[k][1][i]) || + (bld.factor[k][0][j] == bld.factor[k][1][i] && + bld.factor[k][1][j] == bld.factor[k][0][i])) + break; + } + + if(j < i) + bld.term[k][i] = bld.term[k][j]; + else + bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]); + } + + /* + * Combine terms + */ + + /* See if this function has been previously applied */ + for(j = 0; j < i; ++j) { + unsigned prev_func = j < 3 ? blend->rgb_func : blend->alpha_func; + unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func); + + if((!func_reverse && + bld.term[0][j] == bld.term[0][i] && + bld.term[1][j] == bld.term[1][i]) || + ((func_commutative || func_reverse) && + bld.term[0][j] == bld.term[1][i] && + bld.term[1][j] == bld.term[0][i])) + break; + } + + if(j < i) + res[i] = res[j]; + else + res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]); + } + else { + res[i] = src[i]; + } + } + else { + res[i] = dst[i]; + } + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.c b/src/gallium/drivers/llvmpipe/lp_bld_const.c new file mode 100644 index 00000000000..21487365eae --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_const.c @@ -0,0 +1,369 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helper functions for constant building. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include <float.h> + +#include "util/u_debug.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" + + +unsigned +lp_mantissa(union lp_type type) +{ + assert(type.floating); + + if(type.floating) { + switch(type.width) { + case 32: + return 23; + case 64: + return 53; + default: + assert(0); + return 0; + } + } + else { + if(type.sign) + return type.width - 1; + else + return type.width; + } +} + + +/** + * Shift of the unity. + * + * Same as lp_const_scale(), but in terms of shifts. + */ +unsigned +lp_const_shift(union lp_type type) +{ + if(type.floating) + return 0; + else if(type.fixed) + return type.width/2; + else if(type.norm) + return type.sign ? type.width - 1 : type.width; + else + return 0; +} + + +unsigned +lp_const_offset(union lp_type type) +{ + if(type.floating || type.fixed) + return 0; + else if(type.norm) + return 1; + else + return 0; +} + + +/** + * Scaling factor between the LLVM native value and its interpretation. + * + * This is 1.0 for all floating types and unnormalized integers, and something + * else for the fixed points types and normalized integers. + */ +double +lp_const_scale(union lp_type type) +{ + unsigned long long llscale; + double dscale; + + llscale = (unsigned long long)1 << lp_const_shift(type); + llscale -= lp_const_offset(type); + dscale = (double)llscale; + assert((unsigned long long)dscale == llscale); + + return dscale; +} + + +/** + * Minimum value representable by the type. + */ +double +lp_const_min(union lp_type type) +{ + unsigned bits; + + if(!type.sign) + return 0.0; + + if(type.norm) + return -1.0; + + if (type.floating) { + switch(type.width) { + case 32: + return -FLT_MAX; + case 64: + return -DBL_MAX; + default: + assert(0); + return 0.0; + } + } + + if(type.fixed) + /* FIXME: consider the fractional bits? */ + bits = type.width / 2 - 1; + else + bits = type.width - 1; + + return (double)-((long long)1 << bits); +} + + +/** + * Maximum value representable by the type. + */ +double +lp_const_max(union lp_type type) +{ + unsigned bits; + + if(type.norm) + return 1.0; + + if (type.floating) { + switch(type.width) { + case 32: + return FLT_MAX; + case 64: + return DBL_MAX; + default: + assert(0); + return 0.0; + } + } + + if(type.fixed) + bits = type.width / 2; + else + bits = type.width; + + if(type.sign) + bits -= 1; + + return (double)(((unsigned long long)1 << bits) - 1); +} + + +double +lp_const_eps(union lp_type type) +{ + if (type.floating) { + switch(type.width) { + case 32: + return FLT_EPSILON; + case 64: + return DBL_EPSILON; + default: + assert(0); + return 0.0; + } + } + else { + double scale = lp_const_scale(type); + return 1.0/scale; + } +} + + +LLVMValueRef +lp_build_undef(union lp_type type) +{ + LLVMTypeRef vec_type = lp_build_vec_type(type); + return LLVMGetUndef(vec_type); +} + + +LLVMValueRef +lp_build_zero(union lp_type type) +{ + LLVMTypeRef vec_type = lp_build_vec_type(type); + return LLVMConstNull(vec_type); +} + + +LLVMValueRef +lp_build_one(union lp_type type) +{ + LLVMTypeRef elem_type; + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(type.length <= LP_MAX_VECTOR_LENGTH); + + elem_type = lp_build_elem_type(type); + + if(type.floating) + elems[0] = LLVMConstReal(elem_type, 1.0); + else if(type.fixed) + elems[0] = LLVMConstInt(elem_type, 1LL << (type.width/2), 0); + else if(!type.norm) + elems[0] = LLVMConstInt(elem_type, 1, 0); + else if(type.sign) + elems[0] = LLVMConstInt(elem_type, (1LL << (type.width - 1)) - 1, 0); + else { + /* special case' -- 1.0 for normalized types is more easily attained if + * we start with a vector consisting of all bits set */ + LLVMTypeRef vec_type = LLVMVectorType(elem_type, type.length); + LLVMValueRef vec = LLVMConstAllOnes(vec_type); + +#if 0 + if(type.sign) + /* TODO: Unfortunately this caused "Tried to create a shift operation + * on a non-integer type!" */ + vec = LLVMConstLShr(vec, lp_build_int_const_scalar(type, 1)); +#endif + + return vec; + } + + for(i = 1; i < type.length; ++i) + elems[i] = elems[0]; + + return LLVMConstVector(elems, type.length); +} + + +LLVMValueRef +lp_build_const_scalar(union lp_type type, + double val) +{ + LLVMTypeRef elem_type = lp_build_elem_type(type); + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(type.length <= LP_MAX_VECTOR_LENGTH); + + if(type.floating) { + elems[0] = LLVMConstReal(elem_type, val); + } + else { + double dscale = lp_const_scale(type); + + elems[0] = LLVMConstInt(elem_type, val*dscale + 0.5, 0); + } + + for(i = 1; i < type.length; ++i) + elems[i] = elems[0]; + + return LLVMConstVector(elems, type.length); +} + + +LLVMValueRef +lp_build_int_const_scalar(union lp_type type, + long long val) +{ + LLVMTypeRef elem_type = lp_build_int_elem_type(type); + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(type.length <= LP_MAX_VECTOR_LENGTH); + + for(i = 0; i < type.length; ++i) + elems[i] = LLVMConstInt(elem_type, val, type.sign ? 1 : 0); + + return LLVMConstVector(elems, type.length); +} + + +LLVMValueRef +lp_build_const_aos(union lp_type type, + double r, double g, double b, double a, + const unsigned char *swizzle) +{ + const unsigned char default_swizzle[4] = {0, 1, 2, 3}; + LLVMTypeRef elem_type; + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(type.length % 4 == 0); + assert(type.length <= LP_MAX_VECTOR_LENGTH); + + elem_type = lp_build_elem_type(type); + + if(swizzle == NULL) + swizzle = default_swizzle; + + if(type.floating) { + elems[swizzle[0]] = LLVMConstReal(elem_type, r); + elems[swizzle[1]] = LLVMConstReal(elem_type, g); + elems[swizzle[2]] = LLVMConstReal(elem_type, b); + elems[swizzle[3]] = LLVMConstReal(elem_type, a); + } + else { + double dscale = lp_const_scale(type); + + elems[swizzle[0]] = LLVMConstInt(elem_type, r*dscale + 0.5, 0); + elems[swizzle[1]] = LLVMConstInt(elem_type, g*dscale + 0.5, 0); + elems[swizzle[2]] = LLVMConstInt(elem_type, b*dscale + 0.5, 0); + elems[swizzle[3]] = LLVMConstInt(elem_type, a*dscale + 0.5, 0); + } + + for(i = 4; i < type.length; ++i) + elems[i] = elems[i % 4]; + + return LLVMConstVector(elems, type.length); +} + + +LLVMValueRef +lp_build_const_mask_aos(union lp_type type, + boolean cond[4]) +{ + LLVMTypeRef elem_type = LLVMIntType(type.width); + LLVMValueRef masks[LP_MAX_VECTOR_LENGTH]; + unsigned i, j; + + assert(type.length <= LP_MAX_VECTOR_LENGTH); + + for(j = 0; j < type.length; j += 4) + for(i = 0; i < 4; ++i) + masks[j + i] = LLVMConstInt(elem_type, cond[i] ? ~0 : 0, 0); + + return LLVMConstVector(masks, type.length); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.h b/src/gallium/drivers/llvmpipe/lp_bld_const.h new file mode 100644 index 00000000000..1934530ea3c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_const.h @@ -0,0 +1,108 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for constant building. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_CONST_H +#define LP_BLD_CONST_H + + +#include <llvm-c/Core.h> + +#include <pipe/p_compiler.h> + + +union lp_type type; + + +unsigned +lp_mantissa(union lp_type type); + + +unsigned +lp_const_shift(union lp_type type); + + +unsigned +lp_const_offset(union lp_type type); + + +double +lp_const_scale(union lp_type type); + +double +lp_const_min(union lp_type type); + + +double +lp_const_max(union lp_type type); + + +double +lp_const_eps(union lp_type type); + + +LLVMValueRef +lp_build_undef(union lp_type type); + + +LLVMValueRef +lp_build_zero(union lp_type type); + + +LLVMValueRef +lp_build_one(union lp_type type); + + +LLVMValueRef +lp_build_const_scalar(union lp_type type, + double val); + + +LLVMValueRef +lp_build_int_const_scalar(union lp_type type, + long long val); + + +LLVMValueRef +lp_build_const_aos(union lp_type type, + double r, double g, double b, double a, + const unsigned char *swizzle); + + +LLVMValueRef +lp_build_const_mask_aos(union lp_type type, + boolean cond[4]); + + +#endif /* !LP_BLD_CONST_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.c b/src/gallium/drivers/llvmpipe/lp_bld_conv.c new file mode 100644 index 00000000000..c8954c8a34f --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.c @@ -0,0 +1,702 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helper functions for type conversions. + * + * We want to use the fastest type for a given computation whenever feasible. + * The other side of this is that we need to be able convert between several + * types accurately and efficiently. + * + * Conversion between types of different bit width is quite complex since a + * + * To remember there are a few invariants in type conversions: + * + * - register width must remain constant: + * + * src_type.width * src_type.length == dst_type.width * dst_type.length + * + * - total number of elements must remain constant: + * + * src_type.length * num_srcs == dst_type.length * num_dsts + * + * It is not always possible to do the conversion both accurately and + * efficiently, usually due to lack of adequate machine instructions. In these + * cases it is important not to cut shortcuts here and sacrifice accuracy, as + * there this functions can be used anywhere. In the future we might have a + * precision parameter which can gauge the accuracy vs efficiency compromise, + * but for now if the data conversion between two stages happens to be the + * bottleneck, then most likely should just avoid converting at all and run + * both stages with the same type. + * + * Make sure to run lp_test_conv unit test after any change to this file. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_debug.h" +#include "util/u_math.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_intr.h" +#include "lp_bld_arit.h" +#include "lp_bld_conv.h" + + +/** + * Special case for converting clamped IEEE-754 floats to unsigned norms. + * + * The mathematical voodoo below may seem excessive but it is actually + * paramount we do it this way for several reasons. First, there is no single + * precision FP to unsigned integer conversion Intel SSE instruction. Second, + * secondly, even if there was, since the FP's mantissa takes only a fraction + * of register bits the typically scale and cast approach would require double + * precision for accurate results, and therefore half the throughput + * + * Although the result values can be scaled to an arbitrary bit width specified + * by dst_width, the actual result type will have the same width. + */ +LLVMValueRef +lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, + union lp_type src_type, + unsigned dst_width, + LLVMValueRef src) +{ + LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type); + LLVMValueRef res; + unsigned mantissa; + unsigned n; + unsigned long long ubound; + unsigned long long mask; + double scale; + double bias; + + assert(src_type.floating); + + mantissa = lp_mantissa(src_type); + + /* We cannot carry more bits than the mantissa */ + n = MIN2(mantissa, dst_width); + + /* This magic coefficients will make the desired result to appear in the + * lowest significant bits of the mantissa. + */ + ubound = ((unsigned long long)1 << n); + mask = ubound - 1; + scale = (double)mask/ubound; + bias = (double)((unsigned long long)1 << (mantissa - n)); + + res = LLVMBuildMul(builder, src, lp_build_const_scalar(src_type, scale), ""); + res = LLVMBuildAdd(builder, res, lp_build_const_scalar(src_type, bias), ""); + res = LLVMBuildBitCast(builder, res, int_vec_type, ""); + + if(dst_width > n) { + int shift = dst_width - n; + res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), ""); + + /* Fill in the empty lower bits for added precision? */ +#if 0 + { + LLVMValueRef msb; + msb = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, dst_width - 1), ""); + msb = LLVMBuildShl(builder, msb, lp_build_int_const_scalar(src_type, shift), ""); + msb = LLVMBuildSub(builder, msb, lp_build_int_const_scalar(src_type, 1), ""); + res = LLVMBuildOr(builder, res, msb, ""); + } +#elif 0 + while(shift > 0) { + res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, n), ""), ""); + shift -= n; + n *= 2; + } +#endif + } + else + res = LLVMBuildAnd(builder, res, lp_build_int_const_scalar(src_type, mask), ""); + + return res; +} + + +/** + * Inverse of lp_build_clamped_float_to_unsigned_norm above. + */ +LLVMValueRef +lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, + unsigned src_width, + union lp_type dst_type, + LLVMValueRef src) +{ + LLVMTypeRef vec_type = lp_build_vec_type(dst_type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type); + LLVMValueRef bias_; + LLVMValueRef res; + unsigned mantissa; + unsigned n; + unsigned long long ubound; + unsigned long long mask; + double scale; + double bias; + + mantissa = lp_mantissa(dst_type); + + n = MIN2(mantissa, src_width); + + ubound = ((unsigned long long)1 << n); + mask = ubound - 1; + scale = (double)ubound/mask; + bias = (double)((unsigned long long)1 << (mantissa - n)); + + res = src; + + if(src_width > mantissa) { + int shift = src_width - mantissa; + res = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(dst_type, shift), ""); + } + + bias_ = lp_build_const_scalar(dst_type, bias); + + res = LLVMBuildOr(builder, + res, + LLVMBuildBitCast(builder, bias_, int_vec_type, ""), ""); + + res = LLVMBuildBitCast(builder, res, vec_type, ""); + + res = LLVMBuildSub(builder, res, bias_, ""); + res = LLVMBuildMul(builder, res, lp_build_const_scalar(dst_type, scale), ""); + + return res; +} + + +/** + * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions. + */ +static LLVMValueRef +lp_build_const_unpack_shuffle(unsigned n, unsigned lo_hi) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i, j; + + assert(n <= LP_MAX_VECTOR_LENGTH); + assert(lo_hi < 2); + + /* TODO: cache results in a static table */ + + for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) { + elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0); + elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0); + } + + return LLVMConstVector(elems, n); +} + + +/** + * Build shuffle vectors that match PACKxx instructions. + */ +static LLVMValueRef +lp_build_const_pack_shuffle(unsigned n) +{ + LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + assert(n <= LP_MAX_VECTOR_LENGTH); + + /* TODO: cache results in a static table */ + + for(i = 0; i < n; ++i) + elems[i] = LLVMConstInt(LLVMInt32Type(), 2*i, 0); + + return LLVMConstVector(elems, n); +} + + +/** + * Expand the bit width. + * + * This will only change the number of bits the values are represented, not the + * values themselved. + */ +static void +lp_build_expand(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + LLVMValueRef src, + LLVMValueRef *dst, unsigned num_dsts) +{ + unsigned num_tmps; + unsigned i; + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length == dst_type.length * num_dsts); + + num_tmps = 1; + dst[0] = src; + + while(src_type.width < dst_type.width) { + union lp_type new_type = src_type; + LLVMTypeRef new_vec_type; + + new_type.width *= 2; + new_type.length /= 2; + new_vec_type = lp_build_vec_type(new_type); + + for(i = num_tmps; i--; ) { + LLVMValueRef zero; + LLVMValueRef shuffle_lo; + LLVMValueRef shuffle_hi; + LLVMValueRef lo; + LLVMValueRef hi; + + zero = lp_build_zero(src_type); + shuffle_lo = lp_build_const_unpack_shuffle(src_type.length, 0); + shuffle_hi = lp_build_const_unpack_shuffle(src_type.length, 1); + + /* PUNPCKLBW, PUNPCKHBW */ + lo = LLVMBuildShuffleVector(builder, dst[i], zero, shuffle_lo, ""); + hi = LLVMBuildShuffleVector(builder, dst[i], zero, shuffle_hi, ""); + + dst[2*i + 0] = LLVMBuildBitCast(builder, lo, new_vec_type, ""); + dst[2*i + 1] = LLVMBuildBitCast(builder, hi, new_vec_type, ""); + } + + src_type = new_type; + + num_tmps *= 2; + } + + assert(num_tmps == num_dsts); +} + + +/** + * Non-interleaved pack. + * + * This will move values as + * + * lo = __ l0 __ l1 __ l2 __.. __ ln + * hi = __ h0 __ h1 __ h2 __.. __ hn + * res = l0 l1 l2 .. ln h0 h1 h2 .. hn + * + * TODO: handle saturation consistently. + */ +static LLVMValueRef +lp_build_pack2(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + boolean clamped, + LLVMValueRef lo, + LLVMValueRef hi) +{ + LLVMTypeRef src_vec_type = lp_build_vec_type(src_type); + LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type); + LLVMValueRef shuffle; + LLVMValueRef res; + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * 2 == dst_type.length); + + assert(!src_type.floating); + assert(!dst_type.floating); + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(src_type.width * src_type.length == 128) { + /* All X86 non-interleaved pack instructions all take signed inputs and + * saturate them, so saturate beforehand. */ + if(!src_type.sign && !clamped) { + struct lp_build_context bld; + unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width; + LLVMValueRef dst_max = lp_build_int_const_scalar(src_type, ((unsigned long long)1 << dst_bits) - 1); + lp_build_context_init(&bld, builder, src_type); + lo = lp_build_min(&bld, lo, dst_max); + hi = lp_build_min(&bld, hi, dst_max); + } + + switch(src_type.width) { + case 32: + if(dst_type.sign) + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", src_vec_type, lo, hi); + else + /* PACKUSDW is the only instrinsic with a consistent signature */ + return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi); + break; + + case 16: + if(dst_type.sign) + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", src_vec_type, lo, hi); + else + res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", src_vec_type, lo, hi); + break; + + default: + assert(0); + return LLVMGetUndef(dst_vec_type); + break; + } + + res = LLVMBuildBitCast(builder, res, dst_vec_type, ""); + return res; + } +#endif + + lo = LLVMBuildBitCast(builder, lo, dst_vec_type, ""); + hi = LLVMBuildBitCast(builder, hi, dst_vec_type, ""); + + shuffle = lp_build_const_pack_shuffle(dst_type.length); + + res = LLVMBuildShuffleVector(builder, lo, hi, shuffle, ""); + + return res; +} + + +/** + * Truncate the bit width. + * + * TODO: Handle saturation consistently. + */ +static LLVMValueRef +lp_build_trunc(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + boolean clamped, + const LLVMValueRef *src, unsigned num_srcs) +{ + LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * num_srcs == dst_type.length); + + for(i = 0; i < num_srcs; ++i) + tmp[i] = src[i]; + + while(src_type.width > dst_type.width) { + union lp_type new_type = src_type; + + new_type.width /= 2; + new_type.length *= 2; + + /* Take in consideration the sign changes only in the last step */ + if(new_type.width == dst_type.width) + new_type.sign = dst_type.sign; + + num_srcs /= 2; + + for(i = 0; i < num_srcs; ++i) + tmp[i] = lp_build_pack2(builder, src_type, new_type, clamped, + tmp[2*i + 0], tmp[2*i + 1]); + + src_type = new_type; + } + + assert(num_srcs == 1); + + return tmp[0]; +} + + +/** + * Generic type conversion. + * + * TODO: Take a precision argument, or even better, add a new precision member + * to the lp_type union. + */ +void +lp_build_conv(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + const LLVMValueRef *src, unsigned num_srcs, + LLVMValueRef *dst, unsigned num_dsts) +{ + union lp_type tmp_type; + LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH]; + unsigned num_tmps; + unsigned i; + + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * num_srcs == dst_type.length * num_dsts); + + assert(src_type.length <= LP_MAX_VECTOR_LENGTH); + assert(dst_type.length <= LP_MAX_VECTOR_LENGTH); + + tmp_type = src_type; + for(i = 0; i < num_srcs; ++i) + tmp[i] = src[i]; + num_tmps = num_srcs; + + /* + * Clamp if necessary + */ + + if(src_type.value != dst_type.value) { + struct lp_build_context bld; + double src_min = lp_const_min(src_type); + double dst_min = lp_const_min(dst_type); + double src_max = lp_const_max(src_type); + double dst_max = lp_const_max(dst_type); + LLVMValueRef thres; + + lp_build_context_init(&bld, builder, tmp_type); + + if(src_min < dst_min) { + if(dst_min == 0.0) + thres = bld.zero; + else + thres = lp_build_const_scalar(src_type, dst_min); + for(i = 0; i < num_tmps; ++i) + tmp[i] = lp_build_max(&bld, tmp[i], thres); + } + + if(src_max > dst_max) { + if(dst_max == 1.0) + thres = bld.one; + else + thres = lp_build_const_scalar(src_type, dst_max); + for(i = 0; i < num_tmps; ++i) + tmp[i] = lp_build_min(&bld, tmp[i], thres); + } + } + + /* + * Scale to the narrowest range + */ + + if(dst_type.floating) { + /* Nothing to do */ + } + else if(tmp_type.floating) { + if(!dst_type.fixed && !dst_type.sign && dst_type.norm) { + for(i = 0; i < num_tmps; ++i) { + tmp[i] = lp_build_clamped_float_to_unsigned_norm(builder, + tmp_type, + dst_type.width, + tmp[i]); + } + tmp_type.floating = FALSE; + } + else { + double dst_scale = lp_const_scale(dst_type); + LLVMTypeRef tmp_vec_type; + + if (dst_scale != 1.0) { + LLVMValueRef scale = lp_build_const_scalar(tmp_type, dst_scale); + for(i = 0; i < num_tmps; ++i) + tmp[i] = LLVMBuildMul(builder, tmp[i], scale, ""); + } + + /* Use an equally sized integer for intermediate computations */ + tmp_type.floating = FALSE; + tmp_vec_type = lp_build_vec_type(tmp_type); + for(i = 0; i < num_tmps; ++i) { +#if 0 + if(dst_type.sign) + tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, ""); + else + tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, ""); +#else + /* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */ + tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, ""); +#endif + } + } + } + else { + unsigned src_shift = lp_const_shift(src_type); + unsigned dst_shift = lp_const_shift(dst_type); + + /* FIXME: compensate different offsets too */ + if(src_shift > dst_shift) { + LLVMValueRef shift = lp_build_int_const_scalar(tmp_type, src_shift - dst_shift); + for(i = 0; i < num_tmps; ++i) + if(src_type.sign) + tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, ""); + else + tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, ""); + } + } + + /* + * Truncate or expand bit width + */ + + assert(!tmp_type.floating || tmp_type.width == dst_type.width); + + if(tmp_type.width > dst_type.width) { + assert(num_dsts == 1); + tmp[0] = lp_build_trunc(builder, tmp_type, dst_type, TRUE, tmp, num_tmps); + tmp_type.width = dst_type.width; + tmp_type.length = dst_type.length; + num_tmps = 1; + } + + if(tmp_type.width < dst_type.width) { + assert(num_tmps == 1); + lp_build_expand(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts); + tmp_type.width = dst_type.width; + tmp_type.length = dst_type.length; + num_tmps = num_dsts; + } + + assert(tmp_type.width == dst_type.width); + assert(tmp_type.length == dst_type.length); + assert(num_tmps == num_dsts); + + /* + * Scale to the widest range + */ + + if(src_type.floating) { + /* Nothing to do */ + } + else if(!src_type.floating && dst_type.floating) { + if(!src_type.fixed && !src_type.sign && src_type.norm) { + for(i = 0; i < num_tmps; ++i) { + tmp[i] = lp_build_unsigned_norm_to_float(builder, + src_type.width, + dst_type, + tmp[i]); + } + tmp_type.floating = TRUE; + } + else { + double src_scale = lp_const_scale(src_type); + LLVMTypeRef tmp_vec_type; + + /* Use an equally sized integer for intermediate computations */ + tmp_type.floating = TRUE; + tmp_type.sign = TRUE; + tmp_vec_type = lp_build_vec_type(tmp_type); + for(i = 0; i < num_tmps; ++i) { +#if 0 + if(dst_type.sign) + tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, ""); + else + tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, ""); +#else + /* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */ + tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, ""); +#endif + } + + if (src_scale != 1.0) { + LLVMValueRef scale = lp_build_const_scalar(tmp_type, 1.0/src_scale); + for(i = 0; i < num_tmps; ++i) + tmp[i] = LLVMBuildMul(builder, tmp[i], scale, ""); + } + } + } + else { + unsigned src_shift = lp_const_shift(src_type); + unsigned dst_shift = lp_const_shift(dst_type); + + /* FIXME: compensate different offsets too */ + if(src_shift < dst_shift) { + LLVMValueRef shift = lp_build_int_const_scalar(tmp_type, dst_shift - src_shift); + for(i = 0; i < num_tmps; ++i) + tmp[i] = LLVMBuildShl(builder, tmp[i], shift, ""); + } + } + + for(i = 0; i < num_dsts; ++i) + dst[i] = tmp[i]; +} + + +/** + * Bit mask conversion. + * + * This will convert the integer masks that match the given types. + * + * The mask values should 0 or -1, i.e., all bits either set to zero or one. + * Any other value will likely cause in unpredictable results. + * + * This is basically a very trimmed down version of lp_build_conv. + */ +void +lp_build_conv_mask(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + const LLVMValueRef *src, unsigned num_srcs, + LLVMValueRef *dst, unsigned num_dsts) +{ + /* Register width must remain constant */ + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * num_srcs == dst_type.length * num_dsts); + + /* + * Drop + * + * We assume all values are 0 or -1 + */ + + src_type.floating = FALSE; + src_type.fixed = FALSE; + src_type.sign = TRUE; + src_type.norm = FALSE; + + dst_type.floating = FALSE; + dst_type.fixed = FALSE; + dst_type.sign = TRUE; + dst_type.norm = FALSE; + + /* + * Truncate or expand bit width + */ + + if(src_type.width > dst_type.width) { + assert(num_dsts == 1); + dst[0] = lp_build_trunc(builder, src_type, dst_type, TRUE, src, num_srcs); + } + else if(src_type.width < dst_type.width) { + assert(num_srcs == 1); + lp_build_expand(builder, src_type, dst_type, src[0], dst, num_dsts); + } + else { + assert(num_srcs == num_dsts); + memcpy(dst, src, num_dsts * sizeof *dst); + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.h b/src/gallium/drivers/llvmpipe/lp_bld_conv.h new file mode 100644 index 00000000000..05c1ef2a100 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.h @@ -0,0 +1,73 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for type conversions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_CONV_H +#define LP_BLD_CONV_H + + +#include <llvm-c/Core.h> + + +union lp_type type; + + +LLVMValueRef +lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, + union lp_type src_type, + unsigned dst_width, + LLVMValueRef src); + +LLVMValueRef +lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, + unsigned src_width, + union lp_type dst_type, + LLVMValueRef src); + + +void +lp_build_conv(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + const LLVMValueRef *srcs, unsigned num_srcs, + LLVMValueRef *dsts, unsigned num_dsts); + +void +lp_build_conv_mask(LLVMBuilderRef builder, + union lp_type src_type, + union lp_type dst_type, + const LLVMValueRef *src, unsigned num_srcs, + LLVMValueRef *dst, unsigned num_dsts); + +#endif /* !LP_BLD_CONV_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.c b/src/gallium/drivers/llvmpipe/lp_bld_debug.c new file mode 100644 index 00000000000..30925b5f415 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.c @@ -0,0 +1,108 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifdef HAVE_UDIS86 +#include <udis86.h> +#endif + +#include "util/u_debug.h" +#include "lp_bld_debug.h" + + +void +lp_disassemble(const void* func) +{ +#ifdef HAVE_UDIS86 + ud_t ud_obj; + uint64_t max_jmp_pc; + + ud_init(&ud_obj); + + ud_set_input_buffer(&ud_obj, (void*)func, 0xffff); + + max_jmp_pc = (uint64_t) (uintptr_t) func; + ud_set_pc(&ud_obj, max_jmp_pc); + +#ifdef PIPE_ARCH_X86 + ud_set_mode(&ud_obj, 32); +#endif +#ifdef PIPE_ARCH_X86_64 + ud_set_mode(&ud_obj, 64); +#endif + + ud_set_syntax(&ud_obj, UD_SYN_ATT); + + while (ud_disassemble(&ud_obj)) { + +#ifdef PIPE_ARCH_X86 + debug_printf("%08lx: ", (unsigned long)ud_insn_off(&ud_obj)); +#endif +#ifdef PIPE_ARCH_X86_64 + debug_printf("%016llx: ", (unsigned long long)ud_insn_off(&ud_obj)); +#endif + +#if 0 + debug_printf("%-16s ", ud_insn_hex(&ud_obj)); +#endif + + debug_printf("%s\n", ud_insn_asm(&ud_obj)); + + if(ud_obj.mnemonic != UD_Icall) { + unsigned i; + for(i = 0; i < 3; ++i) { + const struct ud_operand *op = &ud_obj.operand[i]; + if (op->type == UD_OP_JIMM){ + uint64_t pc = ud_obj.pc; + + switch (op->size) { + case 8: + pc += op->lval.sbyte; + break; + case 16: + pc += op->lval.sword; + break; + case 32: + pc += op->lval.sdword; + break; + default: + break; + } + if(pc > max_jmp_pc) + max_jmp_pc = pc; + } + } + } + + if (ud_insn_off(&ud_obj) >= max_jmp_pc && ud_obj.mnemonic == UD_Iret) + break; + } + debug_printf("\n"); +#else + (void)func; +#endif +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.h b/src/gallium/drivers/llvmpipe/lp_bld_debug.h new file mode 100644 index 00000000000..ecdafef76d0 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.h @@ -0,0 +1,60 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef LP_BLD_DEBUG_H +#define LP_BLD_DEBUG_H + + +#include <llvm-c/Core.h> + +#include "pipe/p_compiler.h" +#include "util/u_string.h" + + +static INLINE void +lp_build_name(LLVMValueRef val, const char *format, ...) +{ +#ifdef DEBUG + char name[32]; + va_list ap; + va_start(ap, format); + util_vsnprintf(name, sizeof name, format, ap); + va_end(ap); + LLVMSetValueName(val, name); +#else + (void)val; + (void)format; +#endif +} + + +void +lp_disassemble(const void* func); + + +#endif /* !LP_BLD_DEBUG_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.c b/src/gallium/drivers/llvmpipe/lp_bld_depth.c new file mode 100644 index 00000000000..2cd6e6b9217 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c @@ -0,0 +1,214 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Depth/stencil testing to LLVM IR translation. + * + * To be done accurately/efficiently the depth/stencil test must be done with + * the same type/format of the depth/stencil buffer, which implies massaging + * the incoming depths to fit into place. Using a more straightforward + * type/format for depth/stencil values internally and only convert when + * flushing would avoid this, but it would most likely result in depth fighting + * artifacts. + * + * We are free to use a different pixel layout though. Since our basic + * processing unit is a quad (2x2 pixel block) we store the depth/stencil + * values tiled, a quad at time. That is, a depth buffer containing + * + * Z11 Z12 Z13 Z14 ... + * Z21 Z22 Z23 Z24 ... + * Z31 Z32 Z33 Z34 ... + * Z41 Z42 Z43 Z44 ... + * ... ... ... ... ... + * + * will actually be stored in memory as + * + * Z11 Z12 Z21 Z22 Z13 Z14 Z23 Z24 ... + * Z31 Z32 Z41 Z42 Z33 Z34 Z43 Z44 ... + * ... ... ... ... ... ... ... ... ... + * + * FIXME: Code generate stencil test + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_state.h" +#include "util/u_format.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_logic.h" +#include "lp_bld_flow.h" +#include "lp_bld_debug.h" +#include "lp_bld_depth.h" + + +/** + * Return a type appropriate for depth/stencil testing. + */ +union lp_type +lp_depth_type(const struct util_format_description *format_desc, + unsigned length) +{ + union lp_type type; + unsigned swizzle; + + assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS); + assert(format_desc->block.width == 1); + assert(format_desc->block.height == 1); + + swizzle = format_desc->swizzle[0]; + assert(swizzle < 4); + + type.value = 0; + type.width = format_desc->block.bits; + + if(format_desc->channel[swizzle].type == UTIL_FORMAT_TYPE_FLOAT) { + type.floating = TRUE; + assert(swizzle = 0); + assert(format_desc->channel[swizzle].size == format_desc->block.bits); + } + else if(format_desc->channel[swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED) { + assert(format_desc->block.bits <= 32); + if(format_desc->channel[swizzle].normalized) + type.norm = TRUE; + } + else + assert(0); + + assert(type.width <= length); + type.length = length / type.width; + + return type; +} + + +/** + * Depth test. + */ +void +lp_build_depth_test(LLVMBuilderRef builder, + const struct pipe_depth_state *state, + union lp_type type, + const struct util_format_description *format_desc, + struct lp_build_mask_context *mask, + LLVMValueRef src, + LLVMValueRef dst_ptr) +{ + struct lp_build_context bld; + unsigned z_swizzle; + LLVMValueRef dst; + LLVMValueRef z_bitmask = NULL; + LLVMValueRef test; + + if(!state->enabled) + return; + + assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS); + assert(format_desc->block.width == 1); + assert(format_desc->block.height == 1); + + z_swizzle = format_desc->swizzle[0]; + if(z_swizzle == UTIL_FORMAT_SWIZZLE_NONE) + return; + + /* Sanity checking */ + assert(z_swizzle < 4); + assert(format_desc->block.bits == type.width); + if(type.floating) { + assert(z_swizzle == 0); + assert(format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_FLOAT); + assert(format_desc->channel[z_swizzle].size == format_desc->block.bits); + } + else { + assert(format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED); + assert(format_desc->channel[z_swizzle].normalized); + assert(!type.fixed); + assert(!type.sign); + assert(type.norm); + } + + /* Setup build context */ + lp_build_context_init(&bld, builder, type); + + dst = LLVMBuildLoad(builder, dst_ptr, ""); + + lp_build_name(dst, "zsbuf"); + + /* Align the source depth bits with the destination's, and mask out any + * stencil or padding bits from both */ + if(format_desc->channel[z_swizzle].size == format_desc->block.bits) { + assert(z_swizzle == 0); + /* nothing to do */ + } + else { + unsigned padding_left; + unsigned padding_right; + unsigned chan; + + assert(format_desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED); + assert(format_desc->channel[z_swizzle].size <= format_desc->block.bits); + assert(format_desc->channel[z_swizzle].normalized); + + padding_right = 0; + for(chan = 0; chan < z_swizzle; ++chan) + padding_right += format_desc->channel[chan].size; + padding_left = format_desc->block.bits - format_desc->channel[z_swizzle].size; + + if(padding_left || padding_right) { + const long long mask_left = ((long long)1 << (format_desc->block.bits - padding_left)) - 1; + const long long mask_right = ((long long)1 << (padding_right)) - 1; + z_bitmask = lp_build_int_const_scalar(type, mask_left & mask_right); + } + + if(padding_left) + src = LLVMBuildLShr(builder, src, lp_build_int_const_scalar(type, padding_left), ""); + if(padding_right) + src = LLVMBuildAnd(builder, src, z_bitmask, ""); + if(padding_left || padding_right) + dst = LLVMBuildAnd(builder, dst, z_bitmask, ""); + } + + lp_build_name(dst, "zsbuf.z"); + + test = lp_build_cmp(&bld, state->func, src, dst); + lp_build_mask_update(mask, test); + + if(state->writemask) { + if(z_bitmask) + z_bitmask = LLVMBuildAnd(builder, mask->value, z_bitmask, ""); + else + z_bitmask = mask->value; + + dst = lp_build_select(&bld, z_bitmask, src, dst); + LLVMBuildStore(builder, dst, dst_ptr); + } + + assert(!state->occlusion_count); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.h b/src/gallium/drivers/llvmpipe/lp_bld_depth.h new file mode 100644 index 00000000000..5d2e042fcc5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.h @@ -0,0 +1,63 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * Depth/stencil testing to LLVM IR translation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_DEPTH_H +#define LP_BLD_DEPTH_H + + +#include <llvm-c/Core.h> + + +struct pipe_depth_state; +struct util_format_description; +union lp_type; +struct lp_build_mask_context; + + +union lp_type +lp_depth_type(const struct util_format_description *format_desc, + unsigned length); + + +void +lp_build_depth_test(LLVMBuilderRef builder, + const struct pipe_depth_state *state, + union lp_type type, + const struct util_format_description *format_desc, + struct lp_build_mask_context *mask, + LLVMValueRef src, + LLVMValueRef dst_ptr); + + +#endif /* !LP_BLD_DEPTH_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.c b/src/gallium/drivers/llvmpipe/lp_bld_flow.c new file mode 100644 index 00000000000..9d99e1a9d9f --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.c @@ -0,0 +1,173 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * LLVM control flow build helpers. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "util/u_debug.h" + +#include "lp_bld_type.h" +#include "lp_bld_flow.h" + + +void +lp_build_mask_begin(struct lp_build_mask_context *mask, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef value) +{ + memset(mask, 0, sizeof *mask); + + mask->builder = builder; + mask->reg_type = LLVMIntType(type.width * type.length); + mask->value = value; +} + + +void +lp_build_mask_update(struct lp_build_mask_context *mask, + LLVMValueRef value) +{ + + LLVMValueRef cond; + LLVMBasicBlockRef current_block; + LLVMBasicBlockRef next_block; + LLVMBasicBlockRef new_block; + + if(mask->value) + mask->value = LLVMBuildAnd(mask->builder, mask->value, value, ""); + else + mask->value = value; + + /* FIXME: disabled until we have proper control flow helpers */ +#if 0 + cond = LLVMBuildICmp(mask->builder, + LLVMIntEQ, + LLVMBuildBitCast(mask->builder, mask->value, mask->reg_type, ""), + LLVMConstNull(mask->reg_type), + ""); + + current_block = LLVMGetInsertBlock(mask->builder); + + if(!mask->skip_block) { + LLVMValueRef function = LLVMGetBasicBlockParent(current_block); + mask->skip_block = LLVMAppendBasicBlock(function, "skip"); + + mask->phi = LLVMBuildPhi(mask->builder, LLVMTypeOf(mask->value), ""); + } + + next_block = LLVMGetNextBasicBlock(current_block); + assert(next_block); + if(next_block) { + new_block = LLVMInsertBasicBlock(next_block, ""); + } + else { + LLVMValueRef function = LLVMGetBasicBlockParent(current_block); + new_block = LLVMAppendBasicBlock(function, ""); + } + + LLVMAddIncoming(mask->phi, &mask->value, ¤t_block, 1); + LLVMBuildCondBr(mask->builder, cond, mask->skip_block, new_block); + + LLVMPositionBuilderAtEnd(mask->builder, new_block); +#endif +} + + +LLVMValueRef +lp_build_mask_end(struct lp_build_mask_context *mask) +{ + if(mask->skip_block) { + LLVMBasicBlockRef current_block = LLVMGetInsertBlock(mask->builder); + + LLVMAddIncoming(mask->phi, &mask->value, ¤t_block, 1); + LLVMBuildBr(mask->builder, mask->skip_block); + + LLVMPositionBuilderAtEnd(mask->builder, mask->skip_block); + + mask->value = mask->phi; + mask->phi = NULL; + mask->skip_block = NULL; + } + + return mask->value; +} + + + +void +lp_build_loop_begin(LLVMBuilderRef builder, + LLVMValueRef start, + struct lp_build_loop_state *state) +{ + LLVMBasicBlockRef block = LLVMGetInsertBlock(builder); + LLVMValueRef function = LLVMGetBasicBlockParent(block); + + state->block = LLVMAppendBasicBlock(function, "loop"); + + LLVMBuildBr(builder, state->block); + + LLVMPositionBuilderAtEnd(builder, state->block); + + state->counter = LLVMBuildPhi(builder, LLVMTypeOf(start), ""); + + LLVMAddIncoming(state->counter, &start, &block, 1); + +} + + +void +lp_build_loop_end(LLVMBuilderRef builder, + LLVMValueRef end, + LLVMValueRef step, + struct lp_build_loop_state *state) +{ + LLVMBasicBlockRef block = LLVMGetInsertBlock(builder); + LLVMValueRef function = LLVMGetBasicBlockParent(block); + LLVMValueRef next; + LLVMValueRef cond; + LLVMBasicBlockRef after_block; + + if (!step) + step = LLVMConstInt(LLVMTypeOf(end), 1, 0); + + next = LLVMBuildAdd(builder, state->counter, step, ""); + + cond = LLVMBuildICmp(builder, LLVMIntNE, next, end, ""); + + after_block = LLVMAppendBasicBlock(function, ""); + + LLVMBuildCondBr(builder, cond, after_block, state->block); + + LLVMAddIncoming(state->counter, &next, &block, 1); + + LLVMPositionBuilderAtEnd(builder, after_block); +} + diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.h b/src/gallium/drivers/llvmpipe/lp_bld_flow.h new file mode 100644 index 00000000000..1b634ff038d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.h @@ -0,0 +1,103 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * LLVM control flow build helpers. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_FLOW_H +#define LP_BLD_FLOW_H + + +#include <llvm-c/Core.h> + + +union lp_type; + + +struct lp_build_mask_context +{ + LLVMBuilderRef builder; + + LLVMTypeRef reg_type; + + LLVMValueRef value; + + LLVMValueRef phi; + + LLVMBasicBlockRef skip_block; +}; + + +void +lp_build_mask_begin(struct lp_build_mask_context *mask, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef value); + +/** + * Bitwise AND the mask with the given value, if a previous mask was set. + */ +void +lp_build_mask_update(struct lp_build_mask_context *mask, + LLVMValueRef value); + +LLVMValueRef +lp_build_mask_end(struct lp_build_mask_context *mask); + + +/** + * LLVM's IR doesn't represent for-loops directly. Furthermore it + * it requires creating code blocks, branches, phi variables, so it + * requires a fair amount of code. + * + * @sa http://www.llvm.org/docs/tutorial/LangImpl5.html#for + */ +struct lp_build_loop_state +{ + LLVMBasicBlockRef block; + LLVMValueRef counter; +}; + + +void +lp_build_loop_begin(LLVMBuilderRef builder, + LLVMValueRef start, + struct lp_build_loop_state *state); + + +void +lp_build_loop_end(LLVMBuilderRef builder, + LLVMValueRef end, + LLVMValueRef step, + struct lp_build_loop_state *state); + + + +#endif /* !LP_BLD_FLOW_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format.h b/src/gallium/drivers/llvmpipe/lp_bld_format.h new file mode 100644 index 00000000000..01c8a752d18 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_format.h @@ -0,0 +1,101 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_BLD_H +#define LP_BLD_H + + +/** + * @file + * LLVM IR building helpers interfaces. + * + * We use LLVM-C bindings for now. They are not documented, but follow the C++ + * interfaces very closely, and appear to be complete enough for code + * genration. See + * http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html + * for a standalone example. + */ + +#include <llvm-c/Core.h> + +#include "pipe/p_format.h" + + +union lp_type; + + +/** + * Unpack a pixel into its RGBA components. + * + * @param packed integer. + * + * @return RGBA in a 4 floats vector. + */ +LLVMValueRef +lp_build_unpack_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef packed); + + +/** + * Pack a pixel. + * + * @param rgba 4 float vector with the unpacked components. + */ +LLVMValueRef +lp_build_pack_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef rgba); + + +/** + * Load a pixel into its RGBA components. + * + * @param ptr value with the pointer to the packed pixel. Pointer type is + * irrelevant. + * + * @return RGBA in a 4 floats vector. + */ +LLVMValueRef +lp_build_load_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef ptr); + + +/** + * Store a pixel. + * + * @param rgba 4 float vector with the unpacked components. + */ +void +lp_build_store_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef ptr, + LLVMValueRef rgba); + + +#endif /* !LP_BLD_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c new file mode 100644 index 00000000000..dcbc0076c7d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c @@ -0,0 +1,303 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "util/u_format.h" + +#include "lp_bld_format.h" + + +LLVMValueRef +lp_build_unpack_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef packed) +{ + const struct util_format_description *desc; + LLVMTypeRef type; + LLVMValueRef shifted, casted, scaled, masked; + LLVMValueRef shifts[4]; + LLVMValueRef masks[4]; + LLVMValueRef scales[4]; + LLVMValueRef swizzles[4]; + LLVMValueRef aux[4]; + bool normalized; + int empty_channel; + unsigned shift; + unsigned i; + + desc = util_format_description(format); + + /* FIXME: Support more formats */ + assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(desc->block.width == 1); + assert(desc->block.height == 1); + assert(desc->block.bits <= 32); + + type = LLVMIntType(desc->block.bits); + + /* Do the intermediate integer computations with 32bit integers since it + * matches floating point size */ + if (desc->block.bits < 32) + packed = LLVMBuildZExt(builder, packed, LLVMInt32Type(), ""); + + /* Broadcast the packed value to all four channels */ + packed = LLVMBuildInsertElement(builder, + LLVMGetUndef(LLVMVectorType(LLVMInt32Type(), 4)), + packed, + LLVMConstNull(LLVMInt32Type()), + ""); + packed = LLVMBuildShuffleVector(builder, + packed, + LLVMGetUndef(LLVMVectorType(LLVMInt32Type(), 4)), + LLVMConstNull(LLVMVectorType(LLVMInt32Type(), 4)), + ""); + + /* Initialize vector constants */ + normalized = FALSE; + empty_channel = -1; + shift = 0; + for (i = 0; i < 4; ++i) { + unsigned bits = desc->channel[i].size; + + if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) { + shifts[i] = LLVMGetUndef(LLVMInt32Type()); + masks[i] = LLVMConstNull(LLVMInt32Type()); + scales[i] = LLVMConstNull(LLVMFloatType()); + empty_channel = i; + } + else { + unsigned mask = (1 << bits) - 1; + + assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED); + assert(bits < 32); + + shifts[i] = LLVMConstInt(LLVMInt32Type(), shift, 0); + masks[i] = LLVMConstInt(LLVMInt32Type(), mask, 0); + + if (desc->channel[i].normalized) { + scales[i] = LLVMConstReal(LLVMFloatType(), 1.0/mask); + normalized = TRUE; + } + else + scales[i] = LLVMConstReal(LLVMFloatType(), 1.0); + } + + shift += bits; + } + + shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), ""); + masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), ""); + // UIToFP can't be expressed in SSE2 + casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatType(), 4), ""); + + if (normalized) + scaled = LLVMBuildMul(builder, casted, LLVMConstVector(scales, 4), ""); + else + scaled = casted; + + for (i = 0; i < 4; ++i) + aux[i] = LLVMGetUndef(LLVMFloatType()); + + for (i = 0; i < 4; ++i) { + enum util_format_swizzle swizzle = desc->swizzle[i]; + + switch (swizzle) { + case UTIL_FORMAT_SWIZZLE_X: + case UTIL_FORMAT_SWIZZLE_Y: + case UTIL_FORMAT_SWIZZLE_Z: + case UTIL_FORMAT_SWIZZLE_W: + swizzles[i] = LLVMConstInt(LLVMInt32Type(), swizzle, 0); + break; + case UTIL_FORMAT_SWIZZLE_0: + assert(empty_channel >= 0); + swizzles[i] = LLVMConstInt(LLVMInt32Type(), empty_channel, 0); + break; + case UTIL_FORMAT_SWIZZLE_1: + swizzles[i] = LLVMConstInt(LLVMInt32Type(), 4, 0); + aux[0] = LLVMConstReal(LLVMFloatType(), 1.0); + break; + case UTIL_FORMAT_SWIZZLE_NONE: + swizzles[i] = LLVMGetUndef(LLVMFloatType()); + assert(0); + break; + } + } + + return LLVMBuildShuffleVector(builder, scaled, LLVMConstVector(aux, 4), LLVMConstVector(swizzles, 4), ""); +} + + +LLVMValueRef +lp_build_pack_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef rgba) +{ + const struct util_format_description *desc; + LLVMTypeRef type; + LLVMValueRef packed = NULL; + LLVMValueRef swizzles[4]; + LLVMValueRef shifted, casted, scaled, unswizzled; + LLVMValueRef shifts[4]; + LLVMValueRef scales[4]; + bool normalized; + unsigned shift; + unsigned i, j; + + desc = util_format_description(format); + + assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(desc->block.width == 1); + assert(desc->block.height == 1); + + type = LLVMIntType(desc->block.bits); + + /* Unswizzle the color components into the source vector. */ + for (i = 0; i < 4; ++i) { + for (j = 0; j < 4; ++j) { + if (desc->swizzle[j] == i) + break; + } + if (j < 4) + swizzles[i] = LLVMConstInt(LLVMInt32Type(), j, 0); + else + swizzles[i] = LLVMGetUndef(LLVMInt32Type()); + } + + unswizzled = LLVMBuildShuffleVector(builder, rgba, + LLVMGetUndef(LLVMVectorType(LLVMFloatType(), 4)), + LLVMConstVector(swizzles, 4), ""); + + normalized = FALSE; + shift = 0; + for (i = 0; i < 4; ++i) { + unsigned bits = desc->channel[i].size; + + if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) { + shifts[i] = LLVMGetUndef(LLVMInt32Type()); + scales[i] = LLVMGetUndef(LLVMFloatType()); + } + else { + unsigned mask = (1 << bits) - 1; + + assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED); + assert(bits < 32); + + shifts[i] = LLVMConstInt(LLVMInt32Type(), shift, 0); + + if (desc->channel[i].normalized) { + scales[i] = LLVMConstReal(LLVMFloatType(), mask); + normalized = TRUE; + } + else + scales[i] = LLVMConstReal(LLVMFloatType(), 1.0); + } + + shift += bits; + } + + if (normalized) + scaled = LLVMBuildMul(builder, unswizzled, LLVMConstVector(scales, 4), ""); + else + scaled = unswizzled; + + casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32Type(), 4), ""); + + shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), ""); + + /* Bitwise or all components */ + for (i = 0; i < 4; ++i) { + if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) { + LLVMValueRef component = LLVMBuildExtractElement(builder, shifted, LLVMConstInt(LLVMInt32Type(), i, 0), ""); + if (packed) + packed = LLVMBuildOr(builder, packed, component, ""); + else + packed = component; + } + } + + if (!packed) + packed = LLVMGetUndef(LLVMInt32Type()); + + if (desc->block.bits < 32) + packed = LLVMBuildTrunc(builder, packed, type, ""); + + return packed; +} + + +LLVMValueRef +lp_build_load_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef ptr) +{ + const struct util_format_description *desc; + LLVMTypeRef type; + LLVMValueRef packed; + + desc = util_format_description(format); + + /* FIXME: Support more formats */ + assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(desc->block.width == 1); + assert(desc->block.height == 1); + assert(desc->block.bits <= 32); + + type = LLVMIntType(desc->block.bits); + + ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, 0), ""); + + packed = LLVMBuildLoad(builder, ptr, ""); + + return lp_build_unpack_rgba(builder, format, packed); +} + + +void +lp_build_store_rgba(LLVMBuilderRef builder, + enum pipe_format format, + LLVMValueRef ptr, + LLVMValueRef rgba) +{ + const struct util_format_description *desc; + LLVMTypeRef type; + LLVMValueRef packed; + + desc = util_format_description(format); + + assert(desc->layout == UTIL_FORMAT_LAYOUT_ARITH); + assert(desc->block.width == 1); + assert(desc->block.height == 1); + + type = LLVMIntType(desc->block.bits); + + packed = lp_build_pack_rgba(builder, format, rgba); + + ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, 0), ""); + + LLVMBuildStore(builder, packed, ptr); +} + diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c new file mode 100644 index 00000000000..cfe20a0d75b --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c @@ -0,0 +1,377 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Position and shader input interpolation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_shader_tokens.h" +#include "util/u_debug.h" +#include "util/u_memory.h" +#include "util/u_math.h" +#include "tgsi/tgsi_parse.h" +#include "lp_bld_debug.h" +#include "lp_bld_const.h" +#include "lp_bld_arit.h" +#include "lp_bld_swizzle.h" +#include "lp_bld_interp.h" + + +static void +attrib_name(LLVMValueRef val, unsigned attrib, unsigned chan, const char *suffix) +{ + if(attrib == 0) + lp_build_name(val, "pos.%c%s", "xyzw"[chan], suffix); + else + lp_build_name(val, "input%u.%c%s", attrib - 1, "xyzw"[chan], suffix); +} + + +static void +coeffs_init(struct lp_build_interp_soa_context *bld, + LLVMValueRef a0_ptr, + LLVMValueRef dadx_ptr, + LLVMValueRef dady_ptr) +{ + LLVMBuilderRef builder = bld->base.builder; + unsigned attrib; + unsigned chan; + + for(attrib = 0; attrib < bld->num_attribs; ++attrib) { + unsigned mask = bld->mask[attrib]; + unsigned mode = bld->mode[attrib]; + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + if(mask & (1 << chan)) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), attrib*NUM_CHANNELS + chan, 0); + LLVMValueRef a0 = NULL; + LLVMValueRef dadx = NULL; + LLVMValueRef dady = NULL; + + switch( mode ) { + case TGSI_INTERPOLATE_PERSPECTIVE: + /* fall-through */ + + case TGSI_INTERPOLATE_LINEAR: + dadx = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dadx_ptr, &index, 1, ""), ""); + dady = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dady_ptr, &index, 1, ""), ""); + dadx = lp_build_broadcast_scalar(&bld->base, dadx); + dady = lp_build_broadcast_scalar(&bld->base, dady); + attrib_name(dadx, attrib, chan, ".dadx"); + attrib_name(dady, attrib, chan, ".dady"); + /* fall-through */ + + case TGSI_INTERPOLATE_CONSTANT: + a0 = LLVMBuildLoad(builder, LLVMBuildGEP(builder, a0_ptr, &index, 1, ""), ""); + a0 = lp_build_broadcast_scalar(&bld->base, a0); + attrib_name(a0, attrib, chan, ".dady"); + break; + + default: + assert(0); + break; + } + + bld->a0 [attrib][chan] = a0; + bld->dadx[attrib][chan] = dadx; + bld->dady[attrib][chan] = dady; + } + } + } +} + + +/** + * Small vector x scale multiplication optimization. + * + * TODO: Should be elsewhere. + */ +static LLVMValueRef +coeff_multiply(struct lp_build_interp_soa_context *bld, + LLVMValueRef coeff, + int step) +{ + LLVMValueRef factor; + + switch(step) { + case 0: + return bld->base.zero; + case 1: + return coeff; + case 2: + return lp_build_add(&bld->base, coeff, coeff); + default: + factor = lp_build_const_scalar(bld->base.type, (double)step); + return lp_build_mul(&bld->base, coeff, factor); + } +} + + +/** + * Multiply the dadx and dady with the xstep and ystep respectively. + */ +static void +coeffs_update(struct lp_build_interp_soa_context *bld) +{ + unsigned attrib; + unsigned chan; + + for(attrib = 0; attrib < bld->num_attribs; ++attrib) { + unsigned mask = bld->mask[attrib]; + unsigned mode = bld->mode[attrib]; + if (mode != TGSI_INTERPOLATE_CONSTANT) { + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + if(mask & (1 << chan)) { + bld->dadx[attrib][chan] = coeff_multiply(bld, bld->dadx[attrib][chan], bld->xstep); + bld->dady[attrib][chan] = coeff_multiply(bld, bld->dady[attrib][chan], bld->ystep); + } + } + } + } +} + + +static void +attribs_init(struct lp_build_interp_soa_context *bld) +{ + LLVMValueRef x = bld->pos[0]; + LLVMValueRef y = bld->pos[1]; + LLVMValueRef oow = NULL; + unsigned attrib; + unsigned chan; + + for(attrib = 0; attrib < bld->num_attribs; ++attrib) { + unsigned mask = bld->mask[attrib]; + unsigned mode = bld->mode[attrib]; + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + if(mask & (1 << chan)) { + LLVMValueRef a0 = bld->a0 [attrib][chan]; + LLVMValueRef dadx = bld->dadx[attrib][chan]; + LLVMValueRef dady = bld->dady[attrib][chan]; + LLVMValueRef res; + + res = a0; + + if (mode != TGSI_INTERPOLATE_CONSTANT) { + res = lp_build_add(&bld->base, res, lp_build_mul(&bld->base, x, dadx)); + res = lp_build_add(&bld->base, res, lp_build_mul(&bld->base, y, dady)); + } + + /* Keep the value of the attribue before perspective divide for faster updates */ + bld->attribs_pre[attrib][chan] = res; + + if (mode == TGSI_INTERPOLATE_PERSPECTIVE) { + LLVMValueRef w = bld->pos[3]; + assert(attrib != 0); + if(!oow) + oow = lp_build_rcp(&bld->base, w); + res = lp_build_mul(&bld->base, res, oow); + } + + attrib_name(res, attrib, chan, ""); + + bld->attribs[attrib][chan] = res; + } + } + } +} + + +static void +attribs_update(struct lp_build_interp_soa_context *bld) +{ + LLVMValueRef oow = NULL; + unsigned attrib; + unsigned chan; + + for(attrib = 0; attrib < bld->num_attribs; ++attrib) { + unsigned mask = bld->mask[attrib]; + unsigned mode = bld->mode[attrib]; + + if (mode != TGSI_INTERPOLATE_CONSTANT) { + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + if(mask & (1 << chan)) { + LLVMValueRef dadx = bld->dadx[attrib][chan]; + LLVMValueRef dady = bld->dady[attrib][chan]; + LLVMValueRef res; + + res = bld->attribs_pre[attrib][chan]; + + if(bld->xstep) + res = lp_build_add(&bld->base, res, dadx); + + if(bld->ystep) + res = lp_build_add(&bld->base, res, dady); + + bld->attribs_pre[attrib][chan] = res; + + if (mode == TGSI_INTERPOLATE_PERSPECTIVE) { + LLVMValueRef w = bld->pos[3]; + assert(attrib != 0); + if(!oow) + oow = lp_build_rcp(&bld->base, w); + res = lp_build_mul(&bld->base, res, oow); + } + + attrib_name(res, attrib, chan, ""); + + bld->attribs[attrib][chan] = res; + } + } + } + } +} + + +/** + * Generate the position vectors. + * + * Parameter x0, y0 are the integer values with the quad upper left coordinates. + */ +static void +pos_init(struct lp_build_interp_soa_context *bld, + LLVMValueRef x0, + LLVMValueRef y0) +{ + lp_build_name(x0, "pos.x"); + lp_build_name(y0, "pos.y"); + + bld->attribs[0][0] = x0; + bld->attribs[0][1] = y0; +} + + +static void +pos_update(struct lp_build_interp_soa_context *bld) +{ + LLVMValueRef x = bld->attribs[0][0]; + LLVMValueRef y = bld->attribs[0][1]; + + if(bld->xstep) + x = lp_build_add(&bld->base, x, lp_build_const_scalar(bld->base.type, bld->xstep)); + + if(bld->ystep) + y = lp_build_add(&bld->base, y, lp_build_const_scalar(bld->base.type, bld->ystep)); + + lp_build_name(x, "pos.x"); + lp_build_name(y, "pos.y"); + + bld->attribs[0][0] = x; + bld->attribs[0][1] = y; +} + + +void +lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld, + const struct tgsi_token *tokens, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef a0_ptr, + LLVMValueRef dadx_ptr, + LLVMValueRef dady_ptr, + LLVMValueRef x0, + LLVMValueRef y0, + int xstep, + int ystep) +{ + struct tgsi_parse_context parse; + struct tgsi_full_declaration *decl; + + memset(bld, 0, sizeof *bld); + + lp_build_context_init(&bld->base, builder, type); + + /* For convenience */ + bld->pos = bld->attribs[0]; + bld->inputs = (const LLVMValueRef (*)[NUM_CHANNELS]) bld->attribs[1]; + + /* Position */ + bld->num_attribs = 1; + bld->mask[0] = TGSI_WRITEMASK_ZW; + bld->mode[0] = TGSI_INTERPOLATE_LINEAR; + + /* Inputs */ + tgsi_parse_init( &parse, tokens ); + while( !tgsi_parse_end_of_tokens( &parse ) ) { + tgsi_parse_token( &parse ); + + switch( parse.FullToken.Token.Type ) { + case TGSI_TOKEN_TYPE_DECLARATION: + decl = &parse.FullToken.FullDeclaration; + if( decl->Declaration.File == TGSI_FILE_INPUT ) { + unsigned first, last, mask; + unsigned attrib; + + first = decl->DeclarationRange.First; + last = decl->DeclarationRange.Last; + mask = decl->Declaration.UsageMask; + + for( attrib = first; attrib <= last; ++attrib ) { + bld->mask[1 + attrib] = mask; + bld->mode[1 + attrib] = decl->Declaration.Interpolate; + } + + bld->num_attribs = MAX2(bld->num_attribs, 1 + last + 1); + } + break; + + case TGSI_TOKEN_TYPE_INSTRUCTION: + case TGSI_TOKEN_TYPE_IMMEDIATE: + break; + + default: + assert( 0 ); + } + } + tgsi_parse_free( &parse ); + + coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr); + + pos_init(bld, x0, y0); + + attribs_init(bld); + + bld->xstep = xstep; + bld->ystep = ystep; + + coeffs_update(bld); +} + + +/** + * Advance the position and inputs with the xstep and ystep. + */ +void +lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld) +{ + pos_update(bld); + + attribs_update(bld); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.h b/src/gallium/drivers/llvmpipe/lp_bld_interp.h new file mode 100644 index 00000000000..9194f6233a7 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.h @@ -0,0 +1,99 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Position and shader input interpolation. + * + * Special attention is given to the interpolation of side by side quads. + * Multiplications are made only for the first quad. Interpolation of + * inputs for posterior quads are done exclusively with additions, and + * perspective divide if necessary. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_INTERP_H +#define LP_BLD_INTERP_H + + +#include <llvm-c/Core.h> + +#include "tgsi/tgsi_exec.h" + +#include "lp_bld_type.h" + + +struct tgsi_token; + + +struct lp_build_interp_soa_context +{ + struct lp_build_context base; + + unsigned num_attribs; + unsigned mask[1 + PIPE_MAX_SHADER_INPUTS]; + unsigned mode[1 + PIPE_MAX_SHADER_INPUTS]; + + LLVMValueRef a0 [1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + LLVMValueRef dadx[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + LLVMValueRef dady[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + + int xstep; + int ystep; + + /* Attribute values before perspective divide */ + LLVMValueRef attribs_pre[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + + LLVMValueRef attribs[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + + /* + * Convenience pointers. Callers may access this one. + */ + const LLVMValueRef *pos; + const LLVMValueRef (*inputs)[NUM_CHANNELS]; +}; + + +void +lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld, + const struct tgsi_token *tokens, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef a0_ptr, + LLVMValueRef dadx_ptr, + LLVMValueRef dady_ptr, + LLVMValueRef x0, + LLVMValueRef y0, + int xstep, + int ystep); + +void +lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld); + + +#endif /* LP_BLD_INTERP_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_intr.c b/src/gallium/drivers/llvmpipe/lp_bld_intr.c new file mode 100644 index 00000000000..9895749d568 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_intr.c @@ -0,0 +1,192 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helpers for emiting intrinsic calls. + * + * LLVM vanilla IR doesn't represent all basic arithmetic operations we care + * about, and it is often necessary to resort target-specific intrinsics for + * performance, convenience. + * + * Ideally we would like to stay away from target specific intrinsics and + * move all the instruction selection logic into upstream LLVM where it belongs. + * + * These functions are also used for calling C functions provided by us from + * generated LLVM code. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_debug.h" + +#include "lp_bld_intr.h" + + +LLVMValueRef +lp_declare_intrinsic(LLVMModuleRef module, + const char *name, + LLVMTypeRef ret_type, + LLVMTypeRef *arg_types, + unsigned num_args) +{ + LLVMTypeRef function_type; + LLVMValueRef function; + + assert(!LLVMGetNamedFunction(module, name)); + + function_type = LLVMFunctionType(ret_type, arg_types, num_args, 0); + function = LLVMAddFunction(module, name, function_type); + + LLVMSetFunctionCallConv(function, LLVMCCallConv); + LLVMSetLinkage(function, LLVMExternalLinkage); + + assert(LLVMIsDeclaration(function)); + + if(name[0] == 'l' && + name[1] == 'l' && + name[2] == 'v' && + name[3] == 'm' && + name[4] == '.') + assert(LLVMGetIntrinsicID(function)); + + return function; +} + + +LLVMValueRef +lp_build_intrinsic(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef *args, + unsigned num_args) +{ + LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder))); + LLVMValueRef function; + + function = LLVMGetNamedFunction(module, name); + if(!function) { + LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS]; + unsigned i; + + assert(num_args <= LP_MAX_FUNC_ARGS); + + for(i = 0; i < num_args; ++i) { + assert(args[i]); + arg_types[i] = LLVMTypeOf(args[i]); + } + + function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args); + } + + return LLVMBuildCall(builder, function, args, num_args, ""); +} + + +LLVMValueRef +lp_build_intrinsic_unary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a) +{ + return lp_build_intrinsic(builder, name, ret_type, &a, 1); +} + + +LLVMValueRef +lp_build_intrinsic_binary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a, + LLVMValueRef b) +{ + LLVMValueRef args[2]; + + args[0] = a; + args[1] = b; + + return lp_build_intrinsic(builder, name, ret_type, args, 2); +} + + +LLVMValueRef +lp_build_intrinsic_map(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef *args, + unsigned num_args) +{ + LLVMTypeRef ret_elem_type = LLVMGetElementType(ret_type); + unsigned n = LLVMGetVectorSize(ret_type); + unsigned i, j; + LLVMValueRef res; + + assert(num_args <= LP_MAX_FUNC_ARGS); + + res = LLVMGetUndef(ret_type); + for(i = 0; i < n; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef arg_elems[LP_MAX_FUNC_ARGS]; + LLVMValueRef res_elem; + for(j = 0; j < num_args; ++j) + arg_elems[j] = LLVMBuildExtractElement(builder, args[j], index, ""); + res_elem = lp_build_intrinsic(builder, name, ret_elem_type, arg_elems, num_args); + res = LLVMBuildInsertElement(builder, res, res_elem, index, ""); + } + + return res; +} + + +LLVMValueRef +lp_build_intrinsic_map_unary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a) +{ + return lp_build_intrinsic_map(builder, name, ret_type, &a, 1); +} + + +LLVMValueRef +lp_build_intrinsic_map_binary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a, + LLVMValueRef b) +{ + LLVMValueRef args[2]; + + args[0] = a; + args[1] = b; + + return lp_build_intrinsic_map(builder, name, ret_type, args, 2); +} + + diff --git a/src/gallium/drivers/llvmpipe/lp_bld_intr.h b/src/gallium/drivers/llvmpipe/lp_bld_intr.h new file mode 100644 index 00000000000..f813f27074b --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_intr.h @@ -0,0 +1,102 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for calling intrinsics. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_INTR_H +#define LP_BLD_INTR_H + + +#include <llvm-c/Core.h> + + +/** + * Max number of arguments in an intrinsic. + */ +#define LP_MAX_FUNC_ARGS 32 + + +LLVMValueRef +lp_declare_intrinsic(LLVMModuleRef module, + const char *name, + LLVMTypeRef ret_type, + LLVMTypeRef *arg_types, + unsigned num_args); + +LLVMValueRef +lp_build_intrinsic(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef *args, + unsigned num_args); + + +LLVMValueRef +lp_build_intrinsic_unary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a); + + +LLVMValueRef +lp_build_intrinsic_binary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a, + LLVMValueRef b); + + +LLVMValueRef +lp_build_intrinsic_map(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef *args, + unsigned num_args); + + +LLVMValueRef +lp_build_intrinsic_map_unary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a); + + +LLVMValueRef +lp_build_intrinsic_map_binary(LLVMBuilderRef builder, + const char *name, + LLVMTypeRef ret_type, + LLVMValueRef a, + LLVMValueRef b); + + +#endif /* !LP_BLD_INTR_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.c b/src/gallium/drivers/llvmpipe/lp_bld_logic.c new file mode 100644 index 00000000000..8631efd6c3e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.c @@ -0,0 +1,397 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for logical operations. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_intr.h" +#include "lp_bld_logic.h" + + +LLVMValueRef +lp_build_cmp(struct lp_build_context *bld, + unsigned func, + LLVMValueRef a, + LLVMValueRef b) +{ + const union lp_type type = bld->type; + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef zeros = LLVMConstNull(int_vec_type); + LLVMValueRef ones = LLVMConstAllOnes(int_vec_type); + LLVMValueRef cond; + LLVMValueRef res; + unsigned i; + + if(func == PIPE_FUNC_NEVER) + return zeros; + if(func == PIPE_FUNC_ALWAYS) + return ones; + + /* TODO: optimize the constant case */ + + /* XXX: It is not clear if we should use the ordered or unordered operators */ + +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + if(type.width * type.length == 128) { + if(type.floating) { + LLVMValueRef args[3]; + unsigned cc; + boolean swap; + + swap = FALSE; + switch(func) { + case PIPE_FUNC_EQUAL: + cc = 0; + break; + case PIPE_FUNC_NOTEQUAL: + cc = 4; + break; + case PIPE_FUNC_LESS: + cc = 1; + break; + case PIPE_FUNC_LEQUAL: + cc = 2; + break; + case PIPE_FUNC_GREATER: + cc = 1; + swap = TRUE; + break; + case PIPE_FUNC_GEQUAL: + cc = 2; + swap = TRUE; + break; + default: + assert(0); + return bld->undef; + } + + if(swap) { + args[0] = b; + args[1] = a; + } + else { + args[0] = a; + args[1] = b; + } + + args[2] = LLVMConstInt(LLVMInt8Type(), cc, 0); + res = lp_build_intrinsic(bld->builder, + "llvm.x86.sse.cmp.ps", + vec_type, + args, 3); + res = LLVMBuildBitCast(bld->builder, res, int_vec_type, ""); + return res; + } + else { + static const struct { + unsigned swap:1; + unsigned eq:1; + unsigned gt:1; + unsigned not:1; + } table[] = { + {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */ + {1, 0, 1, 0}, /* PIPE_FUNC_LESS */ + {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */ + {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */ + {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */ + {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */ + {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */ + {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */ + }; + const char *pcmpeq; + const char *pcmpgt; + LLVMValueRef args[2]; + LLVMValueRef res; + + switch (type.width) { + case 8: + pcmpeq = "llvm.x86.sse2.pcmpeq.b"; + pcmpgt = "llvm.x86.sse2.pcmpgt.b"; + break; + case 16: + pcmpeq = "llvm.x86.sse2.pcmpeq.w"; + pcmpgt = "llvm.x86.sse2.pcmpgt.w"; + break; + case 32: + pcmpeq = "llvm.x86.sse2.pcmpeq.d"; + pcmpgt = "llvm.x86.sse2.pcmpgt.d"; + break; + default: + assert(0); + return bld->undef; + } + + /* There are no signed byte and unsigned word/dword comparison + * instructions. So flip the sign bit so that the results match. + */ + if(table[func].gt && + ((type.width == 8 && type.sign) || + (type.width != 8 && !type.sign))) { + LLVMValueRef msb = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + a = LLVMBuildXor(bld->builder, a, msb, ""); + b = LLVMBuildXor(bld->builder, b, msb, ""); + } + + if(table[func].swap) { + args[0] = b; + args[1] = a; + } + else { + args[0] = a; + args[1] = b; + } + + if(table[func].eq) + res = lp_build_intrinsic(bld->builder, pcmpeq, vec_type, args, 2); + else if (table[func].gt) + res = lp_build_intrinsic(bld->builder, pcmpgt, vec_type, args, 2); + else + res = LLVMConstNull(vec_type); + + if(table[func].not) + res = LLVMBuildNot(bld->builder, res, ""); + + return res; + } + } +#endif + + if(type.floating) { + LLVMRealPredicate op; + switch(func) { + case PIPE_FUNC_NEVER: + op = LLVMRealPredicateFalse; + break; + case PIPE_FUNC_ALWAYS: + op = LLVMRealPredicateTrue; + break; + case PIPE_FUNC_EQUAL: + op = LLVMRealUEQ; + break; + case PIPE_FUNC_NOTEQUAL: + op = LLVMRealUNE; + break; + case PIPE_FUNC_LESS: + op = LLVMRealULT; + break; + case PIPE_FUNC_LEQUAL: + op = LLVMRealULE; + break; + case PIPE_FUNC_GREATER: + op = LLVMRealUGT; + break; + case PIPE_FUNC_GEQUAL: + op = LLVMRealUGE; + break; + default: + assert(0); + return bld->undef; + } + +#if 0 + /* XXX: Although valid IR, no LLVM target currently support this */ + cond = LLVMBuildFCmp(bld->builder, op, a, b, ""); + res = LLVMBuildSelect(bld->builder, cond, ones, zeros, ""); +#else + debug_printf("%s: warning: using slow element-wise vector comparison\n", + __FUNCTION__); + res = LLVMGetUndef(int_vec_type); + for(i = 0; i < type.length; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + cond = LLVMBuildFCmp(bld->builder, op, + LLVMBuildExtractElement(bld->builder, a, index, ""), + LLVMBuildExtractElement(bld->builder, b, index, ""), + ""); + cond = LLVMBuildSelect(bld->builder, cond, + LLVMConstExtractElement(ones, index), + LLVMConstExtractElement(zeros, index), + ""); + res = LLVMBuildInsertElement(bld->builder, res, cond, index, ""); + } +#endif + } + else { + LLVMIntPredicate op; + switch(func) { + case PIPE_FUNC_EQUAL: + op = LLVMIntEQ; + break; + case PIPE_FUNC_NOTEQUAL: + op = LLVMIntNE; + break; + case PIPE_FUNC_LESS: + op = type.sign ? LLVMIntSLT : LLVMIntULT; + break; + case PIPE_FUNC_LEQUAL: + op = type.sign ? LLVMIntSLE : LLVMIntULE; + break; + case PIPE_FUNC_GREATER: + op = type.sign ? LLVMIntSGT : LLVMIntUGT; + break; + case PIPE_FUNC_GEQUAL: + op = type.sign ? LLVMIntSGE : LLVMIntUGE; + break; + default: + assert(0); + return bld->undef; + } + +#if 0 + /* XXX: Although valid IR, no LLVM target currently support this */ + cond = LLVMBuildICmp(bld->builder, op, a, b, ""); + res = LLVMBuildSelect(bld->builder, cond, ones, zeros, ""); +#else + debug_printf("%s: warning: using slow element-wise vector comparison\n", + __FUNCTION__); + res = LLVMGetUndef(int_vec_type); + for(i = 0; i < type.length; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + cond = LLVMBuildICmp(bld->builder, op, + LLVMBuildExtractElement(bld->builder, a, index, ""), + LLVMBuildExtractElement(bld->builder, b, index, ""), + ""); + cond = LLVMBuildSelect(bld->builder, cond, + LLVMConstExtractElement(ones, index), + LLVMConstExtractElement(zeros, index), + ""); + res = LLVMBuildInsertElement(bld->builder, res, cond, index, ""); + } +#endif + } + + return res; +} + + +LLVMValueRef +lp_build_select(struct lp_build_context *bld, + LLVMValueRef mask, + LLVMValueRef a, + LLVMValueRef b) +{ + union lp_type type = bld->type; + LLVMValueRef res; + + if(a == b) + return a; + + if(type.floating) { + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + a = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + b = LLVMBuildBitCast(bld->builder, b, int_vec_type, ""); + } + + /* TODO: On SSE4 we could do this with a single instruction -- PBLENDVB */ + + a = LLVMBuildAnd(bld->builder, a, mask, ""); + + /* This often gets translated to PANDN, but sometimes the NOT is + * pre-computed and stored in another constant. The best strategy depends + * on available registers, so it is not a big deal -- hopefully LLVM does + * the right decision attending the rest of the program. + */ + b = LLVMBuildAnd(bld->builder, b, LLVMBuildNot(bld->builder, mask, ""), ""); + + res = LLVMBuildOr(bld->builder, a, b, ""); + + if(type.floating) { + LLVMTypeRef vec_type = lp_build_vec_type(type); + res = LLVMBuildBitCast(bld->builder, res, vec_type, ""); + } + + return res; +} + + +LLVMValueRef +lp_build_select_aos(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b, + boolean cond[4]) +{ + const union lp_type type = bld->type; + const unsigned n = type.length; + unsigned i, j; + + if(a == b) + return a; + if(cond[0] && cond[1] && cond[2] && cond[3]) + return a; + if(!cond[0] && !cond[1] && !cond[2] && !cond[3]) + return b; + if(a == bld->undef || b == bld->undef) + return bld->undef; + + /* + * There are three major ways of accomplishing this: + * - with a shuffle, + * - with a select, + * - or with a bit mask. + * + * Select isn't supported for vector types yet. + * The flip between these is empirical and might need to be. + */ + if (n <= 4) { + /* + * Shuffle. + */ + LLVMTypeRef elem_type = LLVMInt32Type(); + LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH]; + + for(j = 0; j < n; j += 4) + for(i = 0; i < 4; ++i) + shuffles[j + i] = LLVMConstInt(elem_type, (cond[i] ? 0 : n) + j + i, 0); + + return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), ""); + } +#if 0 + else if(0) { + /* FIXME: Unfortunately select of vectors do not work */ + /* Use a select */ + LLVMTypeRef elem_type = LLVMInt1Type(); + LLVMValueRef cond[LP_MAX_VECTOR_LENGTH]; + + for(j = 0; j < n; j += 4) + for(i = 0; i < 4; ++i) + cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0); + + return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, ""); + } +#endif + else { + LLVMValueRef mask = lp_build_const_mask_aos(type, cond); + return lp_build_select(bld, mask, a, b); + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.h b/src/gallium/drivers/llvmpipe/lp_bld_logic.h new file mode 100644 index 00000000000..29b9e1c45b8 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.h @@ -0,0 +1,72 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for logical operations. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_LOGIC_H +#define LP_BLD_LOGIC_H + + +#include <llvm-c/Core.h> + +#include "pipe/p_defines.h" /* For PIPE_FUNC_xxx */ + + +union lp_type type; +struct lp_build_context; + + +/** + * @param func is one of PIPE_FUNC_xxx + */ +LLVMValueRef +lp_build_cmp(struct lp_build_context *bld, + unsigned func, + LLVMValueRef a, + LLVMValueRef b); + + +LLVMValueRef +lp_build_select(struct lp_build_context *bld, + LLVMValueRef mask, + LLVMValueRef a, + LLVMValueRef b); + +LLVMValueRef +lp_build_select_aos(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b, + boolean cond[4]); + + +#endif /* !LP_BLD_LOGIC_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.c b/src/gallium/drivers/llvmpipe/lp_bld_struct.c new file mode 100644 index 00000000000..14d2b10df9c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.c @@ -0,0 +1,59 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Helper functions for manipulation structures. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_debug.h" +#include "util/u_memory.h" + +#include "lp_bld_debug.h" +#include "lp_bld_struct.h" + + +LLVMValueRef +lp_build_struct_get(LLVMBuilderRef builder, + LLVMValueRef ptr, + unsigned member, + const char *name) +{ + LLVMValueRef indices[2]; + LLVMValueRef member_ptr; + LLVMValueRef res; + indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0); + indices[1] = LLVMConstInt(LLVMInt32Type(), member, 0); + member_ptr = LLVMBuildGEP(builder, ptr, indices, Elements(indices), ""); + res = LLVMBuildLoad(builder, member_ptr, ""); + lp_build_name(res, "%s.%s", LLVMGetValueName(ptr), name); + return res; +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.h b/src/gallium/drivers/llvmpipe/lp_bld_struct.h new file mode 100644 index 00000000000..cbefdc9f815 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.h @@ -0,0 +1,63 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for type conversions. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_STRUCT_H +#define LP_BLD_STRUCT_H + + +#include <llvm-c/Core.h> +#include <llvm-c/Target.h> + +#include "util/u_debug.h" +#include "util/u_memory.h" + + +#define LP_CHECK_STRUCT_SIZE(_ctype, _ltarget, _ltype) \ + assert(LLVMABISizeOfType(_ltarget, _ltype) == \ + sizeof(_ctype)) + +#define LP_CHECK_MEMBER_OFFSET(_ctype, _cmember, _ltarget, _ltype, _lindex) \ + assert(LLVMOffsetOfElement(_ltarget, _ltype, _lindex) == \ + offsetof(_ctype, _cmember)) + + +LLVMValueRef +lp_build_struct_get(LLVMBuilderRef builder, + LLVMValueRef ptr, + unsigned member, + const char *name); + + +#endif /* !LP_BLD_STRUCT_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c new file mode 100644 index 00000000000..ac7eed9379a --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c @@ -0,0 +1,238 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for swizzling/shuffling. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "util/u_debug.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_logic.h" +#include "lp_bld_swizzle.h" + + +LLVMValueRef +lp_build_broadcast(LLVMBuilderRef builder, + LLVMTypeRef vec_type, + LLVMValueRef scalar) +{ + const unsigned n = LLVMGetVectorSize(vec_type); + LLVMValueRef res; + unsigned i; + + res = LLVMGetUndef(vec_type); + for(i = 0; i < n; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + res = LLVMBuildInsertElement(builder, res, scalar, index, ""); + } + + return res; +} + + +LLVMValueRef +lp_build_broadcast_scalar(struct lp_build_context *bld, + LLVMValueRef scalar) +{ + const union lp_type type = bld->type; + LLVMValueRef res; + unsigned i; + + res = bld->undef; + for(i = 0; i < type.length; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + res = LLVMBuildInsertElement(bld->builder, res, scalar, index, ""); + } + + return res; +} + + +LLVMValueRef +lp_build_broadcast_aos(struct lp_build_context *bld, + LLVMValueRef a, + unsigned channel) +{ + const union lp_type type = bld->type; + const unsigned n = type.length; + unsigned i, j; + + if(a == bld->undef || a == bld->zero || a == bld->one) + return a; + + /* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing + * using shuffles here actually causes worst results. More investigation is + * needed. */ + if (n <= 4) { + /* + * Shuffle. + */ + LLVMTypeRef elem_type = LLVMInt32Type(); + LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH]; + + for(j = 0; j < n; j += 4) + for(i = 0; i < 4; ++i) + shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0); + + return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), ""); + } + else { + /* + * Bit mask and recursive shifts + * + * XYZW XYZW .... XYZW <= input + * 0Y00 0Y00 .... 0Y00 + * YY00 YY00 .... YY00 + * YYYY YYYY .... YYYY <= output + */ + union lp_type type4 = type; + const char shifts[4][2] = { + { 1, 2}, + {-1, 2}, + { 1, -2}, + {-1, -2} + }; + boolean cond[4]; + unsigned i; + + memset(cond, 0, sizeof cond); + cond[channel] = 1; + + a = LLVMBuildAnd(bld->builder, a, lp_build_const_mask_aos(type, cond), ""); + + type4.width *= 4; + type4.length /= 4; + + a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), ""); + + for(i = 0; i < 2; ++i) { + LLVMValueRef tmp = NULL; + int shift = shifts[channel][i]; + +#ifdef PIPE_ARCH_LITTLE_ENDIAN + shift = -shift; +#endif + + if(shift > 0) + tmp = LLVMBuildLShr(bld->builder, a, lp_build_int_const_scalar(type4, shift*type.width), ""); + if(shift < 0) + tmp = LLVMBuildShl(bld->builder, a, lp_build_int_const_scalar(type4, -shift*type.width), ""); + + assert(tmp); + if(tmp) + a = LLVMBuildOr(bld->builder, a, tmp, ""); + } + + return LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type), ""); + } +} + + +LLVMValueRef +lp_build_swizzle1_aos(struct lp_build_context *bld, + LLVMValueRef a, + unsigned char swizzle[4]) +{ + const unsigned n = bld->type.length; + unsigned i, j; + + if(a == bld->undef || a == bld->zero || a == bld->one) + return a; + + if(swizzle[0] == swizzle[1] && swizzle[1] == swizzle[2] && swizzle[2] == swizzle[3]) + return lp_build_broadcast_aos(bld, a, swizzle[0]); + + { + /* + * Shuffle. + */ + LLVMTypeRef elem_type = LLVMInt32Type(); + LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH]; + + for(j = 0; j < n; j += 4) + for(i = 0; i < 4; ++i) + shuffles[j + i] = LLVMConstInt(elem_type, j + swizzle[i], 0); + + return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), ""); + } +} + + +LLVMValueRef +lp_build_swizzle2_aos(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b, + unsigned char swizzle[4]) +{ + const unsigned n = bld->type.length; + unsigned i, j; + + if(swizzle[0] < 4 && swizzle[1] < 4 && swizzle[2] < 4 && swizzle[3] < 4) + return lp_build_swizzle1_aos(bld, a, swizzle); + + if(a == b) { + swizzle[0] %= 4; + swizzle[1] %= 4; + swizzle[2] %= 4; + swizzle[3] %= 4; + return lp_build_swizzle1_aos(bld, a, swizzle); + } + + if(swizzle[0] % 4 == 0 && + swizzle[1] % 4 == 1 && + swizzle[2] % 4 == 2 && + swizzle[3] % 4 == 3) { + boolean cond[4]; + cond[0] = swizzle[0] / 4; + cond[1] = swizzle[1] / 4; + cond[2] = swizzle[2] / 4; + cond[3] = swizzle[3] / 4; + return lp_build_select_aos(bld, a, b, cond); + } + + { + /* + * Shuffle. + */ + LLVMTypeRef elem_type = LLVMInt32Type(); + LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH]; + + for(j = 0; j < n; j += 4) + for(i = 0; i < 4; ++i) + shuffles[j + i] = LLVMConstInt(elem_type, j + (swizzle[i] % 4) + (swizzle[i] / 4 * n), 0); + + return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), ""); + } +} + + diff --git a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h new file mode 100644 index 00000000000..d7dd6a8a604 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h @@ -0,0 +1,91 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Helper functions for swizzling/shuffling. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_SWIZZLE_H +#define LP_BLD_SWIZZLE_H + + +#include <llvm-c/Core.h> + + +union lp_type type; +struct lp_build_context; + + +LLVMValueRef +lp_build_broadcast(LLVMBuilderRef builder, + LLVMTypeRef vec_type, + LLVMValueRef scalar); + + +LLVMValueRef +lp_build_broadcast_scalar(struct lp_build_context *bld, + LLVMValueRef scalar); + + +/** + * Broadcast one channel of a vector composed of arrays of XYZW structures into + * all four channel. + */ +LLVMValueRef +lp_build_broadcast_aos(struct lp_build_context *bld, + LLVMValueRef a, + unsigned channel); + + +/** + * Swizzle a vector consisting of an array of XYZW structs. + * + * @param swizzle is the in [0,4[ range. + */ +LLVMValueRef +lp_build_swizzle1_aos(struct lp_build_context *bld, + LLVMValueRef a, + unsigned char swizzle[4]); + + +/** + * Swizzle two vector consisting of an array of XYZW structs. + * + * @param swizzle is the in [0,8[ range. Values in [4,8[ range refer to b. + */ +LLVMValueRef +lp_build_swizzle2_aos(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef b, + unsigned char swizzle[4]); + + +#endif /* !LP_BLD_SWIZZLE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h new file mode 100644 index 00000000000..912db24aecb --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h @@ -0,0 +1,69 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * TGSI to LLVM IR translation. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_BLD_TGSI_H +#define LP_BLD_TGSI_H + +#include <llvm-c/Core.h> + + +struct tgsi_token; +union lp_type; +struct lp_build_context; +struct lp_build_mask_context; + + +typedef void +(*lp_emit_fetch_texel_soa_callback)( LLVMBuilderRef builder, + void *context, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel); + +void +lp_build_tgsi_soa(LLVMBuilderRef builder, + const struct tgsi_token *tokens, + union lp_type type, + struct lp_build_mask_context *mask, + LLVMValueRef consts_ptr, + const LLVMValueRef *pos, + const LLVMValueRef (*inputs)[4], + LLVMValueRef (*outputs)[4], + lp_emit_fetch_texel_soa_callback emit_fetch_texel, + void *emit_fetch_texel_context); + + +#endif /* LP_BLD_TGSI_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c new file mode 100644 index 00000000000..d4d18febec7 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c @@ -0,0 +1,1346 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * TGSI to LLVM IR translation -- SoA. + * + * @author Jose Fonseca <jfonseca@vmware.com> + * + * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell, + * Brian Paul, and others. + */ + +#include "pipe/p_config.h" +#include "pipe/p_shader_tokens.h" +#include "util/u_debug.h" +#include "util/u_math.h" +#include "util/u_memory.h" +#include "tgsi/tgsi_info.h" +#include "tgsi/tgsi_parse.h" +#include "tgsi/tgsi_util.h" +#include "tgsi/tgsi_exec.h" +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_intr.h" +#include "lp_bld_arit.h" +#include "lp_bld_logic.h" +#include "lp_bld_swizzle.h" +#include "lp_bld_flow.h" +#include "lp_bld_tgsi.h" +#include "lp_bld_debug.h" + + +#define LP_MAX_TEMPS 256 +#define LP_MAX_IMMEDIATES 256 + + +#define FOR_EACH_CHANNEL( CHAN )\ + for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++) + +#define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\ + ((INST)->FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN))) + +#define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\ + if (IS_DST0_CHANNEL_ENABLED( INST, CHAN )) + +#define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\ + FOR_EACH_CHANNEL( CHAN )\ + IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN ) + +#define CHAN_X 0 +#define CHAN_Y 1 +#define CHAN_Z 2 +#define CHAN_W 3 + + +struct lp_build_tgsi_soa_context +{ + struct lp_build_context base; + + LLVMValueRef consts_ptr; + const LLVMValueRef *pos; + const LLVMValueRef (*inputs)[NUM_CHANNELS]; + LLVMValueRef (*outputs)[NUM_CHANNELS]; + + lp_emit_fetch_texel_soa_callback emit_fetch_texel; + void *emit_fetch_texel_context; + + LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS]; + LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS]; + + struct lp_build_mask_context *mask; +}; + + +/** + * Register fetch. + */ +static LLVMValueRef +emit_fetch( + struct lp_build_tgsi_soa_context *bld, + const struct tgsi_full_instruction *inst, + unsigned index, + const unsigned chan_index ) +{ + const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[index]; + unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index ); + LLVMValueRef res; + + switch (swizzle) { + case TGSI_EXTSWIZZLE_X: + case TGSI_EXTSWIZZLE_Y: + case TGSI_EXTSWIZZLE_Z: + case TGSI_EXTSWIZZLE_W: + + switch (reg->SrcRegister.File) { + case TGSI_FILE_CONSTANT: { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), reg->SrcRegister.Index*4 + swizzle, 0); + LLVMValueRef scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr, &index, 1, ""); + LLVMValueRef scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, ""); + res = lp_build_broadcast_scalar(&bld->base, scalar); + break; + } + + case TGSI_FILE_IMMEDIATE: + res = bld->immediates[reg->SrcRegister.Index][swizzle]; + assert(res); + break; + + case TGSI_FILE_INPUT: + res = bld->inputs[reg->SrcRegister.Index][swizzle]; + assert(res); + break; + + case TGSI_FILE_TEMPORARY: + res = bld->temps[reg->SrcRegister.Index][swizzle]; + if(!res) + return bld->base.undef; + break; + + default: + assert( 0 ); + return bld->base.undef; + } + break; + + case TGSI_EXTSWIZZLE_ZERO: + res = bld->base.zero; + break; + + case TGSI_EXTSWIZZLE_ONE: + res = bld->base.one; + break; + + default: + assert( 0 ); + return bld->base.undef; + } + + switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) { + case TGSI_UTIL_SIGN_CLEAR: + res = lp_build_abs( &bld->base, res ); + break; + + case TGSI_UTIL_SIGN_SET: + res = lp_build_abs( &bld->base, res ); + res = LLVMBuildNeg( bld->base.builder, res, "" ); + break; + + case TGSI_UTIL_SIGN_TOGGLE: + res = LLVMBuildNeg( bld->base.builder, res, "" ); + break; + + case TGSI_UTIL_SIGN_KEEP: + break; + } + + return res; +} + + +/** + * Register store. + */ +static void +emit_store( + struct lp_build_tgsi_soa_context *bld, + const struct tgsi_full_instruction *inst, + unsigned index, + unsigned chan_index, + LLVMValueRef value) +{ + const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[index]; + + switch( inst->Instruction.Saturate ) { + case TGSI_SAT_NONE: + break; + + case TGSI_SAT_ZERO_ONE: + value = lp_build_max(&bld->base, value, bld->base.zero); + value = lp_build_min(&bld->base, value, bld->base.one); + break; + + case TGSI_SAT_MINUS_PLUS_ONE: + value = lp_build_max(&bld->base, value, lp_build_const_scalar(bld->base.type, -1.0)); + value = lp_build_min(&bld->base, value, bld->base.one); + break; + + default: + assert(0); + } + + switch( reg->DstRegister.File ) { + case TGSI_FILE_OUTPUT: + bld->outputs[reg->DstRegister.Index][chan_index] = value; + break; + + case TGSI_FILE_TEMPORARY: + bld->temps[reg->DstRegister.Index][chan_index] = value; + break; + + case TGSI_FILE_ADDRESS: + /* FIXME */ + assert(0); + break; + + default: + assert( 0 ); + } +} + + +/** + * High-level instruction translators. + */ + +static void +emit_tex( struct lp_build_tgsi_soa_context *bld, + const struct tgsi_full_instruction *inst, + boolean apply_lodbias, + boolean projected) +{ + const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index; + LLVMValueRef lodbias; + LLVMValueRef oow; + LLVMValueRef coords[3]; + LLVMValueRef texel[4]; + unsigned num_coords; + unsigned i; + + switch (inst->InstructionExtTexture.Texture) { + case TGSI_TEXTURE_1D: + num_coords = 1; + break; + case TGSI_TEXTURE_2D: + case TGSI_TEXTURE_RECT: + num_coords = 2; + break; + case TGSI_TEXTURE_SHADOW1D: + case TGSI_TEXTURE_SHADOW2D: + case TGSI_TEXTURE_SHADOWRECT: + case TGSI_TEXTURE_3D: + case TGSI_TEXTURE_CUBE: + num_coords = 3; + break; + default: + assert(0); + return; + } + + if(apply_lodbias) + lodbias = emit_fetch( bld, inst, 0, 3 ); + else + lodbias = bld->base.zero; + + if (projected) { + oow = emit_fetch( bld, inst, 0, 3 ); + oow = lp_build_rcp(&bld->base, oow); + } + + for (i = 0; i < num_coords; i++) { + coords[i] = emit_fetch( bld, inst, 0, i ); + if (projected) + coords[i] = lp_build_mul(&bld->base, coords[i], oow); + } + + bld->emit_fetch_texel(bld->base.builder, bld->emit_fetch_texel_context, + unit, num_coords, coords, lodbias, texel); + + FOR_EACH_DST0_ENABLED_CHANNEL( inst, i ) { + emit_store( bld, inst, 0, i, texel[i] ); + } +} + + +static void +emit_kil( + struct lp_build_tgsi_soa_context *bld, + const struct tgsi_full_instruction *inst ) +{ + const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[0]; + LLVMValueRef terms[NUM_CHANNELS]; + LLVMValueRef mask; + unsigned chan_index; + + memset(&terms, 0, sizeof terms); + + FOR_EACH_CHANNEL( chan_index ) { + unsigned swizzle; + + /* Unswizzle channel */ + swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index ); + + /* Note that we test if the value is less than zero, so 1.0 and 0.0 need + * not to be tested. */ + if(swizzle == TGSI_EXTSWIZZLE_ZERO || swizzle == TGSI_EXTSWIZZLE_ONE) + continue; + + /* Check if the component has not been already tested. */ + assert(swizzle < NUM_CHANNELS); + if( !terms[swizzle] ) + /* TODO: change the comparison operator instead of setting the sign */ + terms[swizzle] = emit_fetch(bld, inst, 0, chan_index ); + } + + mask = NULL; + FOR_EACH_CHANNEL( chan_index ) { + if(terms[chan_index]) { + LLVMValueRef chan_mask; + + chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero); + + if(mask) + mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, ""); + else + mask = chan_mask; + } + } + + if(mask) + lp_build_mask_update(bld->mask, mask); +} + + +static void +emit_kilp( + struct lp_build_tgsi_soa_context *bld ) +{ + /* XXX todo / fix me */ +} + + +/** + * Check if inst src/dest regs use indirect addressing into temporary + * register file. + */ +static boolean +indirect_temp_reference(const struct tgsi_full_instruction *inst) +{ + uint i; + for (i = 0; i < inst->Instruction.NumSrcRegs; i++) { + const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[i]; + if (reg->SrcRegister.File == TGSI_FILE_TEMPORARY && + reg->SrcRegister.Indirect) + return TRUE; + } + for (i = 0; i < inst->Instruction.NumDstRegs; i++) { + const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[i]; + if (reg->DstRegister.File == TGSI_FILE_TEMPORARY && + reg->DstRegister.Indirect) + return TRUE; + } + return FALSE; +} + + +static int +emit_instruction( + struct lp_build_tgsi_soa_context *bld, + struct tgsi_full_instruction *inst ) +{ + unsigned chan_index; + LLVMValueRef src0, src1, src2; + LLVMValueRef tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + LLVMValueRef dst0; + + /* we can't handle indirect addressing into temp register file yet */ + if (indirect_temp_reference(inst)) + return FALSE; + + switch (inst->Instruction.Opcode) { +#if 0 + case TGSI_OPCODE_ARL: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_flr(bld, 0, 0); + emit_f2it( bld, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; +#endif + + case TGSI_OPCODE_MOV: + case TGSI_OPCODE_SWZ: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_LIT: + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) { + emit_store( bld, inst, 0, CHAN_X, bld->base.one); + } + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) { + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + dst0 = lp_build_max( &bld->base, src0, bld->base.zero); + emit_store( bld, inst, 0, CHAN_Y, dst0); + } + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) { + /* XMM[1] = SrcReg[0].yyyy */ + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); + /* XMM[1] = max(XMM[1], 0) */ + tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero); + /* XMM[2] = SrcReg[0].wwww */ + tmp2 = emit_fetch( bld, inst, 0, CHAN_W ); + tmp1 = lp_build_pow( &bld->base, tmp1, tmp2); + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero); + dst0 = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero); + emit_store( bld, inst, 0, CHAN_Z, dst0); + } + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) { + emit_store( bld, inst, 0, CHAN_W, bld->base.one); + } + break; + + case TGSI_OPCODE_RCP: + /* TGSI_OPCODE_RECIP */ + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + dst0 = lp_build_rcp(&bld->base, src0); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, dst0 ); + } + break; + + case TGSI_OPCODE_RSQ: + /* TGSI_OPCODE_RECIPSQRT */ + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + src0 = lp_build_abs(&bld->base, src0); + dst0 = lp_build_rsqrt(&bld->base, src0); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, dst0 ); + } + break; + + case TGSI_OPCODE_EXP: + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) { + LLVMValueRef *p_exp2_int_part = NULL; + LLVMValueRef *p_frac_part = NULL; + LLVMValueRef *p_exp2 = NULL; + + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) + p_exp2_int_part = &tmp0; + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) + p_frac_part = &tmp1; + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) + p_exp2 = &tmp2; + + lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2); + + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) + emit_store( bld, inst, 0, CHAN_X, tmp0); + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) + emit_store( bld, inst, 0, CHAN_Y, tmp1); + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) + emit_store( bld, inst, 0, CHAN_Z, tmp2); + } + /* dst.w = 1.0 */ + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) { + tmp0 = bld->base.one; + emit_store( bld, inst, 0, CHAN_W, tmp0); + } + break; + + case TGSI_OPCODE_LOG: + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) { + LLVMValueRef *p_floor_log2; + LLVMValueRef *p_exp; + LLVMValueRef *p_log2; + + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + src0 = lp_build_abs( &bld->base, src0 ); + + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) + p_floor_log2 = &tmp0; + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) + p_exp = &tmp1; + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) + p_log2 = &tmp2; + + lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2); + + /* dst.x = floor(lg2(abs(src.x))) */ + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X )) + emit_store( bld, inst, 0, CHAN_X, tmp0); + /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */ + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) { + tmp1 = lp_build_div( &bld->base, src0, tmp1); + emit_store( bld, inst, 0, CHAN_Y, tmp1); + } + /* dst.z = lg2(abs(src.x)) */ + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) + emit_store( bld, inst, 0, CHAN_Z, tmp2); + } + /* dst.w = 1.0 */ + if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) { + tmp0 = bld->base.one; + emit_store( bld, inst, 0, CHAN_W, tmp0); + } + break; + + case TGSI_OPCODE_MUL: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + dst0 = lp_build_mul(&bld->base, src0, src1); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_ADD: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + dst0 = lp_build_add(&bld->base, src0, src1); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_DP3: + /* TGSI_OPCODE_DOT3 */ + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Z ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Z ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_DP4: + /* TGSI_OPCODE_DOT4 */ + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Z ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Z ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_W ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_W ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_DST: + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) { + tmp0 = bld->base.one; + emit_store( bld, inst, 0, CHAN_X, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) { + tmp0 = emit_fetch( bld, inst, 0, CHAN_Y ); + tmp1 = emit_fetch( bld, inst, 1, CHAN_Y ); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); + emit_store( bld, inst, 0, CHAN_Y, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { + tmp0 = emit_fetch( bld, inst, 0, CHAN_Z ); + emit_store( bld, inst, 0, CHAN_Z, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { + tmp0 = emit_fetch( bld, inst, 1, CHAN_W ); + emit_store( bld, inst, 0, CHAN_W, tmp0); + } + break; + + case TGSI_OPCODE_MIN: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + dst0 = lp_build_min( &bld->base, src0, src1 ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_MAX: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + dst0 = lp_build_max( &bld->base, src0, src1 ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SLT: + /* TGSI_OPCODE_SETLT */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SGE: + /* TGSI_OPCODE_SETGE */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_MAD: + /* TGSI_OPCODE_MADD */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + tmp1 = emit_fetch( bld, inst, 1, chan_index ); + tmp2 = emit_fetch( bld, inst, 2, chan_index ); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); + tmp0 = lp_build_add( &bld->base, tmp0, tmp2); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_SUB: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + tmp1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_sub( &bld->base, tmp0, tmp1); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_LRP: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + src2 = emit_fetch( bld, inst, 2, chan_index ); + tmp0 = lp_build_sub( &bld->base, src1, src2 ); + tmp0 = lp_build_mul( &bld->base, src0, tmp0 ); + dst0 = lp_build_add( &bld->base, tmp0, src2 ); + emit_store( bld, inst, 0, chan_index, dst0 ); + } + break; + + case TGSI_OPCODE_CND: + return 0; + break; + + case TGSI_OPCODE_DP2A: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); /* xmm0 = src[0].x */ + tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); /* xmm1 = src[1].x */ + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 * xmm1 */ + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); /* xmm1 = src[0].y */ + tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); /* xmm2 = src[1].y */ + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); /* xmm1 = xmm1 * xmm2 */ + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */ + tmp1 = emit_fetch( bld, inst, 2, CHAN_X ); /* xmm1 = src[2].x */ + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); /* dest[ch] = xmm0 */ + } + break; + +#if 0 + case TGSI_OPCODE_FRC: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_frc( bld, 0, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_CLAMP: + return 0; + break; + + case TGSI_OPCODE_FLR: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_flr( bld, 0, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_ROUND: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_rnd( bld, 0, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; +#endif + + case TGSI_OPCODE_EX2: { + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_exp2( &bld->base, tmp0); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + } + + case TGSI_OPCODE_LG2: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_log2( &bld->base, tmp0); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_POW: + src0 = emit_fetch( bld, inst, 0, CHAN_X ); + src1 = emit_fetch( bld, inst, 1, CHAN_X ); + dst0 = lp_build_pow( &bld->base, src0, src1 ); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, dst0 ); + } + break; + + case TGSI_OPCODE_XPD: + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) { + tmp1 = emit_fetch( bld, inst, 1, CHAN_Z ); + tmp3 = emit_fetch( bld, inst, 0, CHAN_Z ); + } + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) { + tmp0 = emit_fetch( bld, inst, 0, CHAN_Y ); + tmp4 = emit_fetch( bld, inst, 1, CHAN_Y ); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) { + tmp2 = tmp0; + tmp2 = lp_build_mul( &bld->base, tmp2, tmp1); + tmp5 = tmp3; + tmp5 = lp_build_mul( &bld->base, tmp5, tmp4); + tmp2 = lp_build_sub( &bld->base, tmp2, tmp5); + emit_store( bld, inst, 0, CHAN_X, tmp2); + } + if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) || + IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) { + tmp2 = emit_fetch( bld, inst, 1, CHAN_X ); + tmp5 = emit_fetch( bld, inst, 0, CHAN_X ); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) { + tmp3 = lp_build_mul( &bld->base, tmp3, tmp2); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp5); + tmp3 = lp_build_sub( &bld->base, tmp3, tmp1); + emit_store( bld, inst, 0, CHAN_Y, tmp3); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { + tmp5 = lp_build_mul( &bld->base, tmp5, tmp4); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp2); + tmp5 = lp_build_sub( &bld->base, tmp5, tmp0); + emit_store( bld, inst, 0, CHAN_Z, tmp5); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { + tmp0 = bld->base.one; + emit_store( bld, inst, 0, CHAN_W, tmp0); + } + break; + + case TGSI_OPCODE_ABS: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + tmp0 = lp_build_abs( &bld->base, tmp0 ) ; + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_RCC: + return 0; + break; + + case TGSI_OPCODE_DPH: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 0, CHAN_Z ); + tmp2 = emit_fetch( bld, inst, 1, CHAN_Z ); + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + tmp1 = emit_fetch( bld, inst, 1, CHAN_W ); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_COS: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_cos( &bld->base, tmp0 ); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_DDX: + return 0; + break; + + case TGSI_OPCODE_DDY: + return 0; + break; + +#if 0 + case TGSI_OPCODE_KILP: + /* predicated kill */ + emit_kilp( bld ); + return 0; /* XXX fix me */ + break; +#endif + + case TGSI_OPCODE_KIL: + /* conditional kill */ + emit_kil( bld, inst ); + break; + + case TGSI_OPCODE_PK2H: + return 0; + break; + + case TGSI_OPCODE_PK2US: + return 0; + break; + + case TGSI_OPCODE_PK4B: + return 0; + break; + + case TGSI_OPCODE_PK4UB: + return 0; + break; + + case TGSI_OPCODE_RFL: + return 0; + break; + + case TGSI_OPCODE_SEQ: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SFL: + return 0; + break; + + case TGSI_OPCODE_SGT: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SIN: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_sin( &bld->base, tmp0 ); + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; + + case TGSI_OPCODE_SLE: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SNE: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 ); + dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero ); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_STR: + return 0; + break; + + case TGSI_OPCODE_TEX: + emit_tex( bld, inst, FALSE, FALSE ); + break; + + case TGSI_OPCODE_TXD: + return 0; + break; + + case TGSI_OPCODE_UP2H: + return 0; + break; + + case TGSI_OPCODE_UP2US: + return 0; + break; + + case TGSI_OPCODE_UP4B: + return 0; + break; + + case TGSI_OPCODE_UP4UB: + return 0; + break; + + case TGSI_OPCODE_X2D: + return 0; + break; + + case TGSI_OPCODE_ARA: + return 0; + break; + +#if 0 + case TGSI_OPCODE_ARR: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_rnd( bld, 0, 0 ); + emit_f2it( bld, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; +#endif + + case TGSI_OPCODE_BRA: + return 0; + break; + + case TGSI_OPCODE_CAL: + return 0; + break; + +#if 0 + case TGSI_OPCODE_RET: + emit_ret( bld ); + break; +#endif + + case TGSI_OPCODE_END: + break; + +#if 0 + case TGSI_OPCODE_SSG: + /* TGSI_OPCODE_SGN */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_sgn( bld, 0, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; +#endif + + case TGSI_OPCODE_CMP: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + src0 = emit_fetch( bld, inst, 0, chan_index ); + src1 = emit_fetch( bld, inst, 1, chan_index ); + src2 = emit_fetch( bld, inst, 2, chan_index ); + tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero ); + dst0 = lp_build_select( &bld->base, tmp0, src1, src2); + emit_store( bld, inst, 0, chan_index, dst0); + } + break; + + case TGSI_OPCODE_SCS: + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) { + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_cos( &bld->base, tmp0 ); + emit_store( bld, inst, 0, CHAN_X, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) { + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); + tmp0 = lp_build_sin( &bld->base, tmp0 ); + emit_store( bld, inst, 0, CHAN_Y, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) { + tmp0 = bld->base.zero; + emit_store( bld, inst, 0, CHAN_Z, tmp0); + } + IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) { + tmp0 = bld->base.one; + emit_store( bld, inst, 0, CHAN_W, tmp0); + } + break; + + case TGSI_OPCODE_TXB: + emit_tex( bld, inst, TRUE, FALSE ); + break; + + case TGSI_OPCODE_NRM: + /* fall-through */ + case TGSI_OPCODE_NRM4: + /* 3 or 4-component normalization */ + { + uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4; + + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) || + IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) || + IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) || + (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) { + + /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */ + + /* xmm4 = src.x */ + /* xmm0 = src.x * src.x */ + tmp0 = emit_fetch(bld, inst, 0, CHAN_X); + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) { + tmp4 = tmp0; + } + tmp0 = lp_build_mul( &bld->base, tmp0, tmp0); + + /* xmm5 = src.y */ + /* xmm0 = xmm0 + src.y * src.y */ + tmp1 = emit_fetch(bld, inst, 0, CHAN_Y); + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) { + tmp5 = tmp1; + } + tmp1 = lp_build_mul( &bld->base, tmp1, tmp1); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + + /* xmm6 = src.z */ + /* xmm0 = xmm0 + src.z * src.z */ + tmp1 = emit_fetch(bld, inst, 0, CHAN_Z); + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) { + tmp6 = tmp1; + } + tmp1 = lp_build_mul( &bld->base, tmp1, tmp1); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + + if (dims == 4) { + /* xmm7 = src.w */ + /* xmm0 = xmm0 + src.w * src.w */ + tmp1 = emit_fetch(bld, inst, 0, CHAN_W); + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) { + tmp7 = tmp1; + } + tmp1 = lp_build_mul( &bld->base, tmp1, tmp1); + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); + } + + /* xmm1 = 1 / sqrt(xmm0) */ + tmp1 = lp_build_rsqrt( &bld->base, tmp0); + + /* dst.x = xmm1 * src.x */ + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) { + tmp4 = lp_build_mul( &bld->base, tmp4, tmp1); + emit_store(bld, inst, 0, CHAN_X, tmp4); + } + + /* dst.y = xmm1 * src.y */ + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) { + tmp5 = lp_build_mul( &bld->base, tmp5, tmp1); + emit_store(bld, inst, 0, CHAN_Y, tmp5); + } + + /* dst.z = xmm1 * src.z */ + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) { + tmp6 = lp_build_mul( &bld->base, tmp6, tmp1); + emit_store(bld, inst, 0, CHAN_Z, tmp6); + } + + /* dst.w = xmm1 * src.w */ + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) { + tmp7 = lp_build_mul( &bld->base, tmp7, tmp1); + emit_store(bld, inst, 0, CHAN_W, tmp7); + } + } + + /* dst0.w = 1.0 */ + if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) { + tmp0 = bld->base.one; + emit_store(bld, inst, 0, CHAN_W, tmp0); + } + } + break; + + case TGSI_OPCODE_DIV: + return 0; + break; + + case TGSI_OPCODE_DP2: + tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); /* xmm0 = src[0].x */ + tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); /* xmm1 = src[1].x */ + tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 * xmm1 */ + tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); /* xmm1 = src[0].y */ + tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); /* xmm2 = src[1].y */ + tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); /* xmm1 = xmm1 * xmm2 */ + tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */ + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + emit_store( bld, inst, 0, chan_index, tmp0); /* dest[ch] = xmm0 */ + } + break; + + case TGSI_OPCODE_TXL: + emit_tex( bld, inst, TRUE, FALSE ); + break; + + case TGSI_OPCODE_TXP: + emit_tex( bld, inst, FALSE, TRUE ); + break; + + case TGSI_OPCODE_BRK: + return 0; + break; + + case TGSI_OPCODE_IF: + return 0; + break; + + case TGSI_OPCODE_BGNFOR: + return 0; + break; + + case TGSI_OPCODE_REP: + return 0; + break; + + case TGSI_OPCODE_ELSE: + return 0; + break; + + case TGSI_OPCODE_ENDIF: + return 0; + break; + + case TGSI_OPCODE_ENDFOR: + return 0; + break; + + case TGSI_OPCODE_ENDREP: + return 0; + break; + + case TGSI_OPCODE_PUSHA: + return 0; + break; + + case TGSI_OPCODE_POPA: + return 0; + break; + + case TGSI_OPCODE_CEIL: + return 0; + break; + + case TGSI_OPCODE_I2F: + return 0; + break; + + case TGSI_OPCODE_NOT: + return 0; + break; + +#if 0 + case TGSI_OPCODE_TRUNC: + FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) { + tmp0 = emit_fetch( bld, inst, 0, chan_index ); + emit_f2it( bld, 0 ); + emit_i2f( bld, 0 ); + emit_store( bld, inst, 0, chan_index, tmp0); + } + break; +#endif + + case TGSI_OPCODE_SHL: + return 0; + break; + + case TGSI_OPCODE_SHR: + return 0; + break; + + case TGSI_OPCODE_AND: + return 0; + break; + + case TGSI_OPCODE_OR: + return 0; + break; + + case TGSI_OPCODE_MOD: + return 0; + break; + + case TGSI_OPCODE_XOR: + return 0; + break; + + case TGSI_OPCODE_SAD: + return 0; + break; + + case TGSI_OPCODE_TXF: + return 0; + break; + + case TGSI_OPCODE_TXQ: + return 0; + break; + + case TGSI_OPCODE_CONT: + return 0; + break; + + case TGSI_OPCODE_EMIT: + return 0; + break; + + case TGSI_OPCODE_ENDPRIM: + return 0; + break; + + default: + return 0; + } + + return 1; +} + + +void +lp_build_tgsi_soa(LLVMBuilderRef builder, + const struct tgsi_token *tokens, + union lp_type type, + struct lp_build_mask_context *mask, + LLVMValueRef consts_ptr, + const LLVMValueRef *pos, + const LLVMValueRef (*inputs)[NUM_CHANNELS], + LLVMValueRef (*outputs)[NUM_CHANNELS], + lp_emit_fetch_texel_soa_callback emit_fetch_texel, + void *emit_fetch_texel_context) +{ + struct lp_build_tgsi_soa_context bld; + struct tgsi_parse_context parse; + uint num_immediates = 0; + unsigned i; + + /* Setup build context */ + memset(&bld, 0, sizeof bld); + lp_build_context_init(&bld.base, builder, type); + bld.mask = mask; + bld.pos = pos; + bld.inputs = inputs; + bld.outputs = outputs; + bld.consts_ptr = consts_ptr; + bld.emit_fetch_texel = emit_fetch_texel; + bld.emit_fetch_texel_context = emit_fetch_texel_context; + + tgsi_parse_init( &parse, tokens ); + + while( !tgsi_parse_end_of_tokens( &parse ) ) { + tgsi_parse_token( &parse ); + + switch( parse.FullToken.Token.Type ) { + case TGSI_TOKEN_TYPE_DECLARATION: + /* Input already interpolated */ + break; + + case TGSI_TOKEN_TYPE_INSTRUCTION: + if (!emit_instruction( &bld, &parse.FullToken.FullInstruction )) { + unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode; + const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode); + _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n", + info ? info->mnemonic : "<invalid>"); + } + break; + + case TGSI_TOKEN_TYPE_IMMEDIATE: + /* simply copy the immediate values into the next immediates[] slot */ + { + const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1; + assert(size <= 4); + assert(num_immediates < LP_MAX_IMMEDIATES); + for( i = 0; i < size; ++i ) + bld.immediates[num_immediates][i] = + lp_build_const_scalar(type, parse.FullToken.FullImmediate.u[i].Float); + for( i = size; i < 4; ++i ) + bld.immediates[num_immediates][i] = bld.base.undef; + num_immediates++; + } + break; + + default: + assert( 0 ); + } + } + + tgsi_parse_free( &parse ); +} + diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.c b/src/gallium/drivers/llvmpipe/lp_bld_type.c new file mode 100644 index 00000000000..8e0026fd973 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_type.c @@ -0,0 +1,170 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "util/u_debug.h" + +#include "lp_bld_type.h" +#include "lp_bld_const.h" + + +LLVMTypeRef +lp_build_elem_type(union lp_type type) +{ + if (type.floating) { + switch(type.width) { + case 32: + return LLVMFloatType(); + break; + case 64: + return LLVMDoubleType(); + break; + default: + assert(0); + return LLVMFloatType(); + } + } + else { + return LLVMIntType(type.width); + } +} + + +LLVMTypeRef +lp_build_vec_type(union lp_type type) +{ + LLVMTypeRef elem_type = lp_build_elem_type(type); + return LLVMVectorType(elem_type, type.length); +} + + +/** + * This function is a mirror of lp_build_elem_type() above. + * + * XXX: I'm not sure if it wouldn't be easier/efficient to just recreate the + * type and check for identity. + */ +boolean +lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type) +{ + LLVMTypeKind elem_kind; + + assert(elem_type); + if(!elem_type) + return FALSE; + + elem_kind = LLVMGetTypeKind(elem_type); + + if (type.floating) { + switch(type.width) { + case 32: + if(elem_kind != LLVMFloatTypeKind) + return FALSE; + break; + case 64: + if(elem_kind != LLVMDoubleTypeKind) + return FALSE; + break; + default: + assert(0); + return FALSE; + } + } + else { + if(elem_kind != LLVMIntegerTypeKind) + return FALSE; + + if(LLVMGetIntTypeWidth(elem_type) != type.width) + return FALSE; + } + + return TRUE; +} + + +boolean +lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type) +{ + LLVMTypeRef elem_type; + + assert(vec_type); + if(!vec_type) + return FALSE; + + if(LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind) + return FALSE; + + if(LLVMGetVectorSize(vec_type) != type.length) + return FALSE; + + elem_type = LLVMGetElementType(vec_type); + + return lp_check_elem_type(type, elem_type); +} + + +boolean +lp_check_value(union lp_type type, LLVMValueRef val) +{ + LLVMTypeRef vec_type; + + assert(val); + if(!val) + return FALSE; + + vec_type = LLVMTypeOf(val); + + return lp_check_vec_type(type, vec_type); +} + + +LLVMTypeRef +lp_build_int_elem_type(union lp_type type) +{ + return LLVMIntType(type.width); +} + + +LLVMTypeRef +lp_build_int_vec_type(union lp_type type) +{ + LLVMTypeRef elem_type = lp_build_int_elem_type(type); + return LLVMVectorType(elem_type, type.length); +} + + +void +lp_build_context_init(struct lp_build_context *bld, + LLVMBuilderRef builder, + union lp_type type) +{ + bld->builder = builder; + bld->type = type; + bld->undef = lp_build_undef(type); + bld->zero = lp_build_zero(type); + bld->one = lp_build_one(type); +} diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.h b/src/gallium/drivers/llvmpipe/lp_bld_type.h new file mode 100644 index 00000000000..3ce566be641 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_bld_type.h @@ -0,0 +1,174 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Convenient representation of SIMD types. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_BLD_TYPE_H +#define LP_BLD_TYPE_H + + +#include <llvm-c/Core.h> + +#include <pipe/p_compiler.h> + + +/** + * Several functions can only cope with vectors of length up to this value. + * You may need to increase that value if you want to represent bigger vectors. + */ +#define LP_MAX_VECTOR_LENGTH 16 + +#define LP_MAX_TYPE_WIDTH 64 + + +/** + * The LLVM type system can't conveniently express all the things we care about + * on the types used for intermediate computations, such as signed vs unsigned, + * normalized values, or fixed point. + */ +union lp_type { + struct { + /** + * Floating-point. Cannot be used with fixed. Integer numbers are + * represented by this zero. + */ + unsigned floating:1; + + /** + * Fixed-point. Cannot be used with floating. Integer numbers are + * represented by this zero. + */ + unsigned fixed:1; + + /** + * Whether it can represent negative values or not. + * + * If this is not set for floating point, it means that all values are + * assumed to be positive. + */ + unsigned sign:1; + + /** + * Whether values are normalized to fit [0, 1] interval, or [-1, 1] + * interval for signed types. + * + * For integer types it means the representable integer range should be + * interpreted as the interval above. + * + * For floating and fixed point formats it means the values should be + * clamped to the interval above. + */ + unsigned norm:1; + + /** + * Element width. + * + * For fixed point values, the fixed point is assumed to be at half the + * width. + */ + unsigned width:14; + + /** + * Vector length. + * + * width*length should be a power of two greater or equal to eight. + * + * @sa LP_MAX_VECTOR_LENGTH + */ + unsigned length:14; + }; + uint32_t value; +}; + + +/** + * We need most of the information here in order to correctly and efficiently + * translate an arithmetic operation into LLVM IR. Putting it here avoids the + * trouble of passing it as parameters. + */ +struct lp_build_context +{ + LLVMBuilderRef builder; + + /** + * This not only describes the input/output LLVM types, but also whether + * to normalize/clamp the results. + */ + union lp_type type; + + /** Same as lp_build_undef(type) */ + LLVMValueRef undef; + + /** Same as lp_build_zero(type) */ + LLVMValueRef zero; + + /** Same as lp_build_one(type) */ + LLVMValueRef one; +}; + + +LLVMTypeRef +lp_build_elem_type(union lp_type type); + + +LLVMTypeRef +lp_build_vec_type(union lp_type type); + + +boolean +lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type); + + +boolean +lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type); + + +boolean +lp_check_value(union lp_type type, LLVMValueRef val); + + +LLVMTypeRef +lp_build_int_elem_type(union lp_type type); + + +LLVMTypeRef +lp_build_int_vec_type(union lp_type type); + + +void +lp_build_context_init(struct lp_build_context *bld, + LLVMBuilderRef builder, + union lp_type type); + + +#endif /* !LP_BLD_TYPE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_buffer.c b/src/gallium/drivers/llvmpipe/lp_buffer.c new file mode 100644 index 00000000000..66f1f8e1383 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_buffer.c @@ -0,0 +1,150 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "util/u_memory.h" +#include "util/u_math.h" + +#include "lp_winsys.h" +#include "lp_screen.h" +#include "lp_texture.h" +#include "lp_buffer.h" + + +static void * +llvmpipe_buffer_map(struct pipe_screen *screen, + struct pipe_buffer *buf, + unsigned flags) +{ + struct llvmpipe_buffer *llvmpipe_buf = llvmpipe_buffer(buf); + return llvmpipe_buf->data; +} + + +static void +llvmpipe_buffer_unmap(struct pipe_screen *screen, + struct pipe_buffer *buf) +{ +} + + +static void +llvmpipe_buffer_destroy(struct pipe_buffer *buf) +{ + struct llvmpipe_buffer *sbuf = llvmpipe_buffer(buf); + + if (!sbuf->userBuffer) + align_free(sbuf->data); + + FREE(sbuf); +} + + +static struct pipe_buffer * +llvmpipe_buffer_create(struct pipe_screen *screen, + unsigned alignment, + unsigned usage, + unsigned size) +{ + struct llvmpipe_buffer *buffer = CALLOC_STRUCT(llvmpipe_buffer); + + pipe_reference_init(&buffer->base.reference, 1); + buffer->base.screen = screen; + buffer->base.alignment = MAX2(alignment, 16); + buffer->base.usage = usage; + buffer->base.size = size; + + buffer->data = align_malloc(size, alignment); + + return &buffer->base; +} + + +/** + * Create buffer which wraps user-space data. + */ +static struct pipe_buffer * +llvmpipe_user_buffer_create(struct pipe_screen *screen, + void *ptr, + unsigned bytes) +{ + struct llvmpipe_buffer *buffer; + + buffer = CALLOC_STRUCT(llvmpipe_buffer); + if(!buffer) + return NULL; + + pipe_reference_init(&buffer->base.reference, 1); + buffer->base.screen = screen; + buffer->base.size = bytes; + buffer->userBuffer = TRUE; + buffer->data = ptr; + + return &buffer->base; +} + + +static void +llvmpipe_fence_reference(struct pipe_screen *screen, + struct pipe_fence_handle **ptr, + struct pipe_fence_handle *fence) +{ +} + + +static int +llvmpipe_fence_signalled(struct pipe_screen *screen, + struct pipe_fence_handle *fence, + unsigned flag) +{ + return 0; +} + + +static int +llvmpipe_fence_finish(struct pipe_screen *screen, + struct pipe_fence_handle *fence, + unsigned flag) +{ + return 0; +} + + +void +llvmpipe_init_screen_buffer_funcs(struct pipe_screen *screen) +{ + screen->buffer_create = llvmpipe_buffer_create; + screen->user_buffer_create = llvmpipe_user_buffer_create; + screen->buffer_map = llvmpipe_buffer_map; + screen->buffer_unmap = llvmpipe_buffer_unmap; + screen->buffer_destroy = llvmpipe_buffer_destroy; + + screen->fence_reference = llvmpipe_fence_reference; + screen->fence_signalled = llvmpipe_fence_signalled; + screen->fence_finish = llvmpipe_fence_finish; + +} diff --git a/src/gallium/drivers/llvmpipe/lp_buffer.h b/src/gallium/drivers/llvmpipe/lp_buffer.h new file mode 100644 index 00000000000..d6b8184a0b9 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_buffer.h @@ -0,0 +1,55 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_BUFFER_H +#define LP_BUFFER_H + +#include "pipe/p_compiler.h" +#include "pipe/p_state.h" + + +struct llvmpipe_buffer +{ + struct pipe_buffer base; + boolean userBuffer; /** Is this a user-space buffer? */ + void *data; +}; + + +/** Cast wrapper */ +static INLINE struct llvmpipe_buffer * +llvmpipe_buffer( struct pipe_buffer *buf ) +{ + return (struct llvmpipe_buffer *)buf; +} + + +void +llvmpipe_init_screen_buffer_funcs(struct pipe_screen *screen); + + +#endif /* LP_BUFFER_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_clear.c b/src/gallium/drivers/llvmpipe/lp_clear.c new file mode 100644 index 00000000000..580cca5b463 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_clear.c @@ -0,0 +1,80 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * Copyright 2009 VMware, Inc. All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Brian Paul + * Michel Dänzer + */ + + +#include "pipe/p_defines.h" +#include "util/u_pack_color.h" +#include "lp_clear.h" +#include "lp_context.h" +#include "lp_surface.h" +#include "lp_state.h" +#include "lp_tile_cache.h" + + +/** + * Clear the given buffers to the specified values. + * No masking, no scissor (clear entire buffer). + */ +void +llvmpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba, + double depth, unsigned stencil) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + unsigned cv; + uint i; + + if (llvmpipe->no_rast) + return; + +#if 0 + llvmpipe_update_derived(llvmpipe); /* not needed?? */ +#endif + + if (buffers & PIPE_CLEAR_COLOR) { + for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) { + struct pipe_surface *ps = llvmpipe->framebuffer.cbufs[i]; + + util_pack_color(rgba, ps->format, &cv); + lp_tile_cache_clear(llvmpipe->cbuf_cache[i], rgba, cv); + } + } + + if (buffers & PIPE_CLEAR_DEPTHSTENCIL) { + struct pipe_surface *ps = llvmpipe->framebuffer.zsbuf; + + cv = util_pack_z_stencil(ps->format, depth, stencil); + + /* non-cached surface */ + pipe->surface_fill(pipe, ps, 0, 0, ps->width, ps->height, cv); + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_clear.h b/src/gallium/drivers/llvmpipe/lp_clear.h new file mode 100644 index 00000000000..6d4ffccdf45 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_clear.h @@ -0,0 +1,43 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Brian Paul + */ + +#ifndef LP_CLEAR_H +#define LP_CLEAR_H + +#include "pipe/p_state.h" +struct pipe_context; + +extern void +llvmpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba, + double depth, unsigned stencil); + + +#endif /* LP_CLEAR_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c new file mode 100644 index 00000000000..233d1df0e10 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_context.c @@ -0,0 +1,297 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * Copyright 2008 VMware, Inc. All rights reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Keith Whitwell <keith@tungstengraphics.com> + */ + +#include "draw/draw_context.h" +#include "pipe/p_defines.h" +#include "util/u_math.h" +#include "util/u_memory.h" +#include "lp_clear.h" +#include "lp_context.h" +#include "lp_flush.h" +#include "lp_prim_setup.h" +#include "lp_prim_vbuf.h" +#include "lp_state.h" +#include "lp_surface.h" +#include "lp_tile_cache.h" +#include "lp_tex_cache.h" +#include "lp_texture.h" +#include "lp_winsys.h" +#include "lp_query.h" + + + +/** + * Map any drawing surfaces which aren't already mapped + */ +void +llvmpipe_map_transfers(struct llvmpipe_context *lp) +{ + struct pipe_screen *screen = lp->pipe.screen; + struct pipe_surface *zsbuf = lp->framebuffer.zsbuf; + unsigned i; + + for (i = 0; i < lp->framebuffer.nr_cbufs; i++) { + lp_tile_cache_map_transfers(lp->cbuf_cache[i]); + } + + if(zsbuf) { + if(!lp->zsbuf_transfer) + lp->zsbuf_transfer = screen->get_tex_transfer(screen, zsbuf->texture, + zsbuf->face, zsbuf->level, zsbuf->zslice, + PIPE_TRANSFER_READ_WRITE, + 0, 0, zsbuf->width, zsbuf->height); + if(lp->zsbuf_transfer && !lp->zsbuf_map) + lp->zsbuf_map = screen->transfer_map(screen, lp->zsbuf_transfer); + + } +} + + +/** + * Unmap any mapped drawing surfaces + */ +void +llvmpipe_unmap_transfers(struct llvmpipe_context *lp) +{ + uint i; + + for (i = 0; i < lp->framebuffer.nr_cbufs; i++) { + lp_tile_cache_unmap_transfers(lp->cbuf_cache[i]); + } + + if(lp->zsbuf_transfer) { + struct pipe_screen *screen = lp->pipe.screen; + + if(lp->zsbuf_map) { + screen->transfer_unmap(screen, lp->zsbuf_transfer); + lp->zsbuf_map = NULL; + } + } +} + + +static void llvmpipe_destroy( struct pipe_context *pipe ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe ); + uint i; + + if (llvmpipe->draw) + draw_destroy( llvmpipe->draw ); + + for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) + lp_destroy_tile_cache(llvmpipe->cbuf_cache[i]); + + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) + lp_destroy_tex_tile_cache(llvmpipe->tex_cache[i]); + + for (i = 0; i < Elements(llvmpipe->constants); i++) { + if (llvmpipe->constants[i].buffer) { + pipe_buffer_reference(&llvmpipe->constants[i].buffer, NULL); + } + } + + align_free( llvmpipe ); +} + +static unsigned int +llvmpipe_is_texture_referenced( struct pipe_context *pipe, + struct pipe_texture *texture, + unsigned face, unsigned level) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe ); + unsigned i; + + if(llvmpipe->dirty_render_cache) { + for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) { + if(llvmpipe->framebuffer.cbufs[i] && + llvmpipe->framebuffer.cbufs[i]->texture == texture) + return PIPE_REFERENCED_FOR_WRITE; + } + if(llvmpipe->framebuffer.zsbuf && + llvmpipe->framebuffer.zsbuf->texture == texture) + return PIPE_REFERENCED_FOR_WRITE; + } + + /* FIXME: we also need to do the same for the texture cache */ + + return PIPE_UNREFERENCED; +} + +static unsigned int +llvmpipe_is_buffer_referenced( struct pipe_context *pipe, + struct pipe_buffer *buf) +{ + return PIPE_UNREFERENCED; +} + +struct pipe_context * +llvmpipe_create( struct pipe_screen *screen ) +{ + struct llvmpipe_context *llvmpipe; + uint i; + + llvmpipe = align_malloc(sizeof(struct llvmpipe_context), 16); + if (!llvmpipe) + return NULL; + + util_init_math(); + + memset(llvmpipe, 0, sizeof *llvmpipe); + + llvmpipe->pipe.winsys = screen->winsys; + llvmpipe->pipe.screen = screen; + llvmpipe->pipe.destroy = llvmpipe_destroy; + + /* state setters */ + llvmpipe->pipe.create_blend_state = llvmpipe_create_blend_state; + llvmpipe->pipe.bind_blend_state = llvmpipe_bind_blend_state; + llvmpipe->pipe.delete_blend_state = llvmpipe_delete_blend_state; + + llvmpipe->pipe.create_sampler_state = llvmpipe_create_sampler_state; + llvmpipe->pipe.bind_sampler_states = llvmpipe_bind_sampler_states; + llvmpipe->pipe.delete_sampler_state = llvmpipe_delete_sampler_state; + + llvmpipe->pipe.create_depth_stencil_alpha_state = llvmpipe_create_depth_stencil_state; + llvmpipe->pipe.bind_depth_stencil_alpha_state = llvmpipe_bind_depth_stencil_state; + llvmpipe->pipe.delete_depth_stencil_alpha_state = llvmpipe_delete_depth_stencil_state; + + llvmpipe->pipe.create_rasterizer_state = llvmpipe_create_rasterizer_state; + llvmpipe->pipe.bind_rasterizer_state = llvmpipe_bind_rasterizer_state; + llvmpipe->pipe.delete_rasterizer_state = llvmpipe_delete_rasterizer_state; + + llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state; + llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state; + llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state; + + llvmpipe->pipe.create_vs_state = llvmpipe_create_vs_state; + llvmpipe->pipe.bind_vs_state = llvmpipe_bind_vs_state; + llvmpipe->pipe.delete_vs_state = llvmpipe_delete_vs_state; + + llvmpipe->pipe.set_blend_color = llvmpipe_set_blend_color; + llvmpipe->pipe.set_clip_state = llvmpipe_set_clip_state; + llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer; + llvmpipe->pipe.set_framebuffer_state = llvmpipe_set_framebuffer_state; + llvmpipe->pipe.set_polygon_stipple = llvmpipe_set_polygon_stipple; + llvmpipe->pipe.set_scissor_state = llvmpipe_set_scissor_state; + llvmpipe->pipe.set_sampler_textures = llvmpipe_set_sampler_textures; + llvmpipe->pipe.set_viewport_state = llvmpipe_set_viewport_state; + + llvmpipe->pipe.set_vertex_buffers = llvmpipe_set_vertex_buffers; + llvmpipe->pipe.set_vertex_elements = llvmpipe_set_vertex_elements; + + llvmpipe->pipe.draw_arrays = llvmpipe_draw_arrays; + llvmpipe->pipe.draw_elements = llvmpipe_draw_elements; + llvmpipe->pipe.draw_range_elements = llvmpipe_draw_range_elements; + llvmpipe->pipe.set_edgeflags = llvmpipe_set_edgeflags; + + + llvmpipe->pipe.clear = llvmpipe_clear; + llvmpipe->pipe.flush = llvmpipe_flush; + + llvmpipe->pipe.is_texture_referenced = llvmpipe_is_texture_referenced; + llvmpipe->pipe.is_buffer_referenced = llvmpipe_is_buffer_referenced; + + llvmpipe_init_query_funcs( llvmpipe ); + llvmpipe_init_texture_funcs( llvmpipe ); + + /* + * Alloc caches for accessing drawing surfaces and textures. + */ + for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) + llvmpipe->cbuf_cache[i] = lp_create_tile_cache( screen ); + + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) + llvmpipe->tex_cache[i] = lp_create_tex_tile_cache( screen ); + + + /* vertex shader samplers */ + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + llvmpipe->tgsi.vert_samplers[i].base.get_samples = lp_get_samples; + llvmpipe->tgsi.vert_samplers[i].processor = TGSI_PROCESSOR_VERTEX; + llvmpipe->tgsi.vert_samplers[i].cache = llvmpipe->tex_cache[i]; + llvmpipe->tgsi.vert_samplers_list[i] = &llvmpipe->tgsi.vert_samplers[i]; + } + + /* fragment shader samplers */ + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + llvmpipe->tgsi.frag_samplers[i].base.get_samples = lp_get_samples; + llvmpipe->tgsi.frag_samplers[i].processor = TGSI_PROCESSOR_FRAGMENT; + llvmpipe->tgsi.frag_samplers[i].cache = llvmpipe->tex_cache[i]; + llvmpipe->tgsi.frag_samplers_list[i] = &llvmpipe->tgsi.frag_samplers[i]; + } + + /* + * Create drawing context and plug our rendering stage into it. + */ + llvmpipe->draw = draw_create(); + if (!llvmpipe->draw) + goto fail; + + draw_texture_samplers(llvmpipe->draw, + PIPE_MAX_SAMPLERS, + (struct tgsi_sampler **) + llvmpipe->tgsi.vert_samplers_list); + + llvmpipe->setup = lp_draw_render_stage(llvmpipe); + if (!llvmpipe->setup) + goto fail; + + if (debug_get_bool_option( "LP_NO_RAST", FALSE )) + llvmpipe->no_rast = TRUE; + + if (debug_get_bool_option( "LP_NO_VBUF", FALSE )) { + /* Deprecated path -- vbuf is the intended interface to the draw module: + */ + draw_set_rasterize_stage(llvmpipe->draw, llvmpipe->setup); + } + else { + lp_init_vbuf(llvmpipe); + } + + /* plug in AA line/point stages */ + draw_install_aaline_stage(llvmpipe->draw, &llvmpipe->pipe); + draw_install_aapoint_stage(llvmpipe->draw, &llvmpipe->pipe); + +#if USE_DRAW_STAGE_PSTIPPLE + /* Do polygon stipple w/ texture map + frag prog? */ + draw_install_pstipple_stage(llvmpipe->draw, &llvmpipe->pipe); +#endif + + lp_init_surface_functions(llvmpipe); + + return &llvmpipe->pipe; + + fail: + llvmpipe_destroy(&llvmpipe->pipe); + return NULL; +} + diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h new file mode 100644 index 00000000000..8d5a0d4f1fc --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_context.h @@ -0,0 +1,155 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + +#ifndef LP_CONTEXT_H +#define LP_CONTEXT_H + +#include "pipe/p_context.h" + +#include "draw/draw_vertex.h" + +#include "lp_tex_sample.h" +#include "lp_jit.h" + + +struct llvmpipe_vbuf_render; +struct draw_context; +struct draw_stage; +struct llvmpipe_tile_cache; +struct llvmpipe_tex_tile_cache; +struct lp_fragment_shader; +struct lp_vertex_shader; +struct lp_blend_state; + + +struct llvmpipe_context { + struct pipe_context pipe; /**< base class */ + + /** Constant state objects */ + const struct pipe_blend_state *blend; + const struct pipe_sampler_state *sampler[PIPE_MAX_SAMPLERS]; + const struct pipe_depth_stencil_alpha_state *depth_stencil; + const struct pipe_rasterizer_state *rasterizer; + struct lp_fragment_shader *fs; + const struct lp_vertex_shader *vs; + + /** Other rendering state */ + struct pipe_blend_color blend_color[4][16]; + struct pipe_clip_state clip; + struct pipe_constant_buffer constants[PIPE_SHADER_TYPES]; + struct pipe_framebuffer_state framebuffer; + struct pipe_poly_stipple poly_stipple; + struct pipe_scissor_state scissor; + struct pipe_texture *texture[PIPE_MAX_SAMPLERS]; + struct pipe_viewport_state viewport; + struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS]; + struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS]; + + unsigned num_samplers; + unsigned num_textures; + unsigned num_vertex_elements; + unsigned num_vertex_buffers; + + unsigned dirty; /**< Mask of LP_NEW_x flags */ + + /* Counter for occlusion queries. Note this supports overlapping + * queries. + */ + uint64_t occlusion_count; + unsigned active_query_count; + + /** Mapped vertex buffers */ + ubyte *mapped_vbuffer[PIPE_MAX_ATTRIBS]; + + /** Mapped constant buffers */ + void *mapped_constants[PIPE_SHADER_TYPES]; + + /** Vertex format */ + struct vertex_info vertex_info; + struct vertex_info vertex_info_vbuf; + + /** Which vertex shader output slot contains point size */ + int psize_slot; + + /* The reduced version of the primitive supplied by the state + * tracker. + */ + unsigned reduced_api_prim; + + /* The reduced primitive after unfilled triangles, wide-line + * decomposition, etc, are taken into account. This is the + * primitive actually rasterized. + */ + unsigned reduced_prim; + + /** Derived from scissor and surface bounds: */ + struct pipe_scissor_state cliprect; + + unsigned line_stipple_counter; + + /** TGSI exec things */ + struct { + struct lp_shader_sampler vert_samplers[PIPE_MAX_SAMPLERS]; + struct lp_shader_sampler *vert_samplers_list[PIPE_MAX_SAMPLERS]; + struct lp_shader_sampler frag_samplers[PIPE_MAX_SAMPLERS]; + struct lp_shader_sampler *frag_samplers_list[PIPE_MAX_SAMPLERS]; + } tgsi; + + /** The primitive drawing context */ + struct draw_context *draw; + struct draw_stage *setup; + struct draw_stage *vbuf; + struct llvmpipe_vbuf_render *vbuf_render; + + boolean dirty_render_cache; + + struct llvmpipe_tile_cache *cbuf_cache[PIPE_MAX_COLOR_BUFS]; + + /* TODO: we shouldn't be using external interfaces internally like this */ + struct pipe_transfer *zsbuf_transfer; + uint8_t *zsbuf_map; + + unsigned tex_timestamp; + struct llvmpipe_tex_tile_cache *tex_cache[PIPE_MAX_SAMPLERS]; + + unsigned no_rast : 1; + + struct lp_jit_context jit_context; +}; + + +static INLINE struct llvmpipe_context * +llvmpipe_context( struct pipe_context *pipe ) +{ + return (struct llvmpipe_context *)pipe; +} + +#endif /* LP_CONTEXT_H */ + diff --git a/src/gallium/drivers/llvmpipe/lp_draw_arrays.c b/src/gallium/drivers/llvmpipe/lp_draw_arrays.c new file mode 100644 index 00000000000..89772e62d31 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_draw_arrays.c @@ -0,0 +1,192 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Brian Paul + * Keith Whitwell + */ + + +#include "pipe/p_defines.h" +#include "pipe/p_context.h" +#include "pipe/internal/p_winsys_screen.h" +#include "pipe/p_inlines.h" +#include "util/u_prim.h" + +#include "lp_buffer.h" +#include "lp_context.h" +#include "lp_state.h" + +#include "draw/draw_context.h" + + + +static void +llvmpipe_map_constant_buffers(struct llvmpipe_context *lp) +{ + struct pipe_screen *screen = lp->pipe.screen; + uint i, size; + + for (i = 0; i < PIPE_SHADER_TYPES; i++) { + if (lp->constants[i].buffer && lp->constants[i].buffer->size) + lp->mapped_constants[i] = screen->buffer_map(screen, lp->constants[i].buffer, + PIPE_BUFFER_USAGE_CPU_READ); + } + + if (lp->constants[PIPE_SHADER_VERTEX].buffer) + size = lp->constants[PIPE_SHADER_VERTEX].buffer->size; + else + size = 0; + + lp->jit_context.constants = lp->mapped_constants[PIPE_SHADER_FRAGMENT]; + + draw_set_mapped_constant_buffer(lp->draw, + lp->mapped_constants[PIPE_SHADER_VERTEX], + size); +} + + +static void +llvmpipe_unmap_constant_buffers(struct llvmpipe_context *lp) +{ + struct pipe_screen *screen = lp->pipe.screen; + uint i; + + /* really need to flush all prims since the vert/frag shaders const buffers + * are going away now. + */ + draw_flush(lp->draw); + + draw_set_mapped_constant_buffer(lp->draw, NULL, 0); + + lp->jit_context.constants = NULL; + + for (i = 0; i < 2; i++) { + if (lp->constants[i].buffer && lp->constants[i].buffer->size) + screen->buffer_unmap(screen, lp->constants[i].buffer); + lp->mapped_constants[i] = NULL; + } +} + + +boolean +llvmpipe_draw_arrays(struct pipe_context *pipe, unsigned mode, + unsigned start, unsigned count) +{ + return llvmpipe_draw_elements(pipe, NULL, 0, mode, start, count); +} + + +/** + * Draw vertex arrays, with optional indexing. + * Basically, map the vertex buffers (and drawing surfaces), then hand off + * the drawing to the 'draw' module. + */ +boolean +llvmpipe_draw_range_elements(struct pipe_context *pipe, + struct pipe_buffer *indexBuffer, + unsigned indexSize, + unsigned min_index, + unsigned max_index, + unsigned mode, unsigned start, unsigned count) +{ + struct llvmpipe_context *lp = llvmpipe_context(pipe); + struct draw_context *draw = lp->draw; + unsigned i; + + lp->reduced_api_prim = u_reduced_prim(mode); + + if (lp->dirty) + llvmpipe_update_derived( lp ); + + llvmpipe_map_transfers(lp); + llvmpipe_map_constant_buffers(lp); + + /* + * Map vertex buffers + */ + for (i = 0; i < lp->num_vertex_buffers; i++) { + void *buf = llvmpipe_buffer(lp->vertex_buffer[i].buffer)->data; + draw_set_mapped_vertex_buffer(draw, i, buf); + } + + /* Map index buffer, if present */ + if (indexBuffer) { + void *mapped_indexes = llvmpipe_buffer(indexBuffer)->data; + draw_set_mapped_element_buffer_range(draw, indexSize, + min_index, + max_index, + mapped_indexes); + } + else { + /* no index/element buffer */ + draw_set_mapped_element_buffer_range(draw, 0, start, + start + count - 1, NULL); + } + + /* draw! */ + draw_arrays(draw, mode, start, count); + + /* + * unmap vertex/index buffers - will cause draw module to flush + */ + for (i = 0; i < lp->num_vertex_buffers; i++) { + draw_set_mapped_vertex_buffer(draw, i, NULL); + } + if (indexBuffer) { + draw_set_mapped_element_buffer(draw, 0, NULL); + } + + + /* Note: leave drawing surfaces mapped */ + llvmpipe_unmap_constant_buffers(lp); + + lp->dirty_render_cache = TRUE; + + return TRUE; +} + + +boolean +llvmpipe_draw_elements(struct pipe_context *pipe, + struct pipe_buffer *indexBuffer, + unsigned indexSize, + unsigned mode, unsigned start, unsigned count) +{ + return llvmpipe_draw_range_elements( pipe, indexBuffer, + indexSize, + 0, 0xffffffff, + mode, start, count ); +} + + +void +llvmpipe_set_edgeflags(struct pipe_context *pipe, const unsigned *edgeflags) +{ + struct llvmpipe_context *lp = llvmpipe_context(pipe); + draw_set_edgeflags(lp->draw, edgeflags); +} diff --git a/src/gallium/drivers/llvmpipe/lp_flush.c b/src/gallium/drivers/llvmpipe/lp_flush.c new file mode 100644 index 00000000000..b5c1c95bb73 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_flush.c @@ -0,0 +1,98 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Keith Whitwell <keith@tungstengraphics.com> + */ + + +#include "pipe/p_defines.h" +#include "draw/draw_context.h" +#include "lp_flush.h" +#include "lp_context.h" +#include "lp_surface.h" +#include "lp_state.h" +#include "lp_tile_cache.h" +#include "lp_tex_cache.h" +#include "lp_winsys.h" + + +void +llvmpipe_flush( struct pipe_context *pipe, + unsigned flags, + struct pipe_fence_handle **fence ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + uint i; + + draw_flush(llvmpipe->draw); + + if (flags & PIPE_FLUSH_SWAPBUFFERS) { + /* If this is a swapbuffers, just flush color buffers. + * + * The zbuffer changes are not discarded, but held in the cache + * in the hope that a later clear will wipe them out. + */ + for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) + if (llvmpipe->cbuf_cache[i]) + lp_flush_tile_cache(llvmpipe->cbuf_cache[i]); + + /* Need this call for hardware buffers before swapbuffers. + * + * there should probably be another/different flush-type function + * that's called before swapbuffers because we don't always want + * to unmap surfaces when flushing. + */ + llvmpipe_unmap_transfers(llvmpipe); + } + else if (flags & PIPE_FLUSH_RENDER_CACHE) { + for (i = 0; i < llvmpipe->framebuffer.nr_cbufs; i++) + if (llvmpipe->cbuf_cache[i]) + lp_flush_tile_cache(llvmpipe->cbuf_cache[i]); + + /* FIXME: untile zsbuf! */ + + llvmpipe->dirty_render_cache = FALSE; + } + + /* Enable to dump BMPs of the color/depth buffers each frame */ +#if 0 + if(flags & PIPE_FLUSH_FRAME) { + static unsigned frame_no = 1; + static char filename[256]; + util_snprintf(filename, sizeof(filename), "cbuf_%u.bmp", frame_no); + debug_dump_surface_bmp(filename, llvmpipe->framebuffer.cbufs[0]); + util_snprintf(filename, sizeof(filename), "zsbuf_%u.bmp", frame_no); + debug_dump_surface_bmp(filename, llvmpipe->framebuffer.zsbuf); + ++frame_no; + } +#endif + + if (fence) + *fence = NULL; +} + diff --git a/src/gallium/drivers/llvmpipe/lp_flush.h b/src/gallium/drivers/llvmpipe/lp_flush.h new file mode 100644 index 00000000000..10b2b525836 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_flush.h @@ -0,0 +1,37 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_FLUSH_H +#define LP_FLUSH_H + +struct pipe_context; +struct pipe_fence_handle; + +void llvmpipe_flush(struct pipe_context *pipe, unsigned flags, + struct pipe_fence_handle **fence); + +#endif diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c new file mode 100644 index 00000000000..d288460a1b8 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_jit.c @@ -0,0 +1,138 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * C - JIT interfaces + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include <llvm-c/Transforms/Scalar.h> + +#include "util/u_memory.h" +#include "lp_screen.h" +#include "lp_bld_intr.h" +#include "lp_jit.h" + + +static void +lp_jit_init_globals(struct llvmpipe_screen *screen) +{ + /* struct lp_jit_context */ + { + LLVMTypeRef elem_types[4]; + LLVMTypeRef context_type; + + elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* constants */ + elem_types[1] = LLVMPointerType(LLVMInt8Type(), 0); /* samplers */ + elem_types[2] = LLVMFloatType(); /* alpha_ref_value */ + elem_types[3] = LLVMPointerType(LLVMInt8Type(), 0); /* blend_color */ + + context_type = LLVMStructType(elem_types, Elements(elem_types), 0); + + LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, constants, + screen->target, context_type, 0); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, samplers, + screen->target, context_type, 1); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, alpha_ref_value, + screen->target, context_type, 2); + LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, blend_color, + screen->target, context_type, 3); + LP_CHECK_STRUCT_SIZE(struct lp_jit_context, + screen->target, context_type); + + LLVMAddTypeName(screen->module, "context", context_type); + + screen->context_ptr_type = LLVMPointerType(context_type, 0); + } + + /* fetch_texel + */ + { + LLVMTypeRef ret_type; + LLVMTypeRef arg_types[3]; + LLVMValueRef fetch_texel; + + ret_type = LLVMVoidType(); + arg_types[0] = LLVMPointerType(LLVMInt8Type(), 0); /* samplers */ + arg_types[1] = LLVMInt32Type(); /* unit */ + arg_types[2] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0); /* store */ + + fetch_texel = lp_declare_intrinsic(screen->module, "fetch_texel", + ret_type, arg_types, Elements(arg_types)); + + LLVMAddGlobalMapping(screen->engine, fetch_texel, lp_fetch_texel_soa); + } + +#ifdef DEBUG + LLVMDumpModule(screen->module); +#endif +} + + +void +lp_jit_screen_cleanup(struct llvmpipe_screen *screen) +{ + if(screen->engine) + LLVMDisposeExecutionEngine(screen->engine); + + if(screen->pass) + LLVMDisposePassManager(screen->pass); +} + + +void +lp_jit_screen_init(struct llvmpipe_screen *screen) +{ + char *error = NULL; + + screen->module = LLVMModuleCreateWithName("llvmpipe"); + + screen->provider = LLVMCreateModuleProviderForExistingModule(screen->module); + + if (LLVMCreateJITCompiler(&screen->engine, screen->provider, 1, &error)) { + fprintf(stderr, "%s\n", error); + LLVMDisposeMessage(error); + abort(); + } + + screen->target = LLVMGetExecutionEngineTargetData(screen->engine); + + screen->pass = LLVMCreateFunctionPassManager(screen->provider); + LLVMAddTargetData(screen->target, screen->pass); + /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, + * but there are more on SVN. */ + LLVMAddConstantPropagationPass(screen->pass); + LLVMAddInstructionCombiningPass(screen->pass); + LLVMAddPromoteMemoryToRegisterPass(screen->pass); + LLVMAddGVNPass(screen->pass); + LLVMAddCFGSimplificationPass(screen->pass); + + lp_jit_init_globals(screen); +} diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h new file mode 100644 index 00000000000..a7fb60f9f5c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_jit.h @@ -0,0 +1,109 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * C - JIT interfaces + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#ifndef LP_JIT_H +#define LP_JIT_H + + +#include "lp_bld_struct.h" + + +struct tgsi_sampler; +struct llvmpipe_screen; + + +/** + * This structure is passed directly to the generated fragment shader. + * + * It contains the derived state. + * + * Changes here must be reflected in the lp_jit_context_* macros and + * lp_jit_init_types function. Changes to the ordering should be avoided. + * + * Only use types with a clear size and padding here, in particular prefer the + * stdint.h types to the basic integer types. + */ +struct lp_jit_context +{ + const float *constants; + + struct tgsi_sampler **samplers; + + /* TODO: alpha reference value */ + float alpha_ref_value; + + /* TODO: blend constant color */ + uint8_t *blend_color; +}; + + +#define lp_jit_context_constants(_builder, _ptr) \ + lp_build_struct_get(_builder, _ptr, 0, "constants") + +#define lp_jit_context_samplers(_builder, _ptr) \ + lp_build_struct_get(_builder, _ptr, 1, "samplers") + +#define lp_jit_context_alpha_ref_value(_builder, _ptr) \ + lp_build_struct_get(_builder, _ptr, 2, "alpha_ref_value") + +#define lp_jit_context_blend_color(_builder, _ptr) \ + lp_build_struct_get(_builder, _ptr, 3, "blend_color") + + +typedef void +(*lp_jit_frag_func)(struct lp_jit_context *context, + uint32_t x, + uint32_t y, + const void *a0, + const void *dadx, + const void *dady, + uint32_t *mask, + void *color, + void *depth); + +void PIPE_CDECL +lp_fetch_texel_soa( struct tgsi_sampler **samplers, + uint32_t unit, + float *store ); + + +void +lp_jit_screen_cleanup(struct llvmpipe_screen *screen); + + +void +lp_jit_screen_init(struct llvmpipe_screen *screen); + + +#endif /* LP_JIT_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_prim_setup.c b/src/gallium/drivers/llvmpipe/lp_prim_setup.c new file mode 100644 index 00000000000..b14f8fb99d9 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_prim_setup.c @@ -0,0 +1,190 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * \brief A draw stage that drives our triangle setup routines from + * within the draw pipeline. One of two ways to drive setup, the + * other being in lp_prim_vbuf.c. + * + * \author Keith Whitwell <keith@tungstengraphics.com> + * \author Brian Paul + */ + + +#include "lp_context.h" +#include "lp_setup.h" +#include "lp_state.h" +#include "lp_prim_setup.h" +#include "draw/draw_pipe.h" +#include "draw/draw_vertex.h" +#include "util/u_memory.h" + +/** + * Triangle setup info (derived from draw_stage). + * Also used for line drawing (taking some liberties). + */ +struct setup_stage { + struct draw_stage stage; /**< This must be first (base class) */ + + struct setup_context *setup; +}; + + + +/** + * Basically a cast wrapper. + */ +static INLINE struct setup_stage *setup_stage( struct draw_stage *stage ) +{ + return (struct setup_stage *)stage; +} + + +typedef const float (*cptrf4)[4]; + +static void +do_tri(struct draw_stage *stage, struct prim_header *prim) +{ + struct setup_stage *setup = setup_stage( stage ); + + llvmpipe_setup_tri( setup->setup, + (cptrf4)prim->v[0]->data, + (cptrf4)prim->v[1]->data, + (cptrf4)prim->v[2]->data ); +} + +static void +do_line(struct draw_stage *stage, struct prim_header *prim) +{ + struct setup_stage *setup = setup_stage( stage ); + + llvmpipe_setup_line( setup->setup, + (cptrf4)prim->v[0]->data, + (cptrf4)prim->v[1]->data ); +} + +static void +do_point(struct draw_stage *stage, struct prim_header *prim) +{ + struct setup_stage *setup = setup_stage( stage ); + + llvmpipe_setup_point( setup->setup, + (cptrf4)prim->v[0]->data ); +} + + + + +static void setup_begin( struct draw_stage *stage ) +{ + struct setup_stage *setup = setup_stage(stage); + + llvmpipe_setup_prepare( setup->setup ); + + stage->point = do_point; + stage->line = do_line; + stage->tri = do_tri; +} + + +static void setup_first_point( struct draw_stage *stage, + struct prim_header *header ) +{ + setup_begin(stage); + stage->point( stage, header ); +} + +static void setup_first_line( struct draw_stage *stage, + struct prim_header *header ) +{ + setup_begin(stage); + stage->line( stage, header ); +} + + +static void setup_first_tri( struct draw_stage *stage, + struct prim_header *header ) +{ + setup_begin(stage); + stage->tri( stage, header ); +} + + + +static void setup_flush( struct draw_stage *stage, + unsigned flags ) +{ + stage->point = setup_first_point; + stage->line = setup_first_line; + stage->tri = setup_first_tri; +} + + +static void reset_stipple_counter( struct draw_stage *stage ) +{ +} + + +static void render_destroy( struct draw_stage *stage ) +{ + struct setup_stage *ssetup = setup_stage(stage); + llvmpipe_setup_destroy_context(ssetup->setup); + FREE( stage ); +} + + +/** + * Create a new primitive setup/render stage. + */ +struct draw_stage *lp_draw_render_stage( struct llvmpipe_context *llvmpipe ) +{ + struct setup_stage *sstage = CALLOC_STRUCT(setup_stage); + + sstage->setup = llvmpipe_setup_create_context(llvmpipe); + sstage->stage.draw = llvmpipe->draw; + sstage->stage.point = setup_first_point; + sstage->stage.line = setup_first_line; + sstage->stage.tri = setup_first_tri; + sstage->stage.flush = setup_flush; + sstage->stage.reset_stipple_counter = reset_stipple_counter; + sstage->stage.destroy = render_destroy; + + return (struct draw_stage *)sstage; +} + +struct setup_context * +lp_draw_setup_context( struct draw_stage *stage ) +{ + struct setup_stage *ssetup = setup_stage(stage); + return ssetup->setup; +} + +void +lp_draw_flush( struct draw_stage *stage ) +{ + stage->flush( stage, 0 ); +} diff --git a/src/gallium/drivers/llvmpipe/lp_prim_setup.h b/src/gallium/drivers/llvmpipe/lp_prim_setup.h new file mode 100644 index 00000000000..da6cae63751 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_prim_setup.h @@ -0,0 +1,85 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef LP_PRIM_SETUP_H +#define LP_PRIM_SETUP_H + + +/** + * vbuf is a special stage to gather the stream of triangles, lines, points + * together and reconstruct vertex buffers for hardware upload. + * + * First attempt, work in progress. + * + * TODO: + * - separate out vertex buffer building and primitive emit, ie >1 draw per vb. + * - tell vbuf stage how to build hw vertices directly + * - pass vbuf stage a buffer pointer for direct emit to agp/vram. + * + * + * + * Vertices are just an array of floats, with all the attributes + * packed. We currently assume a layout like: + * + * attr[0][0..3] - window position + * attr[1..n][0..3] - remaining attributes. + * + * Attributes are assumed to be 4 floats wide but are packed so that + * all the enabled attributes run contiguously. + */ + + +struct draw_stage; +struct llvmpipe_context; + + +typedef void (*vbuf_draw_func)( struct pipe_context *pipe, + unsigned prim, + const ushort *elements, + unsigned nr_elements, + const void *vertex_buffer, + unsigned nr_vertices ); + + +extern struct draw_stage * +lp_draw_render_stage( struct llvmpipe_context *llvmpipe ); + +extern struct setup_context * +lp_draw_setup_context( struct draw_stage * ); + +extern void +lp_draw_flush( struct draw_stage * ); + + +extern struct draw_stage * +lp_draw_vbuf_stage( struct draw_context *draw_context, + struct pipe_context *pipe, + vbuf_draw_func draw ); + + +#endif /* LP_PRIM_SETUP_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c new file mode 100644 index 00000000000..c394dcb61d0 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.c @@ -0,0 +1,607 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Interface between 'draw' module's output and the llvmpipe rasterizer/setup + * code. When the 'draw' module has finished filling a vertex buffer, the + * draw_arrays() functions below will be called. Loop over the vertices and + * call the point/line/tri setup functions. + * + * Authors + * Brian Paul + */ + + +#include "lp_context.h" +#include "lp_state.h" +#include "lp_prim_vbuf.h" +#include "lp_prim_setup.h" +#include "lp_setup.h" +#include "draw/draw_context.h" +#include "draw/draw_vbuf.h" +#include "util/u_memory.h" +#include "util/u_prim.h" + + +#define LP_MAX_VBUF_INDEXES 1024 +#define LP_MAX_VBUF_SIZE 4096 + +typedef const float (*cptrf4)[4]; + +/** + * Subclass of vbuf_render. + */ +struct llvmpipe_vbuf_render +{ + struct vbuf_render base; + struct llvmpipe_context *llvmpipe; + uint prim; + uint vertex_size; + uint nr_vertices; + uint vertex_buffer_size; + void *vertex_buffer; +}; + + +/** cast wrapper */ +static struct llvmpipe_vbuf_render * +llvmpipe_vbuf_render(struct vbuf_render *vbr) +{ + return (struct llvmpipe_vbuf_render *) vbr; +} + + +static const struct vertex_info * +lp_vbuf_get_vertex_info(struct vbuf_render *vbr) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + return llvmpipe_get_vbuf_vertex_info(cvbr->llvmpipe); +} + + +static boolean +lp_vbuf_allocate_vertices(struct vbuf_render *vbr, + ushort vertex_size, ushort nr_vertices) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + unsigned size = vertex_size * nr_vertices; + + if (cvbr->vertex_buffer_size < size) { + align_free(cvbr->vertex_buffer); + cvbr->vertex_buffer = align_malloc(size, 16); + cvbr->vertex_buffer_size = size; + } + + cvbr->vertex_size = vertex_size; + cvbr->nr_vertices = nr_vertices; + + return cvbr->vertex_buffer != NULL; +} + +static void +lp_vbuf_release_vertices(struct vbuf_render *vbr) +{ +#if 0 + { + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + const struct vertex_info *info = + llvmpipe_get_vbuf_vertex_info(cvbr->llvmpipe); + const float *vtx = (const float *) cvbr->vertex_buffer; + uint i, j; + debug_printf("%s (vtx_size = %u, vtx_used = %u)\n", + __FUNCTION__, cvbr->vertex_size, cvbr->nr_vertices); + for (i = 0; i < cvbr->nr_vertices; i++) { + for (j = 0; j < info->num_attribs; j++) { + uint k; + switch (info->attrib[j].emit) { + case EMIT_4F: k = 4; break; + case EMIT_3F: k = 3; break; + case EMIT_2F: k = 2; break; + case EMIT_1F: k = 1; break; + default: assert(0); + } + debug_printf("Vert %u attr %u: ", i, j); + while (k-- > 0) { + debug_printf("%g ", vtx[0]); + vtx++; + } + debug_printf("\n"); + } + } + } +#endif + + /* keep the old allocation for next time */ +} + +static void * +lp_vbuf_map_vertices(struct vbuf_render *vbr) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + return cvbr->vertex_buffer; +} + +static void +lp_vbuf_unmap_vertices(struct vbuf_render *vbr, + ushort min_index, + ushort max_index ) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + assert( cvbr->vertex_buffer_size >= (max_index+1) * cvbr->vertex_size ); + /* do nothing */ +} + + +static boolean +lp_vbuf_set_primitive(struct vbuf_render *vbr, unsigned prim) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + + /* XXX: break this dependency - make setup_context live under + * llvmpipe, rename the old "setup" draw stage to something else. + */ + struct setup_context *setup_ctx = lp_draw_setup_context(cvbr->llvmpipe->setup); + + llvmpipe_setup_prepare( setup_ctx ); + + cvbr->llvmpipe->reduced_prim = u_reduced_prim(prim); + cvbr->prim = prim; + return TRUE; + +} + + +static INLINE cptrf4 get_vert( const void *vertex_buffer, + int index, + int stride ) +{ + return (cptrf4)((char *)vertex_buffer + index * stride); +} + + +/** + * draw elements / indexed primitives + */ +static void +lp_vbuf_draw(struct vbuf_render *vbr, const ushort *indices, uint nr) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + struct llvmpipe_context *llvmpipe = cvbr->llvmpipe; + const unsigned stride = llvmpipe->vertex_info_vbuf.size * sizeof(float); + const void *vertex_buffer = cvbr->vertex_buffer; + unsigned i; + + /* XXX: break this dependency - make setup_context live under + * llvmpipe, rename the old "setup" draw stage to something else. + */ + struct draw_stage *setup = llvmpipe->setup; + struct setup_context *setup_ctx = lp_draw_setup_context(setup); + + switch (cvbr->prim) { + case PIPE_PRIM_POINTS: + for (i = 0; i < nr; i++) { + llvmpipe_setup_point( setup_ctx, + get_vert(vertex_buffer, indices[i-0], stride) ); + } + break; + + case PIPE_PRIM_LINES: + for (i = 1; i < nr; i += 2) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + break; + + case PIPE_PRIM_LINE_STRIP: + for (i = 1; i < nr; i ++) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + break; + + case PIPE_PRIM_LINE_LOOP: + for (i = 1; i < nr; i ++) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + if (nr) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, indices[nr-1], stride), + get_vert(vertex_buffer, indices[0], stride) ); + } + break; + + case PIPE_PRIM_TRIANGLES: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i += 3) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[i-2], stride) ); + } + } + else { + for (i = 2; i < nr; i += 3) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + } + break; + + case PIPE_PRIM_TRIANGLE_STRIP: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i+(i&1)-1], stride), + get_vert(vertex_buffer, indices[i-(i&1)], stride), + get_vert(vertex_buffer, indices[i-2], stride) ); + } + } + else { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i+(i&1)-2], stride), + get_vert(vertex_buffer, indices[i-(i&1)-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + } + break; + + case PIPE_PRIM_TRIANGLE_FAN: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[0], stride), + get_vert(vertex_buffer, indices[i-1], stride) ); + } + } + else { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[0], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + } + break; + + case PIPE_PRIM_QUADS: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 3; i < nr; i += 4) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-3], stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[i-3], stride) ); + } + } + else { + for (i = 3; i < nr; i += 4) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-3], stride), + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + } + break; + + case PIPE_PRIM_QUAD_STRIP: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 3; i < nr; i += 2) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-3], stride)); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[i-3], stride) ); + } + } + else { + for (i = 3; i < nr; i += 2) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-3], stride), + get_vert(vertex_buffer, indices[i-2], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[i-3], stride), + get_vert(vertex_buffer, indices[i-0], stride) ); + } + } + break; + + case PIPE_PRIM_POLYGON: + /* Almost same as tri fan but the _first_ vertex specifies the flat + * shading color. Note that the first polygon vertex is passed as + * the last triangle vertex here. + * flatshade_first state makes no difference. + */ + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, indices[i-0], stride), + get_vert(vertex_buffer, indices[i-1], stride), + get_vert(vertex_buffer, indices[0], stride) ); + } + break; + + default: + assert(0); + } + + /* XXX: why are we calling this??? If we had to call something, it + * would be a function in lp_setup.c: + */ + lp_draw_flush( setup ); +} + + +/** + * This function is hit when the draw module is working in pass-through mode. + * It's up to us to convert the vertex array into point/line/tri prims. + */ +static void +lp_vbuf_draw_arrays(struct vbuf_render *vbr, uint start, uint nr) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + struct llvmpipe_context *llvmpipe = cvbr->llvmpipe; + const unsigned stride = llvmpipe->vertex_info_vbuf.size * sizeof(float); + const void *vertex_buffer = + (void *) get_vert(cvbr->vertex_buffer, start, stride); + unsigned i; + + /* XXX: break this dependency - make setup_context live under + * llvmpipe, rename the old "setup" draw stage to something else. + */ + struct draw_stage *setup = llvmpipe->setup; + struct setup_context *setup_ctx = lp_draw_setup_context(setup); + + switch (cvbr->prim) { + case PIPE_PRIM_POINTS: + for (i = 0; i < nr; i++) { + llvmpipe_setup_point( setup_ctx, + get_vert(vertex_buffer, i-0, stride) ); + } + break; + + case PIPE_PRIM_LINES: + for (i = 1; i < nr; i += 2) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + break; + + case PIPE_PRIM_LINE_STRIP: + for (i = 1; i < nr; i ++) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + break; + + case PIPE_PRIM_LINE_LOOP: + for (i = 1; i < nr; i ++) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + if (nr) { + llvmpipe_setup_line( setup_ctx, + get_vert(vertex_buffer, nr-1, stride), + get_vert(vertex_buffer, 0, stride) ); + } + break; + + case PIPE_PRIM_TRIANGLES: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i += 3) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, i-2, stride) ); + } + } + else { + for (i = 2; i < nr; i += 3) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + } + break; + + case PIPE_PRIM_TRIANGLE_STRIP: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i++) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i+(i&1)-1, stride), + get_vert(vertex_buffer, i-(i&1), stride), + get_vert(vertex_buffer, i-2, stride) ); + } + } + else { + for (i = 2; i < nr; i++) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i+(i&1)-2, stride), + get_vert(vertex_buffer, i-(i&1)-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + } + break; + + case PIPE_PRIM_TRIANGLE_FAN: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, 0, stride), + get_vert(vertex_buffer, i-1, stride) ); + } + } + else { + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, 0, stride), + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + } + break; + + case PIPE_PRIM_QUADS: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 3; i < nr; i += 4) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-3, stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, i-3, stride) ); + } + } + else { + for (i = 3; i < nr; i += 4) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-3, stride), + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-0, stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + } + break; + + case PIPE_PRIM_QUAD_STRIP: + if (llvmpipe->rasterizer->flatshade_first) { + for (i = 3; i < nr; i += 2) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-3, stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, i-3, stride) ); + } + } + else { + for (i = 3; i < nr; i += 2) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-3, stride), + get_vert(vertex_buffer, i-2, stride), + get_vert(vertex_buffer, i-0, stride) ); + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-3, stride), + get_vert(vertex_buffer, i-0, stride) ); + } + } + break; + + case PIPE_PRIM_POLYGON: + /* Almost same as tri fan but the _first_ vertex specifies the flat + * shading color. Note that the first polygon vertex is passed as + * the last triangle vertex here. + * flatshade_first state makes no difference. + */ + for (i = 2; i < nr; i += 1) { + llvmpipe_setup_tri( setup_ctx, + get_vert(vertex_buffer, i-1, stride), + get_vert(vertex_buffer, i-0, stride), + get_vert(vertex_buffer, 0, stride) ); + } + break; + + default: + assert(0); + } +} + + + +static void +lp_vbuf_destroy(struct vbuf_render *vbr) +{ + struct llvmpipe_vbuf_render *cvbr = llvmpipe_vbuf_render(vbr); + cvbr->llvmpipe->vbuf_render = NULL; + FREE(cvbr); +} + + +/** + * Initialize the post-transform vertex buffer information for the given + * context. + */ +void +lp_init_vbuf(struct llvmpipe_context *lp) +{ + assert(lp->draw); + + lp->vbuf_render = CALLOC_STRUCT(llvmpipe_vbuf_render); + + lp->vbuf_render->base.max_indices = LP_MAX_VBUF_INDEXES; + lp->vbuf_render->base.max_vertex_buffer_bytes = LP_MAX_VBUF_SIZE; + + lp->vbuf_render->base.get_vertex_info = lp_vbuf_get_vertex_info; + lp->vbuf_render->base.allocate_vertices = lp_vbuf_allocate_vertices; + lp->vbuf_render->base.map_vertices = lp_vbuf_map_vertices; + lp->vbuf_render->base.unmap_vertices = lp_vbuf_unmap_vertices; + lp->vbuf_render->base.set_primitive = lp_vbuf_set_primitive; + lp->vbuf_render->base.draw = lp_vbuf_draw; + lp->vbuf_render->base.draw_arrays = lp_vbuf_draw_arrays; + lp->vbuf_render->base.release_vertices = lp_vbuf_release_vertices; + lp->vbuf_render->base.destroy = lp_vbuf_destroy; + + lp->vbuf_render->llvmpipe = lp; + + lp->vbuf = draw_vbuf_stage(lp->draw, &lp->vbuf_render->base); + + draw_set_rasterize_stage(lp->draw, lp->vbuf); + + draw_set_render(lp->draw, &lp->vbuf_render->base); +} diff --git a/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h new file mode 100644 index 00000000000..6c4e6063e6d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_prim_vbuf.h @@ -0,0 +1,38 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_VBUF_H +#define LP_VBUF_H + + +struct llvmpipe_context; + +extern void +lp_init_vbuf(struct llvmpipe_context *llvmpipe); + + +#endif /* LP_VBUF_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_quad.h b/src/gallium/drivers/llvmpipe/lp_quad.h new file mode 100644 index 00000000000..7eb05de77a1 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_quad.h @@ -0,0 +1,114 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + +#ifndef LP_QUAD_H +#define LP_QUAD_H + +#include "pipe/p_state.h" +#include "tgsi/tgsi_exec.h" + + +#define QUAD_PRIM_POINT 1 +#define QUAD_PRIM_LINE 2 +#define QUAD_PRIM_TRI 3 + + +/* The rasterizer generates 2x2 quads of fragment and feeds them to + * the current fp_machine (see below). + * Remember that Y=0=top with Y increasing down the window. + */ +#define QUAD_TOP_LEFT 0 +#define QUAD_TOP_RIGHT 1 +#define QUAD_BOTTOM_LEFT 2 +#define QUAD_BOTTOM_RIGHT 3 + +#define MASK_TOP_LEFT (1 << QUAD_TOP_LEFT) +#define MASK_TOP_RIGHT (1 << QUAD_TOP_RIGHT) +#define MASK_BOTTOM_LEFT (1 << QUAD_BOTTOM_LEFT) +#define MASK_BOTTOM_RIGHT (1 << QUAD_BOTTOM_RIGHT) +#define MASK_ALL 0xf + + +/** + * Quad stage inputs (pos, coverage, front/back face, etc) + */ +struct quad_header_input +{ + int x0, y0; /**< quad window pos, always even */ + float coverage[QUAD_SIZE]; /**< fragment coverage for antialiasing */ + unsigned facing:1; /**< Front (0) or back (1) facing? */ + unsigned prim:2; /**< QUAD_PRIM_POINT, LINE, TRI */ +}; + + +/** + * Quad stage inputs/outputs. + */ +struct quad_header_inout +{ + unsigned mask:4; +}; + + +/** + * Quad stage outputs (color & depth). + */ +struct quad_header_output +{ + /** colors in SOA format (rrrr, gggg, bbbb, aaaa) */ + float ALIGN16_ATTRIB color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][QUAD_SIZE]; +}; + + +/** + * Input interpolation coefficients + */ +struct quad_interp_coef +{ + float ALIGN16_ATTRIB a0[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + float ALIGN16_ATTRIB dadx[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; + float ALIGN16_ATTRIB dady[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS]; +}; + + +/** + * Encodes everything we need to know about a 2x2 pixel block. Uses + * "Channel-Serial" or "SoA" layout. + */ +struct quad_header { + struct quad_header_input input; + struct quad_header_inout inout; + + /* Redundant/duplicated: + */ + const struct quad_interp_coef *coef; +}; + +#endif /* LP_QUAD_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_query.c b/src/gallium/drivers/llvmpipe/lp_query.c new file mode 100644 index 00000000000..5554285425d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_query.c @@ -0,0 +1,111 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Keith Whitwell <keith@tungstengraphics.com> + */ + +#include "draw/draw_context.h" +#include "pipe/p_defines.h" +#include "util/u_memory.h" +#include "lp_context.h" +#include "lp_query.h" +#include "lp_state.h" + +struct llvmpipe_query { + uint64_t start; + uint64_t end; +}; + + +static struct llvmpipe_query *llvmpipe_query( struct pipe_query *p ) +{ + return (struct llvmpipe_query *)p; +} + +static struct pipe_query * +llvmpipe_create_query(struct pipe_context *pipe, + unsigned type) +{ + assert(type == PIPE_QUERY_OCCLUSION_COUNTER); + return (struct pipe_query *)CALLOC_STRUCT( llvmpipe_query ); +} + + +static void +llvmpipe_destroy_query(struct pipe_context *pipe, struct pipe_query *q) +{ + FREE(q); +} + + +static void +llvmpipe_begin_query(struct pipe_context *pipe, struct pipe_query *q) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe ); + struct llvmpipe_query *sq = llvmpipe_query(q); + + sq->start = llvmpipe->occlusion_count; + llvmpipe->active_query_count++; + llvmpipe->dirty |= LP_NEW_QUERY; +} + + +static void +llvmpipe_end_query(struct pipe_context *pipe, struct pipe_query *q) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe ); + struct llvmpipe_query *sq = llvmpipe_query(q); + + llvmpipe->active_query_count--; + sq->end = llvmpipe->occlusion_count; + llvmpipe->dirty |= LP_NEW_QUERY; +} + + +static boolean +llvmpipe_get_query_result(struct pipe_context *pipe, + struct pipe_query *q, + boolean wait, + uint64_t *result ) +{ + struct llvmpipe_query *sq = llvmpipe_query(q); + *result = sq->end - sq->start; + return TRUE; +} + + +void llvmpipe_init_query_funcs(struct llvmpipe_context *llvmpipe ) +{ + llvmpipe->pipe.create_query = llvmpipe_create_query; + llvmpipe->pipe.destroy_query = llvmpipe_destroy_query; + llvmpipe->pipe.begin_query = llvmpipe_begin_query; + llvmpipe->pipe.end_query = llvmpipe_end_query; + llvmpipe->pipe.get_query_result = llvmpipe_get_query_result; +} + + diff --git a/src/gallium/drivers/llvmpipe/lp_query.h b/src/gallium/drivers/llvmpipe/lp_query.h new file mode 100644 index 00000000000..fa9fcd87139 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_query.h @@ -0,0 +1,39 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Author: + * Keith Whitwell + */ + +#ifndef LP_QUERY_H +#define LP_QUERY_H + +struct llvmpipe_context; +extern void llvmpipe_init_query_funcs(struct llvmpipe_context * ); + + +#endif /* LP_QUERY_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c new file mode 100644 index 00000000000..125035771e5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_screen.c @@ -0,0 +1,243 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "util/u_memory.h" +#include "util/u_simple_screen.h" +#include "pipe/internal/p_winsys_screen.h" +#include "pipe/p_defines.h" +#include "pipe/p_screen.h" + +#include "lp_texture.h" +#include "lp_buffer.h" +#include "lp_winsys.h" +#include "lp_jit.h" +#include "lp_screen.h" + + +static const char * +llvmpipe_get_vendor(struct pipe_screen *screen) +{ + return "VMware, Inc."; +} + + +static const char * +llvmpipe_get_name(struct pipe_screen *screen) +{ + return "llvmpipe"; +} + + +static int +llvmpipe_get_param(struct pipe_screen *screen, int param) +{ + switch (param) { + case PIPE_CAP_MAX_TEXTURE_IMAGE_UNITS: + return PIPE_MAX_SAMPLERS; + case PIPE_CAP_MAX_VERTEX_TEXTURE_UNITS: + return PIPE_MAX_SAMPLERS; + case PIPE_CAP_NPOT_TEXTURES: + return 1; + case PIPE_CAP_TWO_SIDED_STENCIL: + return 1; + case PIPE_CAP_GLSL: + return 1; + case PIPE_CAP_S3TC: + return 0; + case PIPE_CAP_ANISOTROPIC_FILTER: + return 0; + case PIPE_CAP_POINT_SPRITE: + return 1; + case PIPE_CAP_MAX_RENDER_TARGETS: + return PIPE_MAX_COLOR_BUFS; + case PIPE_CAP_OCCLUSION_QUERY: + return 1; + case PIPE_CAP_TEXTURE_MIRROR_CLAMP: + return 1; + case PIPE_CAP_TEXTURE_MIRROR_REPEAT: + return 1; + case PIPE_CAP_TEXTURE_SHADOW_MAP: + return 1; + case PIPE_CAP_MAX_TEXTURE_2D_LEVELS: + return 13; /* max 4Kx4K */ + case PIPE_CAP_MAX_TEXTURE_3D_LEVELS: + return 8; /* max 128x128x128 */ + case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS: + return 13; /* max 4Kx4K */ + case PIPE_CAP_TGSI_CONT_SUPPORTED: + return 1; + case PIPE_CAP_BLEND_EQUATION_SEPARATE: + return 1; + default: + return 0; + } +} + + +static float +llvmpipe_get_paramf(struct pipe_screen *screen, int param) +{ + switch (param) { + case PIPE_CAP_MAX_LINE_WIDTH: + /* fall-through */ + case PIPE_CAP_MAX_LINE_WIDTH_AA: + return 255.0; /* arbitrary */ + case PIPE_CAP_MAX_POINT_WIDTH: + /* fall-through */ + case PIPE_CAP_MAX_POINT_WIDTH_AA: + return 255.0; /* arbitrary */ + case PIPE_CAP_MAX_TEXTURE_ANISOTROPY: + return 16.0; /* not actually signficant at this time */ + case PIPE_CAP_MAX_TEXTURE_LOD_BIAS: + return 16.0; /* arbitrary */ + default: + return 0; + } +} + + +/** + * Query format support for creating a texture, drawing surface, etc. + * \param format the format to test + * \param type one of PIPE_TEXTURE, PIPE_SURFACE + */ +static boolean +llvmpipe_is_format_supported( struct pipe_screen *_screen, + enum pipe_format format, + enum pipe_texture_target target, + unsigned tex_usage, + unsigned geom_flags ) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + struct llvmpipe_winsys *winsys = screen->winsys; + + assert(target == PIPE_TEXTURE_1D || + target == PIPE_TEXTURE_2D || + target == PIPE_TEXTURE_3D || + target == PIPE_TEXTURE_CUBE); + + if(format == PIPE_FORMAT_Z16_UNORM) + return FALSE; + if(format == PIPE_FORMAT_S8_UNORM) + return FALSE; + + switch(format) { + case PIPE_FORMAT_DXT1_RGB: + case PIPE_FORMAT_DXT1_RGBA: + case PIPE_FORMAT_DXT3_RGBA: + case PIPE_FORMAT_DXT5_RGBA: + return FALSE; + default: + break; + } + + if(tex_usage & PIPE_TEXTURE_USAGE_DISPLAY_TARGET) + return winsys->is_displaytarget_format_supported(winsys, format); + + return TRUE; +} + + +static struct pipe_buffer * +llvmpipe_surface_buffer_create(struct pipe_screen *screen, + unsigned width, unsigned height, + enum pipe_format format, + unsigned tex_usage, + unsigned usage, + unsigned *stride) +{ + /* This function should never be used */ + assert(0); + return NULL; +} + + +static void +llvmpipe_flush_frontbuffer(struct pipe_screen *_screen, + struct pipe_surface *surface, + void *context_private) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + struct llvmpipe_winsys *winsys = screen->winsys; + struct llvmpipe_texture *texture = llvmpipe_texture(surface->texture); + + assert(texture->dt); + if (texture->dt) + winsys->displaytarget_display(winsys, texture->dt, context_private); +} + + +static void +llvmpipe_destroy_screen( struct pipe_screen *_screen ) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + + struct pipe_winsys *winsys = _screen->winsys; + + lp_jit_screen_cleanup(screen); + + if(winsys->destroy) + winsys->destroy(winsys); + + FREE(screen); +} + + + +/** + * Create a new pipe_screen object + * Note: we're not presently subclassing pipe_screen (no llvmpipe_screen). + */ +struct pipe_screen * +llvmpipe_create_screen(struct llvmpipe_winsys *winsys) +{ + struct llvmpipe_screen *screen = CALLOC_STRUCT(llvmpipe_screen); + + if (!screen) + return NULL; + + screen->winsys = winsys; + + screen->base.destroy = llvmpipe_destroy_screen; + + screen->base.get_name = llvmpipe_get_name; + screen->base.get_vendor = llvmpipe_get_vendor; + screen->base.get_param = llvmpipe_get_param; + screen->base.get_paramf = llvmpipe_get_paramf; + screen->base.is_format_supported = llvmpipe_is_format_supported; + + screen->base.surface_buffer_create = llvmpipe_surface_buffer_create; + screen->base.flush_frontbuffer = llvmpipe_flush_frontbuffer; + + llvmpipe_init_screen_texture_funcs(&screen->base); + llvmpipe_init_screen_buffer_funcs(&screen->base); + + lp_jit_screen_init(screen); + + return &screen->base; +} diff --git a/src/gallium/drivers/llvmpipe/lp_screen.h b/src/gallium/drivers/llvmpipe/lp_screen.h new file mode 100644 index 00000000000..4a1b4d6f3e2 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_screen.h @@ -0,0 +1,79 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @author Jose Fonseca <jfonseca@vmware.com> + * @author Keith Whitwell <keith@tungstengraphics.com> + */ + +#ifndef LP_SCREEN_H +#define LP_SCREEN_H + +#include <llvm-c/Core.h> +#include <llvm-c/Analysis.h> +#include <llvm-c/Target.h> +#include <llvm-c/ExecutionEngine.h> + +#include "pipe/p_screen.h" +#include "pipe/p_defines.h" + + +struct llvmpipe_winsys; + + +struct llvmpipe_screen +{ + struct pipe_screen base; + + struct llvmpipe_winsys *winsys; + + LLVMModuleRef module; + LLVMExecutionEngineRef engine; + LLVMModuleProviderRef provider; + LLVMTargetDataRef target; + LLVMPassManagerRef pass; + + LLVMTypeRef context_ptr_type; + + /* Increments whenever textures are modified. Contexts can track + * this. + */ + unsigned timestamp; +}; + + + + +static INLINE struct llvmpipe_screen * +llvmpipe_screen( struct pipe_screen *pipe ) +{ + return (struct llvmpipe_screen *)pipe; +} + + +#endif /* LP_SCREEN_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c new file mode 100644 index 00000000000..d145f6d6bbc --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_setup.c @@ -0,0 +1,1482 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * \brief Primitive rasterization/rendering (points, lines, triangles) + * + * \author Keith Whitwell <keith@tungstengraphics.com> + * \author Brian Paul + */ + +#include "lp_context.h" +#include "lp_prim_setup.h" +#include "lp_quad.h" +#include "lp_setup.h" +#include "lp_state.h" +#include "draw/draw_context.h" +#include "draw/draw_private.h" +#include "draw/draw_vertex.h" +#include "pipe/p_shader_tokens.h" +#include "pipe/p_thread.h" +#include "util/u_math.h" +#include "util/u_memory.h" +#include "lp_tile_cache.h" +#include "lp_tile_soa.h" + + +#define DEBUG_VERTS 0 +#define DEBUG_FRAGS 0 + +/** + * Triangle edge info + */ +struct edge { + float dx; /**< X(v1) - X(v0), used only during setup */ + float dy; /**< Y(v1) - Y(v0), used only during setup */ + float dxdy; /**< dx/dy */ + float sx, sy; /**< first sample point coord */ + int lines; /**< number of lines on this edge */ +}; + + +#define MAX_QUADS 16 + + +/** + * Triangle setup info (derived from draw_stage). + * Also used for line drawing (taking some liberties). + */ +struct setup_context { + struct llvmpipe_context *llvmpipe; + + /* Vertices are just an array of floats making up each attribute in + * turn. Currently fixed at 4 floats, but should change in time. + * Codegen will help cope with this. + */ + const float (*vmax)[4]; + const float (*vmid)[4]; + const float (*vmin)[4]; + const float (*vprovoke)[4]; + + struct edge ebot; + struct edge etop; + struct edge emaj; + + float oneoverarea; + int facing; + + struct quad_header quad[MAX_QUADS]; + struct quad_header *quad_ptrs[MAX_QUADS]; + unsigned count; + + struct quad_interp_coef coef; + + struct { + int left[2]; /**< [0] = row0, [1] = row1 */ + int right[2]; + int y; + } span; + +#if DEBUG_FRAGS + uint numFragsEmitted; /**< per primitive */ + uint numFragsWritten; /**< per primitive */ +#endif + + unsigned winding; /* which winding to cull */ +}; + + + +/** + * Execute fragment shader for the four fragments in the quad. + */ +static void +shade_quads(struct llvmpipe_context *llvmpipe, + struct quad_header *quads[], + unsigned nr) +{ + struct lp_fragment_shader *fs = llvmpipe->fs; + struct quad_header *quad = quads[0]; + const unsigned x = quad->input.x0; + const unsigned y = quad->input.y0; + uint8_t *tile = lp_get_cached_tile(llvmpipe->cbuf_cache[0], x, y); + uint8_t *color; + void *depth; + uint32_t ALIGN16_ATTRIB mask[4][NUM_CHANNELS]; + unsigned chan_index; + unsigned q; + + assert(fs->current); + if(!fs->current) + return; + + /* Sanity checks */ + assert(nr * QUAD_SIZE == TILE_VECTOR_HEIGHT * TILE_VECTOR_WIDTH); + assert(x % TILE_VECTOR_WIDTH == 0); + assert(y % TILE_VECTOR_HEIGHT == 0); + for (q = 0; q < nr; ++q) { + assert(quads[q]->input.x0 == x + q*2); + assert(quads[q]->input.y0 == y); + } + + /* mask */ + for (q = 0; q < 4; ++q) + for (chan_index = 0; chan_index < NUM_CHANNELS; ++chan_index) + mask[q][chan_index] = quads[q]->inout.mask & (1 << chan_index) ? ~0 : 0; + + /* color buffer */ + color = &TILE_PIXEL(tile, x & (TILE_SIZE-1), y & (TILE_SIZE-1), 0); + + /* depth buffer */ + if(llvmpipe->zsbuf_map) { + assert((x % 2) == 0); + assert((y % 2) == 0); + depth = llvmpipe->zsbuf_map + + y*llvmpipe->zsbuf_transfer->stride + + 2*x*llvmpipe->zsbuf_transfer->block.size; + } + else + depth = NULL; + + /* TODO: blend color */ + + assert((((uintptr_t)mask) & 0xf) == 0); + assert((((uintptr_t)depth) & 0xf) == 0); + assert((((uintptr_t)color) & 0xf) == 0); + assert((((uintptr_t)llvmpipe->jit_context.blend_color) & 0xf) == 0); + + /* run shader */ + fs->current->jit_function( &llvmpipe->jit_context, + x, y, + quad->coef->a0, + quad->coef->dadx, + quad->coef->dady, + &mask[0][0], + color, + depth); +} + + + + +/** + * Do triangle cull test using tri determinant (sign indicates orientation) + * \return true if triangle is to be culled. + */ +static INLINE boolean +cull_tri(const struct setup_context *setup, float det) +{ + if (det != 0) { + /* if (det < 0 then Z points toward camera and triangle is + * counter-clockwise winding. + */ + unsigned winding = (det < 0) ? PIPE_WINDING_CCW : PIPE_WINDING_CW; + + if ((winding & setup->winding) == 0) + return FALSE; + } + + /* Culled: + */ + return TRUE; +} + + + +/** + * Clip setup->quad against the scissor/surface bounds. + */ +static INLINE void +quad_clip( struct setup_context *setup, struct quad_header *quad ) +{ + const struct pipe_scissor_state *cliprect = &setup->llvmpipe->cliprect; + const int minx = (int) cliprect->minx; + const int maxx = (int) cliprect->maxx; + const int miny = (int) cliprect->miny; + const int maxy = (int) cliprect->maxy; + + if (quad->input.x0 >= maxx || + quad->input.y0 >= maxy || + quad->input.x0 + 1 < minx || + quad->input.y0 + 1 < miny) { + /* totally clipped */ + quad->inout.mask = 0x0; + return; + } + if (quad->input.x0 < minx) + quad->inout.mask &= (MASK_BOTTOM_RIGHT | MASK_TOP_RIGHT); + if (quad->input.y0 < miny) + quad->inout.mask &= (MASK_BOTTOM_LEFT | MASK_BOTTOM_RIGHT); + if (quad->input.x0 == maxx - 1) + quad->inout.mask &= (MASK_BOTTOM_LEFT | MASK_TOP_LEFT); + if (quad->input.y0 == maxy - 1) + quad->inout.mask &= (MASK_TOP_LEFT | MASK_TOP_RIGHT); +} + + + +/** + * Given an X or Y coordinate, return the block/quad coordinate that it + * belongs to. + */ +static INLINE int block( int x ) +{ + return x & ~(2-1); +} + +static INLINE int block_x( int x ) +{ + return x & ~(TILE_VECTOR_WIDTH - 1); +} + + +/** + * Emit a quad (pass to next stage) with clipping. + */ +static INLINE void +clip_emit_quad( struct setup_context *setup, struct quad_header *quad ) +{ + quad_clip( setup, quad ); + + if (quad->inout.mask) { + struct llvmpipe_context *lp = setup->llvmpipe; + +#if 1 + /* XXX: The blender expects 4 quads. This is far from efficient, but + * until we codegenerate single-quad variants of the fragment pipeline + * we need this hack. */ + const unsigned nr_quads = TILE_VECTOR_HEIGHT*TILE_VECTOR_WIDTH/QUAD_SIZE; + struct quad_header quads[nr_quads]; + struct quad_header *quad_ptrs[nr_quads]; + int x0 = block_x(quad->input.x0); + unsigned i; + + for(i = 0; i < nr_quads; ++i) { + int x = x0 + 2*i; + if(x == quad->input.x0) + memcpy(&quads[i], quad, sizeof quads[i]); + else { + memset(&quads[i], 0, sizeof quads[i]); + quads[i].input.x0 = x; + quads[i].input.y0 = quad->input.y0; + quads[i].coef = quad->coef; + } + quad_ptrs[i] = &quads[i]; + } + + shade_quads( lp, quad_ptrs, nr_quads ); +#else + shade_quads( lp, &quad, 1 ); +#endif + } +} + + +/** + * Render a horizontal span of quads + */ +static void flush_spans( struct setup_context *setup ) +{ + const int step = TILE_VECTOR_WIDTH; + const int xleft0 = setup->span.left[0]; + const int xleft1 = setup->span.left[1]; + const int xright0 = setup->span.right[0]; + const int xright1 = setup->span.right[1]; + + + int minleft = block_x(MIN2(xleft0, xleft1)); + int maxright = MAX2(xright0, xright1); + int x; + + for (x = minleft; x < maxright; x += step) { + unsigned skip_left0 = CLAMP(xleft0 - x, 0, step); + unsigned skip_left1 = CLAMP(xleft1 - x, 0, step); + unsigned skip_right0 = CLAMP(x + step - xright0, 0, step); + unsigned skip_right1 = CLAMP(x + step - xright1, 0, step); + unsigned lx = x; + const unsigned nr_quads = TILE_VECTOR_HEIGHT*TILE_VECTOR_WIDTH/QUAD_SIZE; + unsigned q = 0; + + unsigned skipmask_left0 = (1U << skip_left0) - 1U; + unsigned skipmask_left1 = (1U << skip_left1) - 1U; + + /* These calculations fail when step == 32 and skip_right == 0. + */ + unsigned skipmask_right0 = ~0U << (unsigned)(step - skip_right0); + unsigned skipmask_right1 = ~0U << (unsigned)(step - skip_right1); + + unsigned mask0 = ~skipmask_left0 & ~skipmask_right0; + unsigned mask1 = ~skipmask_left1 & ~skipmask_right1; + + if (mask0 | mask1) { + for(q = 0; q < nr_quads; ++q) { + unsigned quadmask = (mask0 & 3) | ((mask1 & 3) << 2); + setup->quad[q].input.x0 = lx; + setup->quad[q].input.y0 = setup->span.y; + setup->quad[q].inout.mask = quadmask; + setup->quad_ptrs[q] = &setup->quad[q]; + mask0 >>= 2; + mask1 >>= 2; + lx += 2; + } + assert(!(mask0 | mask1)); + + shade_quads(setup->llvmpipe, setup->quad_ptrs, nr_quads ); + } + } + + + setup->span.y = 0; + setup->span.right[0] = 0; + setup->span.right[1] = 0; + setup->span.left[0] = 1000000; /* greater than right[0] */ + setup->span.left[1] = 1000000; /* greater than right[1] */ +} + + +#if DEBUG_VERTS +static void print_vertex(const struct setup_context *setup, + const float (*v)[4]) +{ + int i; + debug_printf(" Vertex: (%p)\n", v); + for (i = 0; i < setup->quad[0].nr_attrs; i++) { + debug_printf(" %d: %f %f %f %f\n", i, + v[i][0], v[i][1], v[i][2], v[i][3]); + if (util_is_inf_or_nan(v[i][0])) { + debug_printf(" NaN!\n"); + } + } +} +#endif + +/** + * Sort the vertices from top to bottom order, setting up the triangle + * edge fields (ebot, emaj, etop). + * \return FALSE if coords are inf/nan (cull the tri), TRUE otherwise + */ +static boolean setup_sort_vertices( struct setup_context *setup, + float det, + const float (*v0)[4], + const float (*v1)[4], + const float (*v2)[4] ) +{ + setup->vprovoke = v2; + + /* determine bottom to top order of vertices */ + { + float y0 = v0[0][1]; + float y1 = v1[0][1]; + float y2 = v2[0][1]; + if (y0 <= y1) { + if (y1 <= y2) { + /* y0<=y1<=y2 */ + setup->vmin = v0; + setup->vmid = v1; + setup->vmax = v2; + } + else if (y2 <= y0) { + /* y2<=y0<=y1 */ + setup->vmin = v2; + setup->vmid = v0; + setup->vmax = v1; + } + else { + /* y0<=y2<=y1 */ + setup->vmin = v0; + setup->vmid = v2; + setup->vmax = v1; + } + } + else { + if (y0 <= y2) { + /* y1<=y0<=y2 */ + setup->vmin = v1; + setup->vmid = v0; + setup->vmax = v2; + } + else if (y2 <= y1) { + /* y2<=y1<=y0 */ + setup->vmin = v2; + setup->vmid = v1; + setup->vmax = v0; + } + else { + /* y1<=y2<=y0 */ + setup->vmin = v1; + setup->vmid = v2; + setup->vmax = v0; + } + } + } + + setup->ebot.dx = setup->vmid[0][0] - setup->vmin[0][0]; + setup->ebot.dy = setup->vmid[0][1] - setup->vmin[0][1]; + setup->emaj.dx = setup->vmax[0][0] - setup->vmin[0][0]; + setup->emaj.dy = setup->vmax[0][1] - setup->vmin[0][1]; + setup->etop.dx = setup->vmax[0][0] - setup->vmid[0][0]; + setup->etop.dy = setup->vmax[0][1] - setup->vmid[0][1]; + + /* + * Compute triangle's area. Use 1/area to compute partial + * derivatives of attributes later. + * + * The area will be the same as prim->det, but the sign may be + * different depending on how the vertices get sorted above. + * + * To determine whether the primitive is front or back facing we + * use the prim->det value because its sign is correct. + */ + { + const float area = (setup->emaj.dx * setup->ebot.dy - + setup->ebot.dx * setup->emaj.dy); + + setup->oneoverarea = 1.0f / area; + + /* + debug_printf("%s one-over-area %f area %f det %f\n", + __FUNCTION__, setup->oneoverarea, area, det ); + */ + if (util_is_inf_or_nan(setup->oneoverarea)) + return FALSE; + } + + /* We need to know if this is a front or back-facing triangle for: + * - the GLSL gl_FrontFacing fragment attribute (bool) + * - two-sided stencil test + */ + setup->facing = + ((det > 0.0) ^ + (setup->llvmpipe->rasterizer->front_winding == PIPE_WINDING_CW)); + + return TRUE; +} + + +/** + * Compute a0, dadx and dady for a linearly interpolated coefficient, + * for a triangle. + */ +static void tri_pos_coeff( struct setup_context *setup, + uint vertSlot, unsigned i) +{ + float botda = setup->vmid[vertSlot][i] - setup->vmin[vertSlot][i]; + float majda = setup->vmax[vertSlot][i] - setup->vmin[vertSlot][i]; + float a = setup->ebot.dy * majda - botda * setup->emaj.dy; + float b = setup->emaj.dx * botda - majda * setup->ebot.dx; + float dadx = a * setup->oneoverarea; + float dady = b * setup->oneoverarea; + + assert(i <= 3); + + setup->coef.dadx[0][i] = dadx; + setup->coef.dady[0][i] = dady; + + /* calculate a0 as the value which would be sampled for the + * fragment at (0,0), taking into account that we want to sample at + * pixel centers, in other words (0.5, 0.5). + * + * this is neat but unfortunately not a good way to do things for + * triangles with very large values of dadx or dady as it will + * result in the subtraction and re-addition from a0 of a very + * large number, which means we'll end up loosing a lot of the + * fractional bits and precision from a0. the way to fix this is + * to define a0 as the sample at a pixel center somewhere near vmin + * instead - i'll switch to this later. + */ + setup->coef.a0[0][i] = (setup->vmin[vertSlot][i] - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); + + /* + debug_printf("attr[%d].%c: %f dx:%f dy:%f\n", + slot, "xyzw"[i], + setup->coef[slot].a0[i], + setup->coef[slot].dadx[i], + setup->coef[slot].dady[i]); + */ +} + + +/** + * Compute a0 for a constant-valued coefficient (GL_FLAT shading). + * The value value comes from vertex[slot][i]. + * The result will be put into setup->coef[slot].a0[i]. + * \param slot which attribute slot + * \param i which component of the slot (0..3) + */ +static void const_pos_coeff( struct setup_context *setup, + uint vertSlot, unsigned i) +{ + setup->coef.dadx[0][i] = 0; + setup->coef.dady[0][i] = 0; + + /* need provoking vertex info! + */ + setup->coef.a0[0][i] = setup->vprovoke[vertSlot][i]; +} + + +/** + * Compute a0 for a constant-valued coefficient (GL_FLAT shading). + * The value value comes from vertex[slot][i]. + * The result will be put into setup->coef[slot].a0[i]. + * \param slot which attribute slot + * \param i which component of the slot (0..3) + */ +static void const_coeff( struct setup_context *setup, + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for (i = 0; i < NUM_CHANNELS; ++i) { + setup->coef.dadx[1 + attrib][i] = 0; + setup->coef.dady[1 + attrib][i] = 0; + + /* need provoking vertex info! + */ + setup->coef.a0[1 + attrib][i] = setup->vprovoke[vertSlot][i]; + } +} + + +/** + * Compute a0, dadx and dady for a linearly interpolated coefficient, + * for a triangle. + */ +static void tri_linear_coeff( struct setup_context *setup, + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for (i = 0; i < NUM_CHANNELS; ++i) { + float botda = setup->vmid[vertSlot][i] - setup->vmin[vertSlot][i]; + float majda = setup->vmax[vertSlot][i] - setup->vmin[vertSlot][i]; + float a = setup->ebot.dy * majda - botda * setup->emaj.dy; + float b = setup->emaj.dx * botda - majda * setup->ebot.dx; + float dadx = a * setup->oneoverarea; + float dady = b * setup->oneoverarea; + + assert(i <= 3); + + setup->coef.dadx[1 + attrib][i] = dadx; + setup->coef.dady[1 + attrib][i] = dady; + + /* calculate a0 as the value which would be sampled for the + * fragment at (0,0), taking into account that we want to sample at + * pixel centers, in other words (0.5, 0.5). + * + * this is neat but unfortunately not a good way to do things for + * triangles with very large values of dadx or dady as it will + * result in the subtraction and re-addition from a0 of a very + * large number, which means we'll end up loosing a lot of the + * fractional bits and precision from a0. the way to fix this is + * to define a0 as the sample at a pixel center somewhere near vmin + * instead - i'll switch to this later. + */ + setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); + + /* + debug_printf("attr[%d].%c: %f dx:%f dy:%f\n", + slot, "xyzw"[i], + setup->coef[slot].a0[i], + setup->coef[slot].dadx[i], + setup->coef[slot].dady[i]); + */ + } +} + + +/** + * Compute a0, dadx and dady for a perspective-corrected interpolant, + * for a triangle. + * We basically multiply the vertex value by 1/w before computing + * the plane coefficients (a0, dadx, dady). + * Later, when we compute the value at a particular fragment position we'll + * divide the interpolated value by the interpolated W at that fragment. + */ +static void tri_persp_coeff( struct setup_context *setup, + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for (i = 0; i < NUM_CHANNELS; ++i) { + /* premultiply by 1/w (v[0][3] is always W): + */ + float mina = setup->vmin[vertSlot][i] * setup->vmin[0][3]; + float mida = setup->vmid[vertSlot][i] * setup->vmid[0][3]; + float maxa = setup->vmax[vertSlot][i] * setup->vmax[0][3]; + float botda = mida - mina; + float majda = maxa - mina; + float a = setup->ebot.dy * majda - botda * setup->emaj.dy; + float b = setup->emaj.dx * botda - majda * setup->ebot.dx; + float dadx = a * setup->oneoverarea; + float dady = b * setup->oneoverarea; + + /* + debug_printf("tri persp %d,%d: %f %f %f\n", vertSlot, i, + setup->vmin[vertSlot][i], + setup->vmid[vertSlot][i], + setup->vmax[vertSlot][i] + ); + */ + assert(i <= 3); + + setup->coef.dadx[1 + attrib][i] = dadx; + setup->coef.dady[1 + attrib][i] = dady; + setup->coef.a0[1 + attrib][i] = (mina - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); + } +} + + +/** + * Special coefficient setup for gl_FragCoord. + * X and Y are trivial, though Y has to be inverted for OpenGL. + * Z and W are copied from posCoef which should have already been computed. + * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask. + */ +static void +setup_fragcoord_coeff(struct setup_context *setup, uint slot) +{ + /*X*/ + setup->coef.a0[1 + slot][0] = 0; + setup->coef.dadx[1 + slot][0] = 1.0; + setup->coef.dady[1 + slot][0] = 0.0; + /*Y*/ + setup->coef.a0[1 + slot][1] = 0.0; + setup->coef.dadx[1 + slot][1] = 0.0; + setup->coef.dady[1 + slot][1] = 1.0; + /*Z*/ + setup->coef.a0[1 + slot][2] = setup->coef.a0[0][2]; + setup->coef.dadx[1 + slot][2] = setup->coef.dadx[0][2]; + setup->coef.dady[1 + slot][2] = setup->coef.dady[0][2]; + /*W*/ + setup->coef.a0[1 + slot][3] = setup->coef.a0[0][3]; + setup->coef.dadx[1 + slot][3] = setup->coef.dadx[0][3]; + setup->coef.dady[1 + slot][3] = setup->coef.dady[0][3]; +} + + + +/** + * Compute the setup->coef[] array dadx, dady, a0 values. + * Must be called after setup->vmin,vmid,vmax,vprovoke are initialized. + */ +static void setup_tri_coefficients( struct setup_context *setup ) +{ + struct llvmpipe_context *llvmpipe = setup->llvmpipe; + const struct lp_fragment_shader *lpfs = llvmpipe->fs; + const struct vertex_info *vinfo = llvmpipe_get_vertex_info(llvmpipe); + uint fragSlot; + + /* z and w are done by linear interpolation: + */ + tri_pos_coeff(setup, 0, 2); + tri_pos_coeff(setup, 0, 3); + + /* setup interpolation for all the remaining attributes: + */ + for (fragSlot = 0; fragSlot < lpfs->info.num_inputs; fragSlot++) { + const uint vertSlot = vinfo->attrib[fragSlot].src_index; + + switch (vinfo->attrib[fragSlot].interp_mode) { + case INTERP_CONSTANT: + const_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_LINEAR: + tri_linear_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_PERSPECTIVE: + tri_persp_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_POS: + setup_fragcoord_coeff(setup, fragSlot); + break; + default: + assert(0); + } + + if (lpfs->info.input_semantic_name[fragSlot] == TGSI_SEMANTIC_FACE) { + setup->coef.a0[1 + fragSlot][0] = 1.0f - setup->facing; + setup->coef.dadx[1 + fragSlot][0] = 0.0; + setup->coef.dady[1 + fragSlot][0] = 0.0; + } + } +} + + + +static void setup_tri_edges( struct setup_context *setup ) +{ + float vmin_x = setup->vmin[0][0] + 0.5f; + float vmid_x = setup->vmid[0][0] + 0.5f; + + float vmin_y = setup->vmin[0][1] - 0.5f; + float vmid_y = setup->vmid[0][1] - 0.5f; + float vmax_y = setup->vmax[0][1] - 0.5f; + + setup->emaj.sy = ceilf(vmin_y); + setup->emaj.lines = (int) ceilf(vmax_y - setup->emaj.sy); + setup->emaj.dxdy = setup->emaj.dx / setup->emaj.dy; + setup->emaj.sx = vmin_x + (setup->emaj.sy - vmin_y) * setup->emaj.dxdy; + + setup->etop.sy = ceilf(vmid_y); + setup->etop.lines = (int) ceilf(vmax_y - setup->etop.sy); + setup->etop.dxdy = setup->etop.dx / setup->etop.dy; + setup->etop.sx = vmid_x + (setup->etop.sy - vmid_y) * setup->etop.dxdy; + + setup->ebot.sy = ceilf(vmin_y); + setup->ebot.lines = (int) ceilf(vmid_y - setup->ebot.sy); + setup->ebot.dxdy = setup->ebot.dx / setup->ebot.dy; + setup->ebot.sx = vmin_x + (setup->ebot.sy - vmin_y) * setup->ebot.dxdy; +} + + +/** + * Render the upper or lower half of a triangle. + * Scissoring/cliprect is applied here too. + */ +static void subtriangle( struct setup_context *setup, + struct edge *eleft, + struct edge *eright, + unsigned lines ) +{ + const struct pipe_scissor_state *cliprect = &setup->llvmpipe->cliprect; + const int minx = (int) cliprect->minx; + const int maxx = (int) cliprect->maxx; + const int miny = (int) cliprect->miny; + const int maxy = (int) cliprect->maxy; + int y, start_y, finish_y; + int sy = (int)eleft->sy; + + assert((int)eleft->sy == (int) eright->sy); + + /* clip top/bottom */ + start_y = sy; + if (start_y < miny) + start_y = miny; + + finish_y = sy + lines; + if (finish_y > maxy) + finish_y = maxy; + + start_y -= sy; + finish_y -= sy; + + /* + debug_printf("%s %d %d\n", __FUNCTION__, start_y, finish_y); + */ + + for (y = start_y; y < finish_y; y++) { + + /* avoid accumulating adds as floats don't have the precision to + * accurately iterate large triangle edges that way. luckily we + * can just multiply these days. + * + * this is all drowned out by the attribute interpolation anyway. + */ + int left = (int)(eleft->sx + y * eleft->dxdy); + int right = (int)(eright->sx + y * eright->dxdy); + + /* clip left/right */ + if (left < minx) + left = minx; + if (right > maxx) + right = maxx; + + if (left < right) { + int _y = sy + y; + if (block(_y) != setup->span.y) { + flush_spans(setup); + setup->span.y = block(_y); + } + + setup->span.left[_y&1] = left; + setup->span.right[_y&1] = right; + } + } + + + /* save the values so that emaj can be restarted: + */ + eleft->sx += lines * eleft->dxdy; + eright->sx += lines * eright->dxdy; + eleft->sy += lines; + eright->sy += lines; +} + + +/** + * Recalculate prim's determinant. This is needed as we don't have + * get this information through the vbuf_render interface & we must + * calculate it here. + */ +static float +calc_det( const float (*v0)[4], + const float (*v1)[4], + const float (*v2)[4] ) +{ + /* edge vectors e = v0 - v2, f = v1 - v2 */ + const float ex = v0[0][0] - v2[0][0]; + const float ey = v0[0][1] - v2[0][1]; + const float fx = v1[0][0] - v2[0][0]; + const float fy = v1[0][1] - v2[0][1]; + + /* det = cross(e,f).z */ + return ex * fy - ey * fx; +} + + +/** + * Do setup for triangle rasterization, then render the triangle. + */ +void llvmpipe_setup_tri( struct setup_context *setup, + const float (*v0)[4], + const float (*v1)[4], + const float (*v2)[4] ) +{ + float det; + +#if DEBUG_VERTS + debug_printf("Setup triangle:\n"); + print_vertex(setup, v0); + print_vertex(setup, v1); + print_vertex(setup, v2); +#endif + + if (setup->llvmpipe->no_rast) + return; + + det = calc_det(v0, v1, v2); + /* + debug_printf("%s\n", __FUNCTION__ ); + */ + +#if DEBUG_FRAGS + setup->numFragsEmitted = 0; + setup->numFragsWritten = 0; +#endif + + if (cull_tri( setup, det )) + return; + + if (!setup_sort_vertices( setup, det, v0, v1, v2 )) + return; + setup_tri_coefficients( setup ); + setup_tri_edges( setup ); + + assert(setup->llvmpipe->reduced_prim == PIPE_PRIM_TRIANGLES); + + setup->span.y = 0; + setup->span.right[0] = 0; + setup->span.right[1] = 0; + /* setup->span.z_mode = tri_z_mode( setup->ctx ); */ + + /* init_constant_attribs( setup ); */ + + if (setup->oneoverarea < 0.0) { + /* emaj on left: + */ + subtriangle( setup, &setup->emaj, &setup->ebot, setup->ebot.lines ); + subtriangle( setup, &setup->emaj, &setup->etop, setup->etop.lines ); + } + else { + /* emaj on right: + */ + subtriangle( setup, &setup->ebot, &setup->emaj, setup->ebot.lines ); + subtriangle( setup, &setup->etop, &setup->emaj, setup->etop.lines ); + } + + flush_spans( setup ); + +#if DEBUG_FRAGS + printf("Tri: %u frags emitted, %u written\n", + setup->numFragsEmitted, + setup->numFragsWritten); +#endif +} + + + +/** + * Compute a0, dadx and dady for a linearly interpolated coefficient, + * for a line. + */ +static void +linear_pos_coeff(struct setup_context *setup, + uint vertSlot, uint i) +{ + const float da = setup->vmax[vertSlot][i] - setup->vmin[vertSlot][i]; + const float dadx = da * setup->emaj.dx * setup->oneoverarea; + const float dady = da * setup->emaj.dy * setup->oneoverarea; + setup->coef.dadx[0][i] = dadx; + setup->coef.dady[0][i] = dady; + setup->coef.a0[0][i] = (setup->vmin[vertSlot][i] - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); +} + + +/** + * Compute a0, dadx and dady for a linearly interpolated coefficient, + * for a line. + */ +static void +line_linear_coeff(struct setup_context *setup, + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for (i = 0; i < NUM_CHANNELS; ++i) { + const float da = setup->vmax[vertSlot][i] - setup->vmin[vertSlot][i]; + const float dadx = da * setup->emaj.dx * setup->oneoverarea; + const float dady = da * setup->emaj.dy * setup->oneoverarea; + setup->coef.dadx[1 + attrib][i] = dadx; + setup->coef.dady[1 + attrib][i] = dady; + setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); + } +} + + +/** + * Compute a0, dadx and dady for a perspective-corrected interpolant, + * for a line. + */ +static void +line_persp_coeff(struct setup_context *setup, + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for (i = 0; i < NUM_CHANNELS; ++i) { + /* XXX double-check/verify this arithmetic */ + const float a0 = setup->vmin[vertSlot][i] * setup->vmin[0][3]; + const float a1 = setup->vmax[vertSlot][i] * setup->vmax[0][3]; + const float da = a1 - a0; + const float dadx = da * setup->emaj.dx * setup->oneoverarea; + const float dady = da * setup->emaj.dy * setup->oneoverarea; + setup->coef.dadx[1 + attrib][i] = dadx; + setup->coef.dady[1 + attrib][i] = dady; + setup->coef.a0[1 + attrib][i] = (setup->vmin[vertSlot][i] - + (dadx * (setup->vmin[0][0] - 0.5f) + + dady * (setup->vmin[0][1] - 0.5f))); + } +} + + +/** + * Compute the setup->coef[] array dadx, dady, a0 values. + * Must be called after setup->vmin,vmax are initialized. + */ +static INLINE boolean +setup_line_coefficients(struct setup_context *setup, + const float (*v0)[4], + const float (*v1)[4]) +{ + struct llvmpipe_context *llvmpipe = setup->llvmpipe; + const struct lp_fragment_shader *lpfs = llvmpipe->fs; + const struct vertex_info *vinfo = llvmpipe_get_vertex_info(llvmpipe); + uint fragSlot; + float area; + + /* use setup->vmin, vmax to point to vertices */ + if (llvmpipe->rasterizer->flatshade_first) + setup->vprovoke = v0; + else + setup->vprovoke = v1; + setup->vmin = v0; + setup->vmax = v1; + + setup->emaj.dx = setup->vmax[0][0] - setup->vmin[0][0]; + setup->emaj.dy = setup->vmax[0][1] - setup->vmin[0][1]; + + /* NOTE: this is not really area but something proportional to it */ + area = setup->emaj.dx * setup->emaj.dx + setup->emaj.dy * setup->emaj.dy; + if (area == 0.0f || util_is_inf_or_nan(area)) + return FALSE; + setup->oneoverarea = 1.0f / area; + + /* z and w are done by linear interpolation: + */ + linear_pos_coeff(setup, 0, 2); + linear_pos_coeff(setup, 0, 3); + + /* setup interpolation for all the remaining attributes: + */ + for (fragSlot = 0; fragSlot < lpfs->info.num_inputs; fragSlot++) { + const uint vertSlot = vinfo->attrib[fragSlot].src_index; + + switch (vinfo->attrib[fragSlot].interp_mode) { + case INTERP_CONSTANT: + const_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_LINEAR: + line_linear_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_PERSPECTIVE: + line_persp_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_POS: + setup_fragcoord_coeff(setup, fragSlot); + break; + default: + assert(0); + } + + if (lpfs->info.input_semantic_name[fragSlot] == TGSI_SEMANTIC_FACE) { + setup->coef.a0[1 + fragSlot][0] = 1.0f - setup->facing; + setup->coef.dadx[1 + fragSlot][0] = 0.0; + setup->coef.dady[1 + fragSlot][0] = 0.0; + } + } + return TRUE; +} + + +/** + * Plot a pixel in a line segment. + */ +static INLINE void +plot(struct setup_context *setup, int x, int y) +{ + const int iy = y & 1; + const int ix = x & 1; + const int quadX = x - ix; + const int quadY = y - iy; + const int mask = (1 << ix) << (2 * iy); + + if (quadX != setup->quad[0].input.x0 || + quadY != setup->quad[0].input.y0) + { + /* flush prev quad, start new quad */ + + if (setup->quad[0].input.x0 != -1) + clip_emit_quad( setup, &setup->quad[0] ); + + setup->quad[0].input.x0 = quadX; + setup->quad[0].input.y0 = quadY; + setup->quad[0].inout.mask = 0x0; + } + + setup->quad[0].inout.mask |= mask; +} + + +/** + * Do setup for line rasterization, then render the line. + * Single-pixel width, no stipple, etc. We rely on the 'draw' module + * to handle stippling and wide lines. + */ +void +llvmpipe_setup_line(struct setup_context *setup, + const float (*v0)[4], + const float (*v1)[4]) +{ + int x0 = (int) v0[0][0]; + int x1 = (int) v1[0][0]; + int y0 = (int) v0[0][1]; + int y1 = (int) v1[0][1]; + int dx = x1 - x0; + int dy = y1 - y0; + int xstep, ystep; + +#if DEBUG_VERTS + debug_printf("Setup line:\n"); + print_vertex(setup, v0); + print_vertex(setup, v1); +#endif + + if (setup->llvmpipe->no_rast) + return; + + if (dx == 0 && dy == 0) + return; + + if (!setup_line_coefficients(setup, v0, v1)) + return; + + assert(v0[0][0] < 1.0e9); + assert(v0[0][1] < 1.0e9); + assert(v1[0][0] < 1.0e9); + assert(v1[0][1] < 1.0e9); + + if (dx < 0) { + dx = -dx; /* make positive */ + xstep = -1; + } + else { + xstep = 1; + } + + if (dy < 0) { + dy = -dy; /* make positive */ + ystep = -1; + } + else { + ystep = 1; + } + + assert(dx >= 0); + assert(dy >= 0); + assert(setup->llvmpipe->reduced_prim == PIPE_PRIM_LINES); + + setup->quad[0].input.x0 = setup->quad[0].input.y0 = -1; + setup->quad[0].inout.mask = 0x0; + + /* XXX temporary: set coverage to 1.0 so the line appears + * if AA mode happens to be enabled. + */ + setup->quad[0].input.coverage[0] = + setup->quad[0].input.coverage[1] = + setup->quad[0].input.coverage[2] = + setup->quad[0].input.coverage[3] = 1.0; + + if (dx > dy) { + /*** X-major line ***/ + int i; + const int errorInc = dy + dy; + int error = errorInc - dx; + const int errorDec = error - dx; + + for (i = 0; i < dx; i++) { + plot(setup, x0, y0); + + x0 += xstep; + if (error < 0) { + error += errorInc; + } + else { + error += errorDec; + y0 += ystep; + } + } + } + else { + /*** Y-major line ***/ + int i; + const int errorInc = dx + dx; + int error = errorInc - dy; + const int errorDec = error - dy; + + for (i = 0; i < dy; i++) { + plot(setup, x0, y0); + + y0 += ystep; + if (error < 0) { + error += errorInc; + } + else { + error += errorDec; + x0 += xstep; + } + } + } + + /* draw final quad */ + if (setup->quad[0].inout.mask) { + clip_emit_quad( setup, &setup->quad[0] ); + } +} + + +static void +point_persp_coeff(struct setup_context *setup, + const float (*vert)[4], + unsigned attrib, + uint vertSlot) +{ + unsigned i; + for(i = 0; i < NUM_CHANNELS; ++i) { + setup->coef.dadx[1 + attrib][i] = 0.0F; + setup->coef.dady[1 + attrib][i] = 0.0F; + setup->coef.a0[1 + attrib][i] = vert[vertSlot][i] * vert[0][3]; + } +} + + +/** + * Do setup for point rasterization, then render the point. + * Round or square points... + * XXX could optimize a lot for 1-pixel points. + */ +void +llvmpipe_setup_point( struct setup_context *setup, + const float (*v0)[4] ) +{ + struct llvmpipe_context *llvmpipe = setup->llvmpipe; + const struct lp_fragment_shader *lpfs = llvmpipe->fs; + const int sizeAttr = setup->llvmpipe->psize_slot; + const float size + = sizeAttr > 0 ? v0[sizeAttr][0] + : setup->llvmpipe->rasterizer->point_size; + const float halfSize = 0.5F * size; + const boolean round = (boolean) setup->llvmpipe->rasterizer->point_smooth; + const float x = v0[0][0]; /* Note: data[0] is always position */ + const float y = v0[0][1]; + const struct vertex_info *vinfo = llvmpipe_get_vertex_info(llvmpipe); + uint fragSlot; + +#if DEBUG_VERTS + debug_printf("Setup point:\n"); + print_vertex(setup, v0); +#endif + + if (llvmpipe->no_rast) + return; + + assert(setup->llvmpipe->reduced_prim == PIPE_PRIM_POINTS); + + /* For points, all interpolants are constant-valued. + * However, for point sprites, we'll need to setup texcoords appropriately. + * XXX: which coefficients are the texcoords??? + * We may do point sprites as textured quads... + * + * KW: We don't know which coefficients are texcoords - ultimately + * the choice of what interpolation mode to use for each attribute + * should be determined by the fragment program, using + * per-attribute declaration statements that include interpolation + * mode as a parameter. So either the fragment program will have + * to be adjusted for pointsprite vs normal point behaviour, or + * otherwise a special interpolation mode will have to be defined + * which matches the required behaviour for point sprites. But - + * the latter is not a feature of normal hardware, and as such + * probably should be ruled out on that basis. + */ + setup->vprovoke = v0; + + /* setup Z, W */ + const_pos_coeff(setup, 0, 2); + const_pos_coeff(setup, 0, 3); + + for (fragSlot = 0; fragSlot < lpfs->info.num_inputs; fragSlot++) { + const uint vertSlot = vinfo->attrib[fragSlot].src_index; + + switch (vinfo->attrib[fragSlot].interp_mode) { + case INTERP_CONSTANT: + /* fall-through */ + case INTERP_LINEAR: + const_coeff(setup, fragSlot, vertSlot); + break; + case INTERP_PERSPECTIVE: + point_persp_coeff(setup, setup->vprovoke, fragSlot, vertSlot); + break; + case INTERP_POS: + setup_fragcoord_coeff(setup, fragSlot); + break; + default: + assert(0); + } + + if (lpfs->info.input_semantic_name[fragSlot] == TGSI_SEMANTIC_FACE) { + setup->coef.a0[1 + fragSlot][0] = 1.0f - setup->facing; + setup->coef.dadx[1 + fragSlot][0] = 0.0; + setup->coef.dady[1 + fragSlot][0] = 0.0; + } + } + + + if (halfSize <= 0.5 && !round) { + /* special case for 1-pixel points */ + const int ix = ((int) x) & 1; + const int iy = ((int) y) & 1; + setup->quad[0].input.x0 = (int) x - ix; + setup->quad[0].input.y0 = (int) y - iy; + setup->quad[0].inout.mask = (1 << ix) << (2 * iy); + clip_emit_quad( setup, &setup->quad[0] ); + } + else { + if (round) { + /* rounded points */ + const int ixmin = block((int) (x - halfSize)); + const int ixmax = block((int) (x + halfSize)); + const int iymin = block((int) (y - halfSize)); + const int iymax = block((int) (y + halfSize)); + const float rmin = halfSize - 0.7071F; /* 0.7071 = sqrt(2)/2 */ + const float rmax = halfSize + 0.7071F; + const float rmin2 = MAX2(0.0F, rmin * rmin); + const float rmax2 = rmax * rmax; + const float cscale = 1.0F / (rmax2 - rmin2); + int ix, iy; + + for (iy = iymin; iy <= iymax; iy += 2) { + for (ix = ixmin; ix <= ixmax; ix += 2) { + float dx, dy, dist2, cover; + + setup->quad[0].inout.mask = 0x0; + + dx = (ix + 0.5f) - x; + dy = (iy + 0.5f) - y; + dist2 = dx * dx + dy * dy; + if (dist2 <= rmax2) { + cover = 1.0F - (dist2 - rmin2) * cscale; + setup->quad[0].input.coverage[QUAD_TOP_LEFT] = MIN2(cover, 1.0f); + setup->quad[0].inout.mask |= MASK_TOP_LEFT; + } + + dx = (ix + 1.5f) - x; + dy = (iy + 0.5f) - y; + dist2 = dx * dx + dy * dy; + if (dist2 <= rmax2) { + cover = 1.0F - (dist2 - rmin2) * cscale; + setup->quad[0].input.coverage[QUAD_TOP_RIGHT] = MIN2(cover, 1.0f); + setup->quad[0].inout.mask |= MASK_TOP_RIGHT; + } + + dx = (ix + 0.5f) - x; + dy = (iy + 1.5f) - y; + dist2 = dx * dx + dy * dy; + if (dist2 <= rmax2) { + cover = 1.0F - (dist2 - rmin2) * cscale; + setup->quad[0].input.coverage[QUAD_BOTTOM_LEFT] = MIN2(cover, 1.0f); + setup->quad[0].inout.mask |= MASK_BOTTOM_LEFT; + } + + dx = (ix + 1.5f) - x; + dy = (iy + 1.5f) - y; + dist2 = dx * dx + dy * dy; + if (dist2 <= rmax2) { + cover = 1.0F - (dist2 - rmin2) * cscale; + setup->quad[0].input.coverage[QUAD_BOTTOM_RIGHT] = MIN2(cover, 1.0f); + setup->quad[0].inout.mask |= MASK_BOTTOM_RIGHT; + } + + if (setup->quad[0].inout.mask) { + setup->quad[0].input.x0 = ix; + setup->quad[0].input.y0 = iy; + clip_emit_quad( setup, &setup->quad[0] ); + } + } + } + } + else { + /* square points */ + const int xmin = (int) (x + 0.75 - halfSize); + const int ymin = (int) (y + 0.25 - halfSize); + const int xmax = xmin + (int) size; + const int ymax = ymin + (int) size; + /* XXX could apply scissor to xmin,ymin,xmax,ymax now */ + const int ixmin = block(xmin); + const int ixmax = block(xmax - 1); + const int iymin = block(ymin); + const int iymax = block(ymax - 1); + int ix, iy; + + /* + debug_printf("(%f, %f) -> X:%d..%d Y:%d..%d\n", x, y, xmin, xmax,ymin,ymax); + */ + for (iy = iymin; iy <= iymax; iy += 2) { + uint rowMask = 0xf; + if (iy < ymin) { + /* above the top edge */ + rowMask &= (MASK_BOTTOM_LEFT | MASK_BOTTOM_RIGHT); + } + if (iy + 1 >= ymax) { + /* below the bottom edge */ + rowMask &= (MASK_TOP_LEFT | MASK_TOP_RIGHT); + } + + for (ix = ixmin; ix <= ixmax; ix += 2) { + uint mask = rowMask; + + if (ix < xmin) { + /* fragment is past left edge of point, turn off left bits */ + mask &= (MASK_BOTTOM_RIGHT | MASK_TOP_RIGHT); + } + if (ix + 1 >= xmax) { + /* past the right edge */ + mask &= (MASK_BOTTOM_LEFT | MASK_TOP_LEFT); + } + + setup->quad[0].inout.mask = mask; + setup->quad[0].input.x0 = ix; + setup->quad[0].input.y0 = iy; + clip_emit_quad( setup, &setup->quad[0] ); + } + } + } + } +} + +void llvmpipe_setup_prepare( struct setup_context *setup ) +{ + struct llvmpipe_context *lp = setup->llvmpipe; + + if (lp->dirty) { + llvmpipe_update_derived(lp); + } + + if (lp->reduced_api_prim == PIPE_PRIM_TRIANGLES && + lp->rasterizer->fill_cw == PIPE_POLYGON_MODE_FILL && + lp->rasterizer->fill_ccw == PIPE_POLYGON_MODE_FILL) { + /* we'll do culling */ + setup->winding = lp->rasterizer->cull_mode; + } + else { + /* 'draw' will do culling */ + setup->winding = PIPE_WINDING_NONE; + } +} + + + +void llvmpipe_setup_destroy_context( struct setup_context *setup ) +{ + align_free( setup ); +} + + +/** + * Create a new primitive setup/render stage. + */ +struct setup_context *llvmpipe_setup_create_context( struct llvmpipe_context *llvmpipe ) +{ + struct setup_context *setup; + unsigned i; + + setup = align_malloc(sizeof(struct setup_context), 16); + if (!setup) + return NULL; + + memset(setup, 0, sizeof *setup); + setup->llvmpipe = llvmpipe; + + for (i = 0; i < MAX_QUADS; i++) { + setup->quad[i].coef = &setup->coef; + } + + setup->span.left[0] = 1000000; /* greater than right[0] */ + setup->span.left[1] = 1000000; /* greater than right[1] */ + + return setup; +} + diff --git a/src/gallium/drivers/llvmpipe/lp_setup.h b/src/gallium/drivers/llvmpipe/lp_setup.h new file mode 100644 index 00000000000..89c43da0460 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_setup.h @@ -0,0 +1,53 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ +#ifndef LP_SETUP_H +#define LP_SETUP_H + +struct setup_context; +struct llvmpipe_context; + +void +llvmpipe_setup_tri( struct setup_context *setup, + const float (*v0)[4], + const float (*v1)[4], + const float (*v2)[4] ); + +void +llvmpipe_setup_line(struct setup_context *setup, + const float (*v0)[4], + const float (*v1)[4]); + +void +llvmpipe_setup_point( struct setup_context *setup, + const float (*v0)[4] ); + + +struct setup_context *llvmpipe_setup_create_context( struct llvmpipe_context *llvmpipe ); +void llvmpipe_setup_prepare( struct setup_context *setup ); +void llvmpipe_setup_destroy_context( struct setup_context *setup ); + +#endif diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h new file mode 100644 index 00000000000..fb10329887d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state.h @@ -0,0 +1,225 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + +#ifndef LP_STATE_H +#define LP_STATE_H + +#include <llvm-c/Core.h> + +#include "pipe/p_state.h" +#include "tgsi/tgsi_scan.h" +#include "lp_jit.h" + + +#define LP_NEW_VIEWPORT 0x1 +#define LP_NEW_RASTERIZER 0x2 +#define LP_NEW_FS 0x4 +#define LP_NEW_BLEND 0x8 +#define LP_NEW_CLIP 0x10 +#define LP_NEW_SCISSOR 0x20 +#define LP_NEW_STIPPLE 0x40 +#define LP_NEW_FRAMEBUFFER 0x80 +#define LP_NEW_DEPTH_STENCIL_ALPHA 0x100 +#define LP_NEW_CONSTANTS 0x200 +#define LP_NEW_SAMPLER 0x400 +#define LP_NEW_TEXTURE 0x800 +#define LP_NEW_VERTEX 0x1000 +#define LP_NEW_VS 0x2000 +#define LP_NEW_QUERY 0x4000 + + +struct tgsi_sampler; +struct vertex_info; + + +struct lp_fragment_shader; + + +struct lp_fragment_shader_variant_key +{ + struct pipe_depth_state depth; + struct pipe_alpha_state alpha; + struct pipe_blend_state blend; +}; + + +struct lp_fragment_shader_variant +{ + struct lp_fragment_shader *shader; + + struct lp_fragment_shader_variant_key key; + + LLVMValueRef function; + + lp_jit_frag_func jit_function; + + struct lp_fragment_shader_variant *next; +}; + + +/** + * Subclass of pipe_shader_state (though it doesn't really need to be). + * + * This is starting to look an awful lot like a quad pipeline stage... + */ +struct lp_fragment_shader +{ + struct pipe_shader_state base; + + struct tgsi_shader_info info; + + struct lp_fragment_shader_variant *variants; + + struct lp_fragment_shader_variant *current; +}; + + +/** Subclass of pipe_shader_state */ +struct lp_vertex_shader { + struct pipe_shader_state shader; + struct draw_vertex_shader *draw_data; +}; + + + +void * +llvmpipe_create_blend_state(struct pipe_context *, + const struct pipe_blend_state *); +void llvmpipe_bind_blend_state(struct pipe_context *, + void *); +void llvmpipe_delete_blend_state(struct pipe_context *, + void *); + +void * +llvmpipe_create_sampler_state(struct pipe_context *, + const struct pipe_sampler_state *); +void llvmpipe_bind_sampler_states(struct pipe_context *, unsigned, void **); +void llvmpipe_delete_sampler_state(struct pipe_context *, void *); + +void * +llvmpipe_create_depth_stencil_state(struct pipe_context *, + const struct pipe_depth_stencil_alpha_state *); +void llvmpipe_bind_depth_stencil_state(struct pipe_context *, void *); +void llvmpipe_delete_depth_stencil_state(struct pipe_context *, void *); + +void * +llvmpipe_create_rasterizer_state(struct pipe_context *, + const struct pipe_rasterizer_state *); +void llvmpipe_bind_rasterizer_state(struct pipe_context *, void *); +void llvmpipe_delete_rasterizer_state(struct pipe_context *, void *); + +void llvmpipe_set_framebuffer_state( struct pipe_context *, + const struct pipe_framebuffer_state * ); + +void llvmpipe_set_blend_color( struct pipe_context *pipe, + const struct pipe_blend_color *blend_color ); + +void llvmpipe_set_clip_state( struct pipe_context *, + const struct pipe_clip_state * ); + +void llvmpipe_set_constant_buffer(struct pipe_context *, + uint shader, uint index, + const struct pipe_constant_buffer *buf); + +void *llvmpipe_create_fs_state(struct pipe_context *, + const struct pipe_shader_state *); +void llvmpipe_bind_fs_state(struct pipe_context *, void *); +void llvmpipe_delete_fs_state(struct pipe_context *, void *); +void *llvmpipe_create_vs_state(struct pipe_context *, + const struct pipe_shader_state *); +void llvmpipe_bind_vs_state(struct pipe_context *, void *); +void llvmpipe_delete_vs_state(struct pipe_context *, void *); + +void llvmpipe_set_polygon_stipple( struct pipe_context *, + const struct pipe_poly_stipple * ); + +void llvmpipe_set_scissor_state( struct pipe_context *, + const struct pipe_scissor_state * ); + +void llvmpipe_set_sampler_textures( struct pipe_context *, + unsigned num, + struct pipe_texture ** ); + +void llvmpipe_set_viewport_state( struct pipe_context *, + const struct pipe_viewport_state * ); + +void llvmpipe_set_vertex_elements(struct pipe_context *, + unsigned count, + const struct pipe_vertex_element *); + +void llvmpipe_set_vertex_buffers(struct pipe_context *, + unsigned count, + const struct pipe_vertex_buffer *); + +void llvmpipe_update_fs(struct llvmpipe_context *lp); + +void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe ); + + +boolean llvmpipe_draw_arrays(struct pipe_context *pipe, unsigned mode, + unsigned start, unsigned count); + +boolean llvmpipe_draw_elements(struct pipe_context *pipe, + struct pipe_buffer *indexBuffer, + unsigned indexSize, + unsigned mode, unsigned start, unsigned count); +boolean +llvmpipe_draw_range_elements(struct pipe_context *pipe, + struct pipe_buffer *indexBuffer, + unsigned indexSize, + unsigned min_index, + unsigned max_index, + unsigned mode, unsigned start, unsigned count); + +void +llvmpipe_set_edgeflags(struct pipe_context *pipe, const unsigned *edgeflags); + + +void +llvmpipe_map_transfers(struct llvmpipe_context *lp); + +void +llvmpipe_unmap_transfers(struct llvmpipe_context *lp); + +void +llvmpipe_map_texture_surfaces(struct llvmpipe_context *lp); + +void +llvmpipe_unmap_texture_surfaces(struct llvmpipe_context *lp); + + +struct vertex_info * +llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe); + +struct vertex_info * +llvmpipe_get_vbuf_vertex_info(struct llvmpipe_context *llvmpipe); + + +#endif diff --git a/src/gallium/drivers/llvmpipe/lp_state_blend.c b/src/gallium/drivers/llvmpipe/lp_state_blend.c new file mode 100644 index 00000000000..3f03bd00571 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_blend.c @@ -0,0 +1,114 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @author Jose Fonseca <jfonseca@vmware.com> + * @author Keith Whitwell <keith@tungstengraphics.com> + */ + +#include "util/u_memory.h" +#include "util/u_math.h" +#include "util/u_debug_dump.h" +#include "lp_screen.h" +#include "lp_context.h" +#include "lp_state.h" + + +void * +llvmpipe_create_blend_state(struct pipe_context *pipe, + const struct pipe_blend_state *blend) +{ + return mem_dup(blend, sizeof(*blend)); +} + +void llvmpipe_bind_blend_state( struct pipe_context *pipe, + void *blend ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + llvmpipe->blend = blend; + + llvmpipe->dirty |= LP_NEW_BLEND; +} + +void llvmpipe_delete_blend_state(struct pipe_context *pipe, + void *blend) +{ + FREE( blend ); +} + + +void llvmpipe_set_blend_color( struct pipe_context *pipe, + const struct pipe_blend_color *blend_color ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + unsigned i, j; + + memcpy(&llvmpipe->blend_color, blend_color, sizeof *blend_color); + + if(!llvmpipe->jit_context.blend_color) + llvmpipe->jit_context.blend_color = align_malloc(4 * 16, 16); + for (i = 0; i < 4; ++i) { + uint8_t c = float_to_ubyte(blend_color->color[i]); + for (j = 0; j < 16; ++j) + llvmpipe->jit_context.blend_color[i*4 + j] = c; + } +} + + +/** XXX move someday? Or consolidate all these simple state setters + * into one file. + */ + + +void * +llvmpipe_create_depth_stencil_state(struct pipe_context *pipe, + const struct pipe_depth_stencil_alpha_state *depth_stencil) +{ + return mem_dup(depth_stencil, sizeof(*depth_stencil)); +} + +void +llvmpipe_bind_depth_stencil_state(struct pipe_context *pipe, + void *depth_stencil) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + llvmpipe->depth_stencil = (const struct pipe_depth_stencil_alpha_state *)depth_stencil; + + if(llvmpipe->depth_stencil) + llvmpipe->jit_context.alpha_ref_value = llvmpipe->depth_stencil->alpha.ref_value; + + llvmpipe->dirty |= LP_NEW_DEPTH_STENCIL_ALPHA; +} + +void +llvmpipe_delete_depth_stencil_state(struct pipe_context *pipe, void *depth) +{ + FREE( depth ); +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_clip.c b/src/gallium/drivers/llvmpipe/lp_state_clip.c new file mode 100644 index 00000000000..df68f27acc9 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_clip.c @@ -0,0 +1,79 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ +#include "lp_context.h" +#include "lp_state.h" +#include "draw/draw_context.h" + + +void llvmpipe_set_clip_state( struct pipe_context *pipe, + const struct pipe_clip_state *clip ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + /* pass the clip state to the draw module */ + draw_set_clip_state(llvmpipe->draw, clip); +} + + +void llvmpipe_set_viewport_state( struct pipe_context *pipe, + const struct pipe_viewport_state *viewport ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + /* pass the viewport info to the draw module */ + draw_set_viewport_state(llvmpipe->draw, viewport); + + llvmpipe->viewport = *viewport; /* struct copy */ + llvmpipe->dirty |= LP_NEW_VIEWPORT; +} + + +void llvmpipe_set_scissor_state( struct pipe_context *pipe, + const struct pipe_scissor_state *scissor ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + draw_flush(llvmpipe->draw); + + llvmpipe->scissor = *scissor; /* struct copy */ + llvmpipe->dirty |= LP_NEW_SCISSOR; +} + + +void llvmpipe_set_polygon_stipple( struct pipe_context *pipe, + const struct pipe_poly_stipple *stipple ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + draw_flush(llvmpipe->draw); + + llvmpipe->poly_stipple = *stipple; /* struct copy */ + llvmpipe->dirty |= LP_NEW_STIPPLE; +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c new file mode 100644 index 00000000000..6fbb057937e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c @@ -0,0 +1,258 @@ +/************************************************************************** + * + * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#include "util/u_math.h" +#include "util/u_memory.h" +#include "pipe/p_shader_tokens.h" +#include "draw/draw_context.h" +#include "draw/draw_vertex.h" +#include "draw/draw_private.h" +#include "lp_context.h" +#include "lp_screen.h" +#include "lp_tex_cache.h" +#include "lp_state.h" + + +/** + * Mark the current vertex layout as "invalid". + * We'll validate the vertex layout later, when we start to actually + * render a point or line or tri. + */ +static void +invalidate_vertex_layout(struct llvmpipe_context *llvmpipe) +{ + llvmpipe->vertex_info.num_attribs = 0; +} + + +/** + * The vertex info describes how to convert the post-transformed vertices + * (simple float[][4]) used by the 'draw' module into vertices for + * rasterization. + * + * This function validates the vertex layout and returns a pointer to a + * vertex_info object. + */ +struct vertex_info * +llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe) +{ + struct vertex_info *vinfo = &llvmpipe->vertex_info; + + if (vinfo->num_attribs == 0) { + /* compute vertex layout now */ + const struct lp_fragment_shader *lpfs = llvmpipe->fs; + const enum interp_mode colorInterp + = llvmpipe->rasterizer->flatshade ? INTERP_CONSTANT : INTERP_LINEAR; + uint i; + + if (llvmpipe->vbuf) { + /* if using the post-transform vertex buffer, tell draw_vbuf to + * simply emit the whole post-xform vertex as-is: + */ + struct vertex_info *vinfo_vbuf = &llvmpipe->vertex_info_vbuf; + const uint num = draw_num_vs_outputs(llvmpipe->draw); + uint i; + + /* No longer any need to try and emit draw vertex_header info. + */ + vinfo_vbuf->num_attribs = 0; + for (i = 0; i < num; i++) { + draw_emit_vertex_attr(vinfo_vbuf, EMIT_4F, INTERP_PERSPECTIVE, i); + } + draw_compute_vertex_size(vinfo_vbuf); + } + + /* + * Loop over fragment shader inputs, searching for the matching output + * from the vertex shader. + */ + vinfo->num_attribs = 0; + for (i = 0; i < lpfs->info.num_inputs; i++) { + int src; + switch (lpfs->info.input_semantic_name[i]) { + case TGSI_SEMANTIC_POSITION: + src = draw_find_vs_output(llvmpipe->draw, + TGSI_SEMANTIC_POSITION, 0); + draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_POS, src); + break; + + case TGSI_SEMANTIC_COLOR: + src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_COLOR, + lpfs->info.input_semantic_index[i]); + draw_emit_vertex_attr(vinfo, EMIT_4F, colorInterp, src); + break; + + case TGSI_SEMANTIC_FOG: + src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_FOG, 0); + draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src); + break; + + case TGSI_SEMANTIC_GENERIC: + case TGSI_SEMANTIC_FACE: + /* this includes texcoords and varying vars */ + src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_GENERIC, + lpfs->info.input_semantic_index[i]); + draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src); + break; + + default: + assert(0); + } + } + + llvmpipe->psize_slot = draw_find_vs_output(llvmpipe->draw, + TGSI_SEMANTIC_PSIZE, 0); + if (llvmpipe->psize_slot > 0) { + draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_CONSTANT, + llvmpipe->psize_slot); + } + + draw_compute_vertex_size(vinfo); + } + + return vinfo; +} + + +/** + * Called from vbuf module. + * + * Note that there's actually two different vertex layouts in llvmpipe. + * + * The normal one is computed in llvmpipe_get_vertex_info() above and is + * used by the point/line/tri "setup" code. + * + * The other one (this one) is only used by the vbuf module (which is + * not normally used by default but used in testing). For the vbuf module, + * we basically want to pass-through the draw module's vertex layout as-is. + * When the llvmpipe vbuf code begins drawing, the normal vertex layout + * will come into play again. + */ +struct vertex_info * +llvmpipe_get_vbuf_vertex_info(struct llvmpipe_context *llvmpipe) +{ + (void) llvmpipe_get_vertex_info(llvmpipe); + return &llvmpipe->vertex_info_vbuf; +} + + +/** + * Recompute cliprect from scissor bounds, scissor enable and surface size. + */ +static void +compute_cliprect(struct llvmpipe_context *lp) +{ + /* LP_NEW_FRAMEBUFFER + */ + uint surfWidth = lp->framebuffer.width; + uint surfHeight = lp->framebuffer.height; + + /* LP_NEW_RASTERIZER + */ + if (lp->rasterizer->scissor) { + + /* LP_NEW_SCISSOR + * + * clip to scissor rect: + */ + lp->cliprect.minx = MAX2(lp->scissor.minx, 0); + lp->cliprect.miny = MAX2(lp->scissor.miny, 0); + lp->cliprect.maxx = MIN2(lp->scissor.maxx, surfWidth); + lp->cliprect.maxy = MIN2(lp->scissor.maxy, surfHeight); + } + else { + /* clip to surface bounds */ + lp->cliprect.minx = 0; + lp->cliprect.miny = 0; + lp->cliprect.maxx = surfWidth; + lp->cliprect.maxy = surfHeight; + } +} + + +static void +update_tgsi_samplers( struct llvmpipe_context *llvmpipe ) +{ + unsigned i; + + /* vertex shader samplers */ + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + llvmpipe->tgsi.vert_samplers[i].sampler = llvmpipe->sampler[i]; + llvmpipe->tgsi.vert_samplers[i].texture = llvmpipe->texture[i]; + llvmpipe->tgsi.frag_samplers[i].base.get_samples = lp_get_samples; + } + + /* fragment shader samplers */ + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + llvmpipe->tgsi.frag_samplers[i].sampler = llvmpipe->sampler[i]; + llvmpipe->tgsi.frag_samplers[i].texture = llvmpipe->texture[i]; + llvmpipe->tgsi.frag_samplers[i].base.get_samples = lp_get_samples; + } + + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + lp_tex_tile_cache_validate_texture( llvmpipe->tex_cache[i] ); + } + + llvmpipe->jit_context.samplers = (struct tgsi_sampler **)llvmpipe->tgsi.frag_samplers_list; +} + +/* Hopefully this will remain quite simple, otherwise need to pull in + * something like the state tracker mechanism. + */ +void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe ) +{ + struct llvmpipe_screen *lp_screen = llvmpipe_screen(llvmpipe->pipe.screen); + + /* Check for updated textures. + */ + if (llvmpipe->tex_timestamp != lp_screen->timestamp) { + llvmpipe->tex_timestamp = lp_screen->timestamp; + llvmpipe->dirty |= LP_NEW_TEXTURE; + } + + if (llvmpipe->dirty & (LP_NEW_SAMPLER | + LP_NEW_TEXTURE)) + update_tgsi_samplers( llvmpipe ); + + if (llvmpipe->dirty & (LP_NEW_RASTERIZER | + LP_NEW_FS | + LP_NEW_VS)) + invalidate_vertex_layout( llvmpipe ); + + if (llvmpipe->dirty & (LP_NEW_SCISSOR | + LP_NEW_RASTERIZER | + LP_NEW_FRAMEBUFFER)) + compute_cliprect(llvmpipe); + + if (llvmpipe->dirty & (LP_NEW_FS | + LP_NEW_BLEND | + LP_NEW_DEPTH_STENCIL_ALPHA)) + llvmpipe_update_fs( llvmpipe ); + + + llvmpipe->dirty = 0; +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c new file mode 100644 index 00000000000..94170bd7161 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c @@ -0,0 +1,804 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Code generate the whole fragment pipeline. + * + * The fragment pipeline consists of the following stages: + * - stipple (TBI) + * - early depth test + * - fragment shader + * - alpha test + * - depth/stencil test (stencil TBI) + * - blending + * + * This file has only the glue to assembly the fragment pipeline. The actual + * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the + * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we + * muster the LLVM JIT execution engine to create a function that follows an + * established binary interface and that can be called from C directly. + * + * A big source of complexity here is that we often want to run different + * stages with different precisions and data types and precisions. For example, + * the fragment shader needs typically to be done in floats, but the + * depth/stencil test and blending is better done in the type that most closely + * matches the depth/stencil and color buffer respectively. + * + * Since the width of a SIMD vector register stays the same regardless of the + * element type, different types imply different number of elements, so we must + * code generate more instances of the stages with larger types to be able to + * feed/consume the stages with smaller types. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + +#include "pipe/p_defines.h" +#include "util/u_memory.h" +#include "util/u_format.h" +#include "util/u_debug_dump.h" +#include "pipe/internal/p_winsys_screen.h" +#include "pipe/p_shader_tokens.h" +#include "draw/draw_context.h" +#include "tgsi/tgsi_dump.h" +#include "tgsi/tgsi_scan.h" +#include "tgsi/tgsi_parse.h" +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_conv.h" +#include "lp_bld_intr.h" +#include "lp_bld_logic.h" +#include "lp_bld_depth.h" +#include "lp_bld_interp.h" +#include "lp_bld_tgsi.h" +#include "lp_bld_alpha.h" +#include "lp_bld_blend.h" +#include "lp_bld_swizzle.h" +#include "lp_bld_flow.h" +#include "lp_bld_debug.h" +#include "lp_screen.h" +#include "lp_context.h" +#include "lp_state.h" +#include "lp_quad.h" + + +static const unsigned char quad_offset_x[4] = {0, 1, 0, 1}; +static const unsigned char quad_offset_y[4] = {0, 0, 1, 1}; + + +/* + * Derive from the quad's upper left scalar coordinates the coordinates for + * all other quad pixels + */ +static void +generate_pos0(LLVMBuilderRef builder, + LLVMValueRef x, + LLVMValueRef y, + LLVMValueRef *x0, + LLVMValueRef *y0) +{ + LLVMTypeRef int_elem_type = LLVMInt32Type(); + LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE); + LLVMTypeRef elem_type = LLVMFloatType(); + LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE); + LLVMValueRef x_offsets[QUAD_SIZE]; + LLVMValueRef y_offsets[QUAD_SIZE]; + unsigned i; + + x = lp_build_broadcast(builder, int_vec_type, x); + y = lp_build_broadcast(builder, int_vec_type, y); + + for(i = 0; i < QUAD_SIZE; ++i) { + x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0); + y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0); + } + + x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), ""); + y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), ""); + + *x0 = LLVMBuildSIToFP(builder, x, vec_type, ""); + *y0 = LLVMBuildSIToFP(builder, y, vec_type, ""); +} + + +/** + * Generate the depth test. + */ +static void +generate_depth(struct llvmpipe_context *lp, + LLVMBuilderRef builder, + const struct pipe_depth_state *state, + union lp_type src_type, + struct lp_build_mask_context *mask, + LLVMValueRef src, + LLVMValueRef dst_ptr) +{ + const struct util_format_description *format_desc; + union lp_type dst_type; + + if(!lp->framebuffer.zsbuf) + return; + + format_desc = util_format_description(lp->framebuffer.zsbuf->format); + assert(format_desc); + + /* Pick the depth type. */ + dst_type = lp_depth_type(format_desc, src_type.width*src_type.length); + + /* FIXME: Cope with a depth test type with a different bit width. */ + assert(dst_type.width == src_type.width); + assert(dst_type.length == src_type.length); + +#if 1 + src = lp_build_clamped_float_to_unsigned_norm(builder, + src_type, + dst_type.width, + src); +#else + lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1); +#endif + + lp_build_depth_test(builder, + state, + dst_type, + format_desc, + mask, + src, + dst_ptr); +} + + +struct build_fetch_texel_context +{ + LLVMValueRef context_ptr; + + LLVMValueRef samplers_ptr; + + /** Coords/texels store */ + LLVMValueRef store_ptr; +}; + + +void PIPE_CDECL +lp_fetch_texel_soa( struct tgsi_sampler **samplers, + uint32_t unit, + float *store ) +{ + struct tgsi_sampler *sampler = samplers[unit]; + +#if 0 + uint j; + + debug_printf("%s sampler: %p (%p) store: %p\n", + __FUNCTION__, + sampler, *sampler, + store ); + + debug_printf("lodbias %f\n", store[12]); + + for (j = 0; j < 4; j++) + debug_printf("sample %d texcoord %f %f\n", + j, + store[0+j], + store[4+j]); +#endif + + { + float rgba[NUM_CHANNELS][QUAD_SIZE]; + sampler->get_samples(sampler, + &store[0], + &store[4], + &store[8], + 0.0f, /*store[12], lodbias */ + rgba); + memcpy(store, rgba, sizeof rgba); + } + +#if 0 + for (j = 0; j < 4; j++) + debug_printf("sample %d result %f %f %f %f\n", + j, + store[0+j], + store[4+j], + store[8+j], + store[12+j]); +#endif +} + + +static void +emit_fetch_texel( LLVMBuilderRef builder, + void *context, + unsigned unit, + unsigned num_coords, + const LLVMValueRef *coords, + LLVMValueRef lodbias, + LLVMValueRef *texel) +{ + struct build_fetch_texel_context *bld = context; + LLVMTypeRef vec_type = LLVMTypeOf(coords[0]); + LLVMValueRef args[3]; + unsigned i; + + if(!bld->samplers_ptr) + bld->samplers_ptr = lp_jit_context_samplers(builder, bld->context_ptr); + + if(!bld->store_ptr) + bld->store_ptr = LLVMBuildArrayAlloca(builder, + vec_type, + LLVMConstInt(LLVMInt32Type(), 4, 0), + "texel_store"); + + for (i = 0; i < num_coords; i++) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef coord_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, ""); + LLVMBuildStore(builder, coords[i], coord_ptr); + } + + args[0] = bld->samplers_ptr; + args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0); + args[2] = bld->store_ptr; + + lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3); + + for (i = 0; i < NUM_CHANNELS; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef texel_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, ""); + texel[i] = LLVMBuildLoad(builder, texel_ptr, ""); + } +} + + +/** + * Generate the fragment shader, depth/stencil test, and alpha tests. + */ +static void +generate_fs(struct llvmpipe_context *lp, + struct lp_fragment_shader *shader, + const struct lp_fragment_shader_variant_key *key, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef context_ptr, + unsigned i, + const struct lp_build_interp_soa_context *interp, + struct build_fetch_texel_context *sampler, + LLVMValueRef *pmask, + LLVMValueRef *color, + LLVMValueRef depth_ptr) +{ + const struct tgsi_token *tokens = shader->base.tokens; + LLVMTypeRef elem_type; + LLVMTypeRef vec_type; + LLVMTypeRef int_vec_type; + LLVMValueRef consts_ptr; + LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS]; + LLVMValueRef z = interp->pos[2]; + struct lp_build_mask_context mask; + boolean early_depth_test; + unsigned attrib; + unsigned chan; + + elem_type = lp_build_elem_type(type); + vec_type = lp_build_vec_type(type); + int_vec_type = lp_build_int_vec_type(type); + + consts_ptr = lp_jit_context_constants(builder, context_ptr); + + lp_build_mask_begin(&mask, builder, type, *pmask); + + early_depth_test = + lp->depth_stencil->depth.enabled && + lp->framebuffer.zsbuf && + !lp->depth_stencil->alpha.enabled && + !lp->fs->info.uses_kill && + !lp->fs->info.writes_z; + + if(early_depth_test) + generate_depth(lp, builder, &key->depth, + type, &mask, + z, depth_ptr); + + memset(outputs, 0, sizeof outputs); + + lp_build_tgsi_soa(builder, tokens, type, &mask, + consts_ptr, interp->pos, interp->inputs, + outputs, emit_fetch_texel, sampler); + + for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) { + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + if(outputs[attrib][chan]) { + lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]); + + switch (shader->info.output_semantic_name[attrib]) { + case TGSI_SEMANTIC_COLOR: + { + unsigned cbuf = shader->info.output_semantic_index[attrib]; + + lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]); + + /* Alpha test */ + /* XXX: should the alpha reference value be passed separately? */ + if(cbuf == 0 && chan == 3) { + LLVMValueRef alpha = outputs[attrib][chan]; + LLVMValueRef alpha_ref_value; + alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr); + alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value); + lp_build_alpha_test(builder, &key->alpha, type, + &mask, alpha, alpha_ref_value); + } + + if(cbuf == 0) + color[chan] = outputs[attrib][chan]; + + break; + } + + case TGSI_SEMANTIC_POSITION: + if(chan == 2) + z = outputs[attrib][chan]; + break; + } + } + } + } + + if(!early_depth_test) + generate_depth(lp, builder, &key->depth, + type, &mask, + z, depth_ptr); + + lp_build_mask_end(&mask); + + *pmask = mask.value; + +} + + +/** + * Generate color blending and color output. + */ +static void +generate_blend(const struct pipe_blend_state *blend, + LLVMBuilderRef builder, + union lp_type type, + LLVMValueRef context_ptr, + LLVMValueRef mask, + LLVMValueRef *src, + LLVMValueRef dst_ptr) +{ + struct lp_build_context bld; + LLVMTypeRef vec_type; + LLVMTypeRef int_vec_type; + LLVMValueRef const_ptr; + LLVMValueRef con[4]; + LLVMValueRef dst[4]; + LLVMValueRef res[4]; + unsigned chan; + + vec_type = lp_build_vec_type(type); + int_vec_type = lp_build_int_vec_type(type); + + lp_build_context_init(&bld, builder, type); + + const_ptr = lp_jit_context_blend_color(builder, context_ptr); + const_ptr = LLVMBuildBitCast(builder, const_ptr, + LLVMPointerType(vec_type, 0), ""); + + for(chan = 0; chan < 4; ++chan) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); + con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), ""); + + dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), ""); + + lp_build_name(con[chan], "con.%c", "rgba"[chan]); + lp_build_name(dst[chan], "dst.%c", "rgba"[chan]); + } + + lp_build_blend_soa(builder, blend, type, src, dst, con, res); + + for(chan = 0; chan < 4; ++chan) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); + lp_build_name(res[chan], "res.%c", "rgba"[chan]); + res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); + LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); + } +} + + +/** + * Generate the runtime callable function for the whole fragment pipeline. + */ +static struct lp_fragment_shader_variant * +generate_fragment(struct llvmpipe_context *lp, + struct lp_fragment_shader *shader, + const struct lp_fragment_shader_variant_key *key) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen); + struct lp_fragment_shader_variant *variant; + union lp_type fs_type; + union lp_type blend_type; + LLVMTypeRef fs_elem_type; + LLVMTypeRef fs_vec_type; + LLVMTypeRef fs_int_vec_type; + LLVMTypeRef blend_vec_type; + LLVMTypeRef blend_int_vec_type; + LLVMTypeRef arg_types[9]; + LLVMTypeRef func_type; + LLVMValueRef context_ptr; + LLVMValueRef x; + LLVMValueRef y; + LLVMValueRef a0_ptr; + LLVMValueRef dadx_ptr; + LLVMValueRef dady_ptr; + LLVMValueRef mask_ptr; + LLVMValueRef color_ptr; + LLVMValueRef depth_ptr; + LLVMBasicBlockRef block; + LLVMBuilderRef builder; + LLVMValueRef x0; + LLVMValueRef y0; + struct build_fetch_texel_context sampler; + struct lp_build_interp_soa_context interp; + LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH]; + LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH]; + LLVMValueRef blend_mask; + LLVMValueRef blend_in_color[NUM_CHANNELS]; + unsigned num_fs; + unsigned i; + unsigned chan; + +#ifdef DEBUG + tgsi_dump(shader->base.tokens, 0); + if(key->depth.enabled) { + debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE)); + debug_printf("depth.writemask = %u\n", key->depth.writemask); + debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count); + } + if(key->alpha.enabled) { + debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE)); + debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value); + } + if(key->blend.logicop_enable) { + debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func); + } + else if(key->blend.blend_enable) { + debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE)); + debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE)); + debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE)); + debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE)); + debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE)); + debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE)); + } + debug_printf("blend.colormask = 0x%x\n", key->blend.colormask); +#endif + + variant = CALLOC_STRUCT(lp_fragment_shader_variant); + if(!variant) + return NULL; + + variant->shader = shader; + memcpy(&variant->key, key, sizeof *key); + + /* TODO: actually pick these based on the fs and color buffer + * characteristics. */ + + fs_type.value = 0; + fs_type.floating = TRUE; /* floating point values */ + fs_type.sign = TRUE; /* values are signed */ + fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ + fs_type.width = 32; /* 32-bit float */ + fs_type.length = 4; /* 4 element per vector */ + num_fs = 4; + + blend_type.value = 0; + blend_type.floating = FALSE; /* values are integers */ + blend_type.sign = FALSE; /* values are unsigned */ + blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */ + blend_type.width = 8; /* 8-bit ubyte values */ + blend_type.length = 16; /* 16 elements per vector */ + + /* + * Generate the function prototype. Any change here must be reflected in + * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa. + */ + + fs_elem_type = lp_build_elem_type(fs_type); + fs_vec_type = lp_build_vec_type(fs_type); + fs_int_vec_type = lp_build_int_vec_type(fs_type); + + blend_vec_type = lp_build_vec_type(blend_type); + blend_int_vec_type = lp_build_int_vec_type(blend_type); + + arg_types[0] = screen->context_ptr_type; /* context */ + arg_types[1] = LLVMInt32Type(); /* x */ + arg_types[2] = LLVMInt32Type(); /* y */ + arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */ + arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */ + arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */ + arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */ + arg_types[7] = LLVMPointerType(blend_vec_type, 0); /* color */ + arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */ + + func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0); + + variant->function = LLVMAddFunction(screen->module, "shader", func_type); + LLVMSetFunctionCallConv(variant->function, LLVMCCallConv); + for(i = 0; i < Elements(arg_types); ++i) + if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) + LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute); + + context_ptr = LLVMGetParam(variant->function, 0); + x = LLVMGetParam(variant->function, 1); + y = LLVMGetParam(variant->function, 2); + a0_ptr = LLVMGetParam(variant->function, 3); + dadx_ptr = LLVMGetParam(variant->function, 4); + dady_ptr = LLVMGetParam(variant->function, 5); + mask_ptr = LLVMGetParam(variant->function, 6); + color_ptr = LLVMGetParam(variant->function, 7); + depth_ptr = LLVMGetParam(variant->function, 8); + + lp_build_name(context_ptr, "context"); + lp_build_name(x, "x"); + lp_build_name(y, "y"); + lp_build_name(a0_ptr, "a0"); + lp_build_name(dadx_ptr, "dadx"); + lp_build_name(dady_ptr, "dady"); + lp_build_name(mask_ptr, "mask"); + lp_build_name(color_ptr, "color"); + lp_build_name(depth_ptr, "depth"); + + /* + * Function body + */ + + block = LLVMAppendBasicBlock(variant->function, "entry"); + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, block); + + generate_pos0(builder, x, y, &x0, &y0); + + lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type, + a0_ptr, dadx_ptr, dady_ptr, + x0, y0, 2, 0); + + memset(&sampler, 0, sizeof sampler); + sampler.context_ptr = context_ptr; + + for(i = 0; i < num_fs; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef out_color[NUM_CHANNELS]; + LLVMValueRef depth_ptr_i; + + if(i != 0) + lp_build_interp_soa_update(&interp); + + fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), ""); + depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, ""); + + generate_fs(lp, shader, key, + builder, + fs_type, + context_ptr, + i, + &interp, + &sampler, + &fs_mask[i], + out_color, + depth_ptr_i); + + for(chan = 0; chan < NUM_CHANNELS; ++chan) + fs_out_color[chan][i] = out_color[chan]; + } + + /* + * Convert the fs's output color and mask to fit to the blending type. + */ + + for(chan = 0; chan < NUM_CHANNELS; ++chan) { + lp_build_conv(builder, fs_type, blend_type, + fs_out_color[chan], num_fs, + &blend_in_color[chan], 1); + lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]); + + } + + lp_build_conv_mask(builder, fs_type, blend_type, + fs_mask, num_fs, + &blend_mask, 1); + + /* + * Blending. + */ + + generate_blend(&key->blend, + builder, + blend_type, + context_ptr, + blend_mask, + blend_in_color, + color_ptr); + + LLVMBuildRetVoid(builder); + + LLVMDisposeBuilder(builder); + + /* + * Translate the LLVM IR into machine code. + */ + + LLVMRunFunctionPassManager(screen->pass, variant->function); + +#ifdef DEBUG + LLVMDumpValue(variant->function); + debug_printf("\n"); +#endif + + if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) { + LLVMDumpValue(variant->function); + abort(); + } + + variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function); + +#ifdef DEBUG + lp_disassemble(variant->jit_function); +#endif + + variant->next = shader->variants; + shader->variants = variant; + + return variant; +} + + +void * +llvmpipe_create_fs_state(struct pipe_context *pipe, + const struct pipe_shader_state *templ) +{ + struct lp_fragment_shader *shader; + + shader = CALLOC_STRUCT(lp_fragment_shader); + if (!shader) + return NULL; + + /* get/save the summary info for this shader */ + tgsi_scan_shader(templ->tokens, &shader->info); + + /* we need to keep a local copy of the tokens */ + shader->base.tokens = tgsi_dup_tokens(templ->tokens); + + return shader; +} + + +void +llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + llvmpipe->fs = (struct lp_fragment_shader *) fs; + + llvmpipe->dirty |= LP_NEW_FS; +} + + +void +llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen); + struct lp_fragment_shader *shader = fs; + struct lp_fragment_shader_variant *variant; + + assert(fs != llvmpipe->fs); + + variant = shader->variants; + while(variant) { + struct lp_fragment_shader_variant *next = variant->next; + + if(variant->function) { + if(variant->jit_function) + LLVMFreeMachineCodeForFunction(screen->engine, variant->function); + LLVMDeleteFunction(variant->function); + } + + FREE(variant); + + variant = next; + } + + FREE((void *) shader->base.tokens); + FREE(shader); +} + + + +void +llvmpipe_set_constant_buffer(struct pipe_context *pipe, + uint shader, uint index, + const struct pipe_constant_buffer *buf) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + assert(shader < PIPE_SHADER_TYPES); + assert(index == 0); + + /* note: reference counting */ + pipe_buffer_reference(&llvmpipe->constants[shader].buffer, + buf ? buf->buffer : NULL); + + llvmpipe->dirty |= LP_NEW_CONSTANTS; +} + + +/** + * We need to generate several variants of the fragment pipeline to match + * all the combinations of the contributing state atoms. + * + * TODO: there is actually no reason to tie this to context state -- the + * generated code could be cached globally in the screen. + */ +static void +make_variant_key(struct llvmpipe_context *lp, + struct lp_fragment_shader_variant_key *key) +{ + memset(key, 0, sizeof *key); + + memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); + + key->alpha.enabled = lp->depth_stencil->alpha.enabled; + if(key->alpha.enabled) + key->alpha.func = lp->depth_stencil->alpha.func; + /* alpha.ref_value is passed in jit_context */ + + memcpy(&key->blend, lp->blend, sizeof key->blend); +} + + +void +llvmpipe_update_fs(struct llvmpipe_context *lp) +{ + struct lp_fragment_shader *shader = lp->fs; + struct lp_fragment_shader_variant_key key; + struct lp_fragment_shader_variant *variant; + + make_variant_key(lp, &key); + + variant = shader->variants; + while(variant) { + if(memcmp(&variant->key, &key, sizeof key) == 0) + break; + + variant = variant->next; + } + + if(!variant) + variant = generate_fragment(lp, shader, &key); + + shader->current = variant; +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c b/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c new file mode 100644 index 00000000000..4561c6b8456 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c @@ -0,0 +1,62 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#include "pipe/p_defines.h" +#include "util/u_memory.h" +#include "lp_context.h" +#include "lp_state.h" +#include "draw/draw_context.h" + + + +void * +llvmpipe_create_rasterizer_state(struct pipe_context *pipe, + const struct pipe_rasterizer_state *rast) +{ + return mem_dup(rast, sizeof(*rast)); +} + +void llvmpipe_bind_rasterizer_state(struct pipe_context *pipe, + void *setup) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + /* pass-through to draw module */ + draw_set_rasterizer_state(llvmpipe->draw, setup); + + llvmpipe->rasterizer = (struct pipe_rasterizer_state *)setup; + + llvmpipe->dirty |= LP_NEW_RASTERIZER; +} + +void llvmpipe_delete_rasterizer_state(struct pipe_context *pipe, + void *rasterizer) +{ + FREE( rasterizer ); +} + + diff --git a/src/gallium/drivers/llvmpipe/lp_state_sampler.c b/src/gallium/drivers/llvmpipe/lp_state_sampler.c new file mode 100644 index 00000000000..4fef541b1e3 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_sampler.c @@ -0,0 +1,117 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: + * Brian Paul + */ + +#include "util/u_memory.h" + +#include "draw/draw_context.h" + +#include "lp_context.h" +#include "lp_context.h" +#include "lp_state.h" +#include "lp_texture.h" +#include "lp_tex_cache.h" +#include "draw/draw_context.h" + + + +void * +llvmpipe_create_sampler_state(struct pipe_context *pipe, + const struct pipe_sampler_state *sampler) +{ + return mem_dup(sampler, sizeof(*sampler)); +} + + +void +llvmpipe_bind_sampler_states(struct pipe_context *pipe, + unsigned num, void **sampler) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + unsigned i; + + assert(num <= PIPE_MAX_SAMPLERS); + + /* Check for no-op */ + if (num == llvmpipe->num_samplers && + !memcmp(llvmpipe->sampler, sampler, num * sizeof(void *))) + return; + + draw_flush(llvmpipe->draw); + + for (i = 0; i < num; ++i) + llvmpipe->sampler[i] = sampler[i]; + for (i = num; i < PIPE_MAX_SAMPLERS; ++i) + llvmpipe->sampler[i] = NULL; + + llvmpipe->num_samplers = num; + + llvmpipe->dirty |= LP_NEW_SAMPLER; +} + + +void +llvmpipe_set_sampler_textures(struct pipe_context *pipe, + unsigned num, struct pipe_texture **texture) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + uint i; + + assert(num <= PIPE_MAX_SAMPLERS); + + /* Check for no-op */ + if (num == llvmpipe->num_textures && + !memcmp(llvmpipe->texture, texture, num * sizeof(struct pipe_texture *))) + return; + + draw_flush(llvmpipe->draw); + + for (i = 0; i < PIPE_MAX_SAMPLERS; i++) { + struct pipe_texture *tex = i < num ? texture[i] : NULL; + + pipe_texture_reference(&llvmpipe->texture[i], tex); + lp_tex_tile_cache_set_texture(llvmpipe->tex_cache[i], tex); + } + + llvmpipe->num_textures = num; + + llvmpipe->dirty |= LP_NEW_TEXTURE; +} + + +void +llvmpipe_delete_sampler_state(struct pipe_context *pipe, + void *sampler) +{ + FREE( sampler ); +} + + + diff --git a/src/gallium/drivers/llvmpipe/lp_state_surface.c b/src/gallium/drivers/llvmpipe/lp_state_surface.c new file mode 100644 index 00000000000..177a26b7b1f --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_surface.c @@ -0,0 +1,106 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + +#include "lp_context.h" +#include "lp_state.h" +#include "lp_surface.h" +#include "lp_tile_cache.h" + +#include "draw/draw_context.h" + + +/** + * XXX this might get moved someday + * Set the framebuffer surface info: color buffers, zbuffer, stencil buffer. + * Here, we flush the old surfaces and update the tile cache to point to the new + * surfaces. + */ +void +llvmpipe_set_framebuffer_state(struct pipe_context *pipe, + const struct pipe_framebuffer_state *fb) +{ + struct llvmpipe_context *lp = llvmpipe_context(pipe); + uint i; + + for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) { + /* check if changing cbuf */ + if (lp->framebuffer.cbufs[i] != fb->cbufs[i]) { + /* flush old */ + lp_flush_tile_cache(lp->cbuf_cache[i]); + + /* assign new */ + lp->framebuffer.cbufs[i] = fb->cbufs[i]; + + /* update cache */ + lp_tile_cache_set_surface(lp->cbuf_cache[i], fb->cbufs[i]); + } + } + + lp->framebuffer.nr_cbufs = fb->nr_cbufs; + + /* zbuf changing? */ + if (lp->framebuffer.zsbuf != fb->zsbuf) { + + if(lp->zsbuf_transfer) { + struct pipe_screen *screen = pipe->screen; + + if(lp->zsbuf_map) { + screen->transfer_unmap(screen, lp->zsbuf_transfer); + lp->zsbuf_map = NULL; + } + + screen->tex_transfer_destroy(lp->zsbuf_transfer); + lp->zsbuf_transfer = NULL; + } + + /* assign new */ + lp->framebuffer.zsbuf = fb->zsbuf; + + /* Tell draw module how deep the Z/depth buffer is */ + if (lp->framebuffer.zsbuf) { + int depth_bits; + double mrd; + depth_bits = pf_get_component_bits(lp->framebuffer.zsbuf->format, + PIPE_FORMAT_COMP_Z); + if (depth_bits > 16) { + mrd = 0.0000001; + } + else { + mrd = 0.00002; + } + draw_set_mrd(lp->draw, mrd); + } + } + + lp->framebuffer.width = fb->width; + lp->framebuffer.height = fb->height; + + lp->dirty |= LP_NEW_FRAMEBUFFER; +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_vertex.c b/src/gallium/drivers/llvmpipe/lp_state_vertex.c new file mode 100644 index 00000000000..1a17631a4c6 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_vertex.c @@ -0,0 +1,73 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + + +#include "lp_context.h" +#include "lp_state.h" +#include "lp_surface.h" + +#include "draw/draw_context.h" + + +void +llvmpipe_set_vertex_elements(struct pipe_context *pipe, + unsigned count, + const struct pipe_vertex_element *attribs) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + assert(count <= PIPE_MAX_ATTRIBS); + + memcpy(llvmpipe->vertex_element, attribs, + count * sizeof(struct pipe_vertex_element)); + llvmpipe->num_vertex_elements = count; + + llvmpipe->dirty |= LP_NEW_VERTEX; + + draw_set_vertex_elements(llvmpipe->draw, count, attribs); +} + + +void +llvmpipe_set_vertex_buffers(struct pipe_context *pipe, + unsigned count, + const struct pipe_vertex_buffer *buffers) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + assert(count <= PIPE_MAX_ATTRIBS); + + memcpy(llvmpipe->vertex_buffer, buffers, count * sizeof(buffers[0])); + llvmpipe->num_vertex_buffers = count; + + llvmpipe->dirty |= LP_NEW_VERTEX; + + draw_set_vertex_buffers(llvmpipe->draw, count, buffers); +} diff --git a/src/gallium/drivers/llvmpipe/lp_state_vs.c b/src/gallium/drivers/llvmpipe/lp_state_vs.c new file mode 100644 index 00000000000..15c30296144 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_state_vs.c @@ -0,0 +1,96 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include "pipe/p_defines.h" +#include "tgsi/tgsi_parse.h" +#include "util/u_memory.h" +#include "draw/draw_context.h" + +#include "lp_context.h" +#include "lp_state.h" + + +void * +llvmpipe_create_vs_state(struct pipe_context *pipe, + const struct pipe_shader_state *templ) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + struct lp_vertex_shader *state; + + state = CALLOC_STRUCT(lp_vertex_shader); + if (state == NULL ) + goto fail; + + /* copy shader tokens, the ones passed in will go away. + */ + state->shader.tokens = tgsi_dup_tokens(templ->tokens); + if (state->shader.tokens == NULL) + goto fail; + + state->draw_data = draw_create_vertex_shader(llvmpipe->draw, templ); + if (state->draw_data == NULL) + goto fail; + + return state; + +fail: + if (state) { + FREE( (void *)state->shader.tokens ); + FREE( state->draw_data ); + FREE( state ); + } + return NULL; +} + + +void +llvmpipe_bind_vs_state(struct pipe_context *pipe, void *vs) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + llvmpipe->vs = (const struct lp_vertex_shader *)vs; + + draw_bind_vertex_shader(llvmpipe->draw, + (llvmpipe->vs ? llvmpipe->vs->draw_data : NULL)); + + llvmpipe->dirty |= LP_NEW_VS; +} + + +void +llvmpipe_delete_vs_state(struct pipe_context *pipe, void *vs) +{ + struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); + + struct lp_vertex_shader *state = + (struct lp_vertex_shader *)vs; + + draw_delete_vertex_shader(llvmpipe->draw, state->draw_data); + FREE( state ); +} diff --git a/src/gallium/drivers/llvmpipe/lp_surface.c b/src/gallium/drivers/llvmpipe/lp_surface.c new file mode 100644 index 00000000000..6110b0a193e --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_surface.c @@ -0,0 +1,50 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#include "util/u_rect.h" +#include "lp_context.h" +#include "lp_surface.h" + + +static void +lp_surface_copy(struct pipe_context *pipe, + struct pipe_surface *dest, unsigned destx, unsigned desty, + struct pipe_surface *src, unsigned srcx, unsigned srcy, + unsigned width, unsigned height) +{ + util_surface_copy(pipe, FALSE, + dest, destx, desty, + src, srcx, srcy, + width, height); +} + +void +lp_init_surface_functions(struct llvmpipe_context *lp) +{ + lp->pipe.surface_copy = lp_surface_copy; + lp->pipe.surface_fill = util_surface_fill; +} diff --git a/src/gallium/drivers/llvmpipe/lp_surface.h b/src/gallium/drivers/llvmpipe/lp_surface.h new file mode 100644 index 00000000000..4d78a53c4f5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_surface.h @@ -0,0 +1,42 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/* Authors: Keith Whitwell <keith@tungstengraphics.com> + */ + +#ifndef LP_SURFACE_H +#define LP_SURFACE_H + + +struct llvmpipe_context; + + +extern void +lp_init_surface_functions(struct llvmpipe_context *lp); + + +#endif /* LP_SURFACE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_test.h b/src/gallium/drivers/llvmpipe/lp_test.h new file mode 100644 index 00000000000..69aaae26e0a --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test.h @@ -0,0 +1,128 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * Shared testing code. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#ifndef LP_TEST_H +#define LP_TEST_H + + +#include <stdlib.h> +#include <stdio.h> +#include <float.h> + +#include <llvm-c/Core.h> +#include <llvm-c/Analysis.h> +#include <llvm-c/ExecutionEngine.h> +#include <llvm-c/Target.h> +#include <llvm-c/BitWriter.h> +#include <llvm-c/Transforms/Scalar.h> + +#include "pipe/p_state.h" +#include "util/u_format.h" +#include "util/u_math.h" +#include "util/u_debug_dump.h" + +#include "lp_bld_type.h" + + +void +write_tsv_header(FILE *fp); + + +boolean +test_some(unsigned verbose, FILE *fp, unsigned long n); + + +boolean +test_all(unsigned verbose, FILE *fp); + + +static INLINE uint64_t +rdtsc(void) +{ +#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) + uint32_t hi, lo; + __asm__ __volatile__ ("rdtsc" : "=a"(lo), "=d"(hi)); + return ((uint64_t)lo) | (((uint64_t)hi) << 32); +#else + return 0; +#endif +} + + +float +random_float(void); + + +void +dump_type(FILE *fp, union lp_type type); + + +double +read_elem(union lp_type type, const void *src, unsigned index); + + +void +write_elem(union lp_type type, void *dst, unsigned index, double src); + + +void +random_elem(union lp_type type, void *dst, unsigned index); + + +void +read_vec(union lp_type type, const void *src, double *dst); + + +void +write_vec(union lp_type type, void *dst, const double *src); + + +void +random_vec(union lp_type type, void *dst); + + +boolean +compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps); + + +boolean +compare_vec(union lp_type type, const void *res, const void *ref); + + +void +dump_vec(FILE *fp, union lp_type type, const void *src); + + +#endif /* !LP_TEST_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c new file mode 100644 index 00000000000..8dfad468e3c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c @@ -0,0 +1,881 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Unit tests for blend LLVM IR generation + * + * @author Jose Fonseca <jfonseca@vmware.com> + * + * Blend computation code derived from code written by + * @author Brian Paul <brian@vmware.com> + */ + + +#include "lp_bld_type.h" +#include "lp_bld_arit.h" +#include "lp_bld_blend.h" +#include "lp_bld_debug.h" +#include "lp_test.h" + + +enum vector_mode +{ + AoS = 0, + SoA = 1 +}; + + +typedef void (*blend_test_ptr_t)(const void *src, const void *dst, const void *con, void *res); + + +void +write_tsv_header(FILE *fp) +{ + fprintf(fp, + "result\t" + "cycles_per_channel\t" + "mode\t" + "type\t" + "sep_func\t" + "sep_src_factor\t" + "sep_dst_factor\t" + "rgb_func\t" + "rgb_src_factor\t" + "rgb_dst_factor\t" + "alpha_func\t" + "alpha_src_factor\t" + "alpha_dst_factor\n"); + + fflush(fp); +} + + +static void +write_tsv_row(FILE *fp, + const struct pipe_blend_state *blend, + enum vector_mode mode, + union lp_type type, + double cycles, + boolean success) +{ + fprintf(fp, "%s\t", success ? "pass" : "fail"); + + if (mode == AoS) { + fprintf(fp, "%.1f\t", cycles / type.length); + fprintf(fp, "aos\t"); + } + + if (mode == SoA) { + fprintf(fp, "%.1f\t", cycles / (4 * type.length)); + fprintf(fp, "soa\t"); + } + + fprintf(fp, "%s%u%sx%u\t", + type.floating ? "f" : (type.fixed ? "h" : (type.sign ? "s" : "u")), + type.width, + type.norm ? "n" : "", + type.length); + + fprintf(fp, + "%s\t%s\t%s\t", + blend->rgb_func != blend->alpha_func ? "true" : "false", + blend->rgb_src_factor != blend->alpha_src_factor ? "true" : "false", + blend->rgb_dst_factor != blend->alpha_dst_factor ? "true" : "false"); + + fprintf(fp, + "%s\t%s\t%s\t%s\t%s\t%s\n", + debug_dump_blend_func(blend->rgb_func, TRUE), + debug_dump_blend_factor(blend->rgb_src_factor, TRUE), + debug_dump_blend_factor(blend->rgb_dst_factor, TRUE), + debug_dump_blend_func(blend->alpha_func, TRUE), + debug_dump_blend_factor(blend->alpha_src_factor, TRUE), + debug_dump_blend_factor(blend->alpha_dst_factor, TRUE)); + + fflush(fp); +} + + +static void +dump_blend_type(FILE *fp, + const struct pipe_blend_state *blend, + enum vector_mode mode, + union lp_type type) +{ + fprintf(fp, "%s", mode ? "soa" : "aos"); + + fprintf(fp, " type=%s%u%sx%u", + type.floating ? "f" : (type.fixed ? "h" : (type.sign ? "s" : "u")), + type.width, + type.norm ? "n" : "", + type.length); + + fprintf(fp, + " %s=%s %s=%s %s=%s %s=%s %s=%s %s=%s", + "rgb_func", debug_dump_blend_func(blend->rgb_func, TRUE), + "rgb_src_factor", debug_dump_blend_factor(blend->rgb_src_factor, TRUE), + "rgb_dst_factor", debug_dump_blend_factor(blend->rgb_dst_factor, TRUE), + "alpha_func", debug_dump_blend_func(blend->alpha_func, TRUE), + "alpha_src_factor", debug_dump_blend_factor(blend->alpha_src_factor, TRUE), + "alpha_dst_factor", debug_dump_blend_factor(blend->alpha_dst_factor, TRUE)); + + fprintf(fp, " ...\n"); + fflush(fp); +} + + +static LLVMValueRef +add_blend_test(LLVMModuleRef module, + const struct pipe_blend_state *blend, + enum vector_mode mode, + union lp_type type) +{ + LLVMTypeRef ret_type; + LLVMTypeRef vec_type; + LLVMTypeRef args[4]; + LLVMValueRef func; + LLVMValueRef src_ptr; + LLVMValueRef dst_ptr; + LLVMValueRef const_ptr; + LLVMValueRef res_ptr; + LLVMBasicBlockRef block; + LLVMBuilderRef builder; + + ret_type = LLVMInt64Type(); + vec_type = lp_build_vec_type(type); + + args[3] = args[2] = args[1] = args[0] = LLVMPointerType(vec_type, 0); + func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidType(), args, 4, 0)); + LLVMSetFunctionCallConv(func, LLVMCCallConv); + src_ptr = LLVMGetParam(func, 0); + dst_ptr = LLVMGetParam(func, 1); + const_ptr = LLVMGetParam(func, 2); + res_ptr = LLVMGetParam(func, 3); + + block = LLVMAppendBasicBlock(func, "entry"); + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, block); + + if (mode == AoS) { + LLVMValueRef src; + LLVMValueRef dst; + LLVMValueRef con; + LLVMValueRef res; + + src = LLVMBuildLoad(builder, src_ptr, "src"); + dst = LLVMBuildLoad(builder, dst_ptr, "dst"); + con = LLVMBuildLoad(builder, const_ptr, "const"); + + res = lp_build_blend_aos(builder, blend, type, src, dst, con, 3); + + lp_build_name(res, "res"); + + LLVMBuildStore(builder, res, res_ptr); + } + + if (mode == SoA) { + LLVMValueRef src[4]; + LLVMValueRef dst[4]; + LLVMValueRef con[4]; + LLVMValueRef res[4]; + unsigned i; + + for(i = 0; i < 4; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + src[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, src_ptr, &index, 1, ""), ""); + dst[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), ""); + con[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), ""); + lp_build_name(src[i], "src.%c", "rgba"[i]); + lp_build_name(con[i], "con.%c", "rgba"[i]); + lp_build_name(dst[i], "dst.%c", "rgba"[i]); + } + + lp_build_blend_soa(builder, blend, type, src, dst, con, res); + + for(i = 0; i < 4; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + lp_build_name(res[i], "res.%c", "rgba"[i]); + LLVMBuildStore(builder, res[i], LLVMBuildGEP(builder, res_ptr, &index, 1, "")); + } + } + + LLVMBuildRetVoid(builder);; + + LLVMDisposeBuilder(builder); + return func; +} + + +/** Add and limit result to ceiling of 1.0 */ +#define ADD_SAT(R, A, B) \ +do { \ + R = (A) + (B); if (R > 1.0f) R = 1.0f; \ +} while (0) + +/** Subtract and limit result to floor of 0.0 */ +#define SUB_SAT(R, A, B) \ +do { \ + R = (A) - (B); if (R < 0.0f) R = 0.0f; \ +} while (0) + + +static void +compute_blend_ref_term(unsigned rgb_factor, + unsigned alpha_factor, + const double *factor, + const double *src, + const double *dst, + const double *con, + double *term) +{ + double temp; + + switch (rgb_factor) { + case PIPE_BLENDFACTOR_ONE: + term[0] = factor[0]; /* R */ + term[1] = factor[1]; /* G */ + term[2] = factor[2]; /* B */ + break; + case PIPE_BLENDFACTOR_SRC_COLOR: + term[0] = factor[0] * src[0]; /* R */ + term[1] = factor[1] * src[1]; /* G */ + term[2] = factor[2] * src[2]; /* B */ + break; + case PIPE_BLENDFACTOR_SRC_ALPHA: + term[0] = factor[0] * src[3]; /* R */ + term[1] = factor[1] * src[3]; /* G */ + term[2] = factor[2] * src[3]; /* B */ + break; + case PIPE_BLENDFACTOR_DST_COLOR: + term[0] = factor[0] * dst[0]; /* R */ + term[1] = factor[1] * dst[1]; /* G */ + term[2] = factor[2] * dst[2]; /* B */ + break; + case PIPE_BLENDFACTOR_DST_ALPHA: + term[0] = factor[0] * dst[3]; /* R */ + term[1] = factor[1] * dst[3]; /* G */ + term[2] = factor[2] * dst[3]; /* B */ + break; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + temp = MIN2(src[3], 1.0f - dst[3]); + term[0] = factor[0] * temp; /* R */ + term[1] = factor[1] * temp; /* G */ + term[2] = factor[2] * temp; /* B */ + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + term[0] = factor[0] * con[0]; /* R */ + term[1] = factor[1] * con[1]; /* G */ + term[2] = factor[2] * con[2]; /* B */ + break; + case PIPE_BLENDFACTOR_CONST_ALPHA: + term[0] = factor[0] * con[3]; /* R */ + term[1] = factor[1] * con[3]; /* G */ + term[2] = factor[2] * con[3]; /* B */ + break; + case PIPE_BLENDFACTOR_SRC1_COLOR: + assert(0); /* to do */ + break; + case PIPE_BLENDFACTOR_SRC1_ALPHA: + assert(0); /* to do */ + break; + case PIPE_BLENDFACTOR_ZERO: + term[0] = 0.0f; /* R */ + term[1] = 0.0f; /* G */ + term[2] = 0.0f; /* B */ + break; + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + term[0] = factor[0] * (1.0f - src[0]); /* R */ + term[1] = factor[1] * (1.0f - src[1]); /* G */ + term[2] = factor[2] * (1.0f - src[2]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + term[0] = factor[0] * (1.0f - src[3]); /* R */ + term[1] = factor[1] * (1.0f - src[3]); /* G */ + term[2] = factor[2] * (1.0f - src[3]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + term[0] = factor[0] * (1.0f - dst[3]); /* R */ + term[1] = factor[1] * (1.0f - dst[3]); /* G */ + term[2] = factor[2] * (1.0f - dst[3]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + term[0] = factor[0] * (1.0f - dst[0]); /* R */ + term[1] = factor[1] * (1.0f - dst[1]); /* G */ + term[2] = factor[2] * (1.0f - dst[2]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + term[0] = factor[0] * (1.0f - con[0]); /* R */ + term[1] = factor[1] * (1.0f - con[1]); /* G */ + term[2] = factor[2] * (1.0f - con[2]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + term[0] = factor[0] * (1.0f - con[3]); /* R */ + term[1] = factor[1] * (1.0f - con[3]); /* G */ + term[2] = factor[2] * (1.0f - con[3]); /* B */ + break; + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + assert(0); /* to do */ + break; + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + assert(0); /* to do */ + break; + default: + assert(0); + } + + /* + * Compute src/first term A + */ + switch (alpha_factor) { + case PIPE_BLENDFACTOR_ONE: + term[3] = factor[3]; /* A */ + break; + case PIPE_BLENDFACTOR_SRC_COLOR: + case PIPE_BLENDFACTOR_SRC_ALPHA: + term[3] = factor[3] * src[3]; /* A */ + break; + case PIPE_BLENDFACTOR_DST_COLOR: + case PIPE_BLENDFACTOR_DST_ALPHA: + term[3] = factor[3] * dst[3]; /* A */ + break; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + term[3] = src[3]; /* A */ + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + case PIPE_BLENDFACTOR_CONST_ALPHA: + term[3] = factor[3] * con[3]; /* A */ + break; + case PIPE_BLENDFACTOR_ZERO: + term[3] = 0.0f; /* A */ + break; + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + term[3] = factor[3] * (1.0f - src[3]); /* A */ + break; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + term[3] = factor[3] * (1.0f - dst[3]); /* A */ + break; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + term[3] = factor[3] * (1.0f - con[3]); + break; + default: + assert(0); + } +} + + +static void +compute_blend_ref(const struct pipe_blend_state *blend, + const double *src, + const double *dst, + const double *con, + double *res) +{ + double src_term[4]; + double dst_term[4]; + + compute_blend_ref_term(blend->rgb_src_factor, blend->alpha_src_factor, src, src, dst, con, src_term); + compute_blend_ref_term(blend->rgb_dst_factor, blend->alpha_dst_factor, dst, src, dst, con, dst_term); + + /* + * Combine RGB terms + */ + switch (blend->rgb_func) { + case PIPE_BLEND_ADD: + ADD_SAT(res[0], src_term[0], dst_term[0]); /* R */ + ADD_SAT(res[1], src_term[1], dst_term[1]); /* G */ + ADD_SAT(res[2], src_term[2], dst_term[2]); /* B */ + break; + case PIPE_BLEND_SUBTRACT: + SUB_SAT(res[0], src_term[0], dst_term[0]); /* R */ + SUB_SAT(res[1], src_term[1], dst_term[1]); /* G */ + SUB_SAT(res[2], src_term[2], dst_term[2]); /* B */ + break; + case PIPE_BLEND_REVERSE_SUBTRACT: + SUB_SAT(res[0], dst_term[0], src_term[0]); /* R */ + SUB_SAT(res[1], dst_term[1], src_term[1]); /* G */ + SUB_SAT(res[2], dst_term[2], src_term[2]); /* B */ + break; + case PIPE_BLEND_MIN: + res[0] = MIN2(src_term[0], dst_term[0]); /* R */ + res[1] = MIN2(src_term[1], dst_term[1]); /* G */ + res[2] = MIN2(src_term[2], dst_term[2]); /* B */ + break; + case PIPE_BLEND_MAX: + res[0] = MAX2(src_term[0], dst_term[0]); /* R */ + res[1] = MAX2(src_term[1], dst_term[1]); /* G */ + res[2] = MAX2(src_term[2], dst_term[2]); /* B */ + break; + default: + assert(0); + } + + /* + * Combine A terms + */ + switch (blend->alpha_func) { + case PIPE_BLEND_ADD: + ADD_SAT(res[3], src_term[3], dst_term[3]); /* A */ + break; + case PIPE_BLEND_SUBTRACT: + SUB_SAT(res[3], src_term[3], dst_term[3]); /* A */ + break; + case PIPE_BLEND_REVERSE_SUBTRACT: + SUB_SAT(res[3], dst_term[3], src_term[3]); /* A */ + break; + case PIPE_BLEND_MIN: + res[3] = MIN2(src_term[3], dst_term[3]); /* A */ + break; + case PIPE_BLEND_MAX: + res[3] = MAX2(src_term[3], dst_term[3]); /* A */ + break; + default: + assert(0); + } +} + + +static boolean +test_one(unsigned verbose, + FILE *fp, + const struct pipe_blend_state *blend, + enum vector_mode mode, + union lp_type type) +{ + LLVMModuleRef module = NULL; + LLVMValueRef func = NULL; + LLVMExecutionEngineRef engine = NULL; + LLVMModuleProviderRef provider = NULL; + LLVMPassManagerRef pass = NULL; + char *error = NULL; + blend_test_ptr_t blend_test_ptr; + boolean success; + const unsigned n = 32; + int64_t cycles[n]; + double cycles_avg = 0.0; + unsigned i, j; + + if(verbose >= 1) + dump_blend_type(stdout, blend, mode, type); + + module = LLVMModuleCreateWithName("test"); + + func = add_blend_test(module, blend, mode, type); + + if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) { + LLVMDumpModule(module); + abort(); + } + LLVMDisposeMessage(error); + + provider = LLVMCreateModuleProviderForExistingModule(module); + if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) { + if(verbose < 1) + dump_blend_type(stderr, blend, mode, type); + fprintf(stderr, "%s\n", error); + LLVMDisposeMessage(error); + abort(); + } + +#if 0 + pass = LLVMCreatePassManager(); + LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass); + /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, + * but there are more on SVN. */ + LLVMAddConstantPropagationPass(pass); + LLVMAddInstructionCombiningPass(pass); + LLVMAddPromoteMemoryToRegisterPass(pass); + LLVMAddGVNPass(pass); + LLVMAddCFGSimplificationPass(pass); + LLVMRunPassManager(pass, module); +#else + (void)pass; +#endif + + if(verbose >= 2) + LLVMDumpModule(module); + + blend_test_ptr = (blend_test_ptr_t)LLVMGetPointerToGlobal(engine, func); + + if(verbose >= 2) + lp_disassemble(blend_test_ptr); + + success = TRUE; + for(i = 0; i < n && success; ++i) { + if(mode == AoS) { + uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t con[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t res[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + int64_t start_counter = 0; + int64_t end_counter = 0; + + random_vec(type, src); + random_vec(type, dst); + random_vec(type, con); + + { + double fsrc[LP_MAX_VECTOR_LENGTH]; + double fdst[LP_MAX_VECTOR_LENGTH]; + double fcon[LP_MAX_VECTOR_LENGTH]; + double fref[LP_MAX_VECTOR_LENGTH]; + + read_vec(type, src, fsrc); + read_vec(type, dst, fdst); + read_vec(type, con, fcon); + + for(j = 0; j < type.length; j += 4) + compute_blend_ref(blend, fsrc + j, fdst + j, fcon + j, fref + j); + + write_vec(type, ref, fref); + } + + start_counter = rdtsc(); + blend_test_ptr(src, dst, con, res); + end_counter = rdtsc(); + + cycles[i] = end_counter - start_counter; + + if(!compare_vec(type, res, ref)) { + success = FALSE; + + if(verbose < 1) + dump_blend_type(stderr, blend, mode, type); + fprintf(stderr, "MISMATCH\n"); + + fprintf(stderr, " Src: "); + dump_vec(stderr, type, src); + fprintf(stderr, "\n"); + + fprintf(stderr, " Dst: "); + dump_vec(stderr, type, dst); + fprintf(stderr, "\n"); + + fprintf(stderr, " Con: "); + dump_vec(stderr, type, con); + fprintf(stderr, "\n"); + + fprintf(stderr, " Res: "); + dump_vec(stderr, type, res); + fprintf(stderr, "\n"); + + fprintf(stderr, " Ref: "); + dump_vec(stderr, type, ref); + fprintf(stderr, "\n"); + } + } + + if(mode == SoA) { + const unsigned stride = type.length*type.width/8; + uint8_t src[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t dst[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t con[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t res[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + uint8_t ref[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8]; + int64_t start_counter = 0; + int64_t end_counter = 0; + boolean mismatch; + + for(j = 0; j < 4; ++j) { + random_vec(type, src + j*stride); + random_vec(type, dst + j*stride); + random_vec(type, con + j*stride); + } + + { + double fsrc[4]; + double fdst[4]; + double fcon[4]; + double fref[4]; + unsigned k; + + for(k = 0; k < type.length; ++k) { + for(j = 0; j < 4; ++j) { + fsrc[j] = read_elem(type, src + j*stride, k); + fdst[j] = read_elem(type, dst + j*stride, k); + fcon[j] = read_elem(type, con + j*stride, k); + } + + compute_blend_ref(blend, fsrc, fdst, fcon, fref); + + for(j = 0; j < 4; ++j) + write_elem(type, ref + j*stride, k, fref[j]); + } + } + + start_counter = rdtsc(); + blend_test_ptr(src, dst, con, res); + end_counter = rdtsc(); + + cycles[i] = end_counter - start_counter; + + mismatch = FALSE; + for (j = 0; j < 4; ++j) + if(!compare_vec(type, res + j*stride, ref + j*stride)) + mismatch = TRUE; + + if (mismatch) { + success = FALSE; + + if(verbose < 1) + dump_blend_type(stderr, blend, mode, type); + fprintf(stderr, "MISMATCH\n"); + for(j = 0; j < 4; ++j) { + char channel = "RGBA"[j]; + fprintf(stderr, " Src%c: ", channel); + dump_vec(stderr, type, src + j*stride); + fprintf(stderr, "\n"); + + fprintf(stderr, " Dst%c: ", channel); + dump_vec(stderr, type, dst + j*stride); + fprintf(stderr, "\n"); + + fprintf(stderr, " Con%c: ", channel); + dump_vec(stderr, type, con + j*stride); + fprintf(stderr, "\n"); + + fprintf(stderr, " Res%c: ", channel); + dump_vec(stderr, type, res + j*stride); + fprintf(stderr, "\n"); + + fprintf(stderr, " Ref%c: ", channel); + dump_vec(stderr, type, ref + j*stride); + fprintf(stderr, "\n"); + } + } + } + } + + /* + * Unfortunately the output of cycle counter is not very reliable as it comes + * -- sometimes we get outliers (due IRQs perhaps?) which are + * better removed to avoid random or biased data. + */ + { + double sum = 0.0, sum2 = 0.0; + double avg, std; + unsigned m; + + for(i = 0; i < n; ++i) { + sum += cycles[i]; + sum2 += cycles[i]*cycles[i]; + } + + avg = sum/n; + std = sqrtf((sum2 - n*avg*avg)/n); + + m = 0; + sum = 0.0; + for(i = 0; i < n; ++i) { + if(fabs(cycles[i] - avg) <= 4.0*std) { + sum += cycles[i]; + ++m; + } + } + + cycles_avg = sum/m; + + } + + if(fp) + write_tsv_row(fp, blend, mode, type, cycles_avg, success); + + if (!success) { + if(verbose < 2) + LLVMDumpModule(module); + LLVMWriteBitcodeToFile(module, "blend.bc"); + fprintf(stderr, "blend.bc written\n"); + fprintf(stderr, "Invoke as \"llc -o - blend.bc\"\n"); + abort(); + } + + LLVMFreeMachineCodeForFunction(engine, func); + + LLVMDisposeExecutionEngine(engine); + if(pass) + LLVMDisposePassManager(pass); + + return success; +} + + +const unsigned +blend_factors[] = { + PIPE_BLENDFACTOR_ZERO, + PIPE_BLENDFACTOR_ONE, + PIPE_BLENDFACTOR_SRC_COLOR, + PIPE_BLENDFACTOR_SRC_ALPHA, + PIPE_BLENDFACTOR_DST_COLOR, + PIPE_BLENDFACTOR_DST_ALPHA, + PIPE_BLENDFACTOR_CONST_COLOR, + PIPE_BLENDFACTOR_CONST_ALPHA, +#if 0 + PIPE_BLENDFACTOR_SRC1_COLOR, + PIPE_BLENDFACTOR_SRC1_ALPHA, +#endif + PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE, + PIPE_BLENDFACTOR_INV_SRC_COLOR, + PIPE_BLENDFACTOR_INV_SRC_ALPHA, + PIPE_BLENDFACTOR_INV_DST_COLOR, + PIPE_BLENDFACTOR_INV_DST_ALPHA, + PIPE_BLENDFACTOR_INV_CONST_COLOR, + PIPE_BLENDFACTOR_INV_CONST_ALPHA, +#if 0 + PIPE_BLENDFACTOR_INV_SRC1_COLOR, + PIPE_BLENDFACTOR_INV_SRC1_ALPHA, +#endif +}; + + +const unsigned +blend_funcs[] = { + PIPE_BLEND_ADD, + PIPE_BLEND_SUBTRACT, + PIPE_BLEND_REVERSE_SUBTRACT, + PIPE_BLEND_MIN, + PIPE_BLEND_MAX +}; + + +const union lp_type blend_types[] = { + /* float, fixed, sign, norm, width, len */ + {{ TRUE, FALSE, FALSE, TRUE, 32, 4 }}, /* f32 x 4 */ + {{ FALSE, FALSE, FALSE, TRUE, 8, 16 }}, /* u8n x 16 */ +}; + + +const unsigned num_funcs = sizeof(blend_funcs)/sizeof(blend_funcs[0]); +const unsigned num_factors = sizeof(blend_factors)/sizeof(blend_factors[0]); +const unsigned num_types = sizeof(blend_types)/sizeof(blend_types[0]); + + +boolean +test_all(unsigned verbose, FILE *fp) +{ + const unsigned *rgb_func; + const unsigned *rgb_src_factor; + const unsigned *rgb_dst_factor; + const unsigned *alpha_func; + const unsigned *alpha_src_factor; + const unsigned *alpha_dst_factor; + struct pipe_blend_state blend; + enum vector_mode mode; + const union lp_type *type; + bool success = TRUE; + + for(rgb_func = blend_funcs; rgb_func < &blend_funcs[num_funcs]; ++rgb_func) { + for(alpha_func = blend_funcs; alpha_func < &blend_funcs[num_funcs]; ++alpha_func) { + for(rgb_src_factor = blend_factors; rgb_src_factor < &blend_factors[num_factors]; ++rgb_src_factor) { + for(rgb_dst_factor = blend_factors; rgb_dst_factor <= rgb_src_factor; ++rgb_dst_factor) { + for(alpha_src_factor = blend_factors; alpha_src_factor < &blend_factors[num_factors]; ++alpha_src_factor) { + for(alpha_dst_factor = blend_factors; alpha_dst_factor <= alpha_src_factor; ++alpha_dst_factor) { + for(mode = 0; mode < 2; ++mode) { + for(type = blend_types; type < &blend_types[num_types]; ++type) { + + if(*rgb_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE || + *alpha_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) + continue; + + memset(&blend, 0, sizeof blend); + blend.blend_enable = 1; + blend.rgb_func = *rgb_func; + blend.rgb_src_factor = *rgb_src_factor; + blend.rgb_dst_factor = *rgb_dst_factor; + blend.alpha_func = *alpha_func; + blend.alpha_src_factor = *alpha_src_factor; + blend.alpha_dst_factor = *alpha_dst_factor; + blend.colormask = PIPE_MASK_RGBA; + + if(!test_one(verbose, fp, &blend, mode, *type)) + success = FALSE; + + } + } + } + } + } + } + } + } + + return success; +} + + +boolean +test_some(unsigned verbose, FILE *fp, unsigned long n) +{ + const unsigned *rgb_func; + const unsigned *rgb_src_factor; + const unsigned *rgb_dst_factor; + const unsigned *alpha_func; + const unsigned *alpha_src_factor; + const unsigned *alpha_dst_factor; + struct pipe_blend_state blend; + enum vector_mode mode; + const union lp_type *type; + unsigned long i; + bool success = TRUE; + + for(i = 0; i < n; ++i) { + rgb_func = &blend_funcs[random() % num_funcs]; + alpha_func = &blend_funcs[random() % num_funcs]; + rgb_src_factor = &blend_factors[random() % num_factors]; + alpha_src_factor = &blend_factors[random() % num_factors]; + + do { + rgb_dst_factor = &blend_factors[random() % num_factors]; + } while(*rgb_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE); + + do { + alpha_dst_factor = &blend_factors[random() % num_factors]; + } while(*alpha_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE); + + mode = random() & 1; + + type = &blend_types[random() % num_types]; + + memset(&blend, 0, sizeof blend); + blend.blend_enable = 1; + blend.rgb_func = *rgb_func; + blend.rgb_src_factor = *rgb_src_factor; + blend.rgb_dst_factor = *rgb_dst_factor; + blend.alpha_func = *alpha_func; + blend.alpha_src_factor = *alpha_src_factor; + blend.alpha_dst_factor = *alpha_dst_factor; + blend.colormask = PIPE_MASK_RGBA; + + if(!test_one(verbose, fp, &blend, mode, *type)) + success = FALSE; + } + + return success; +} diff --git a/src/gallium/drivers/llvmpipe/lp_test_conv.c b/src/gallium/drivers/llvmpipe/lp_test_conv.c new file mode 100644 index 00000000000..e6489834af5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test_conv.c @@ -0,0 +1,427 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Unit tests for type conversion. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "lp_bld_type.h" +#include "lp_bld_const.h" +#include "lp_bld_conv.h" +#include "lp_bld_debug.h" +#include "lp_test.h" + + +typedef void (*conv_test_ptr_t)(const void *src, const void *dst); + + +void +write_tsv_header(FILE *fp) +{ + fprintf(fp, + "result\t" + "cycles_per_channel\t" + "src_type\t" + "dst_type\n"); + + fflush(fp); +} + + +static void +write_tsv_row(FILE *fp, + union lp_type src_type, + union lp_type dst_type, + double cycles, + boolean success) +{ + fprintf(fp, "%s\t", success ? "pass" : "fail"); + + fprintf(fp, "%.1f\t", cycles / MAX2(src_type.length, dst_type.length)); + + dump_type(fp, src_type); + fprintf(fp, "\t"); + + dump_type(fp, dst_type); + fprintf(fp, "\n"); + + fflush(fp); +} + + +static void +dump_conv_types(FILE *fp, + union lp_type src_type, + union lp_type dst_type) +{ + fprintf(fp, "src_type="); + dump_type(fp, src_type); + + fprintf(fp, " dst_type="); + dump_type(fp, dst_type); + + fprintf(fp, " ...\n"); + fflush(fp); +} + + +static LLVMValueRef +add_conv_test(LLVMModuleRef module, + union lp_type src_type, unsigned num_srcs, + union lp_type dst_type, unsigned num_dsts) +{ + LLVMTypeRef args[2]; + LLVMValueRef func; + LLVMValueRef src_ptr; + LLVMValueRef dst_ptr; + LLVMBasicBlockRef block; + LLVMBuilderRef builder; + LLVMValueRef src[LP_MAX_VECTOR_LENGTH]; + LLVMValueRef dst[LP_MAX_VECTOR_LENGTH]; + unsigned i; + + args[0] = LLVMPointerType(lp_build_vec_type(src_type), 0); + args[1] = LLVMPointerType(lp_build_vec_type(dst_type), 0); + + func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidType(), args, 2, 0)); + LLVMSetFunctionCallConv(func, LLVMCCallConv); + src_ptr = LLVMGetParam(func, 0); + dst_ptr = LLVMGetParam(func, 1); + + block = LLVMAppendBasicBlock(func, "entry"); + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, block); + + for(i = 0; i < num_srcs; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef ptr = LLVMBuildGEP(builder, src_ptr, &index, 1, ""); + src[i] = LLVMBuildLoad(builder, ptr, ""); + } + + lp_build_conv(builder, src_type, dst_type, src, num_srcs, dst, num_dsts); + + for(i = 0; i < num_dsts; ++i) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); + LLVMValueRef ptr = LLVMBuildGEP(builder, dst_ptr, &index, 1, ""); + LLVMBuildStore(builder, dst[i], ptr); + } + + LLVMBuildRetVoid(builder);; + + LLVMDisposeBuilder(builder); + return func; +} + + +static boolean +test_one(unsigned verbose, + FILE *fp, + union lp_type src_type, + union lp_type dst_type) +{ + LLVMModuleRef module = NULL; + LLVMValueRef func = NULL; + LLVMExecutionEngineRef engine = NULL; + LLVMModuleProviderRef provider = NULL; + LLVMPassManagerRef pass = NULL; + char *error = NULL; + conv_test_ptr_t conv_test_ptr; + boolean success; + const unsigned n = 32; + int64_t cycles[n]; + double cycles_avg = 0.0; + unsigned num_srcs; + unsigned num_dsts; + double eps; + unsigned i, j; + + if(verbose >= 1) + dump_conv_types(stdout, src_type, dst_type); + + if(src_type.length > dst_type.length) { + num_srcs = 1; + num_dsts = src_type.length/dst_type.length; + } + else { + num_dsts = 1; + num_srcs = dst_type.length/src_type.length; + } + + assert(src_type.width * src_type.length == dst_type.width * dst_type.length); + + /* We must not loose or gain channels. Only precision */ + assert(src_type.length * num_srcs == dst_type.length * num_dsts); + + eps = MAX2(lp_const_eps(src_type), lp_const_eps(dst_type)); + + module = LLVMModuleCreateWithName("test"); + + func = add_conv_test(module, src_type, num_srcs, dst_type, num_dsts); + + if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) { + LLVMDumpModule(module); + abort(); + } + LLVMDisposeMessage(error); + + provider = LLVMCreateModuleProviderForExistingModule(module); + if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) { + if(verbose < 1) + dump_conv_types(stderr, src_type, dst_type); + fprintf(stderr, "%s\n", error); + LLVMDisposeMessage(error); + abort(); + } + +#if 0 + pass = LLVMCreatePassManager(); + LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass); + /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, + * but there are more on SVN. */ + LLVMAddConstantPropagationPass(pass); + LLVMAddInstructionCombiningPass(pass); + LLVMAddPromoteMemoryToRegisterPass(pass); + LLVMAddGVNPass(pass); + LLVMAddCFGSimplificationPass(pass); + LLVMRunPassManager(pass, module); +#else + (void)pass; +#endif + + if(verbose >= 2) + LLVMDumpModule(module); + + conv_test_ptr = (conv_test_ptr_t)LLVMGetPointerToGlobal(engine, func); + + if(verbose >= 2) + lp_disassemble(conv_test_ptr); + + success = TRUE; + for(i = 0; i < n && success; ++i) { + unsigned src_stride = src_type.length*src_type.width/8; + unsigned dst_stride = dst_type.length*dst_type.width/8; + uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + double fref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; + int64_t start_counter = 0; + int64_t end_counter = 0; + + for(j = 0; j < num_srcs; ++j) { + random_vec(src_type, src + j*src_stride); + read_vec(src_type, src + j*src_stride, fref + j*src_type.length); + } + + for(j = 0; j < num_dsts; ++j) { + write_vec(dst_type, ref + j*dst_stride, fref + j*dst_type.length); + } + + start_counter = rdtsc(); + conv_test_ptr(src, dst); + end_counter = rdtsc(); + + cycles[i] = end_counter - start_counter; + + for(j = 0; j < num_dsts; ++j) { + if(!compare_vec_with_eps(dst_type, dst + j*dst_stride, ref + j*dst_stride, eps)) + success = FALSE; + } + + if (!success) { + if(verbose < 1) + dump_conv_types(stderr, src_type, dst_type); + fprintf(stderr, "MISMATCH\n"); + + for(j = 0; j < num_srcs; ++j) { + fprintf(stderr, " Src%u: ", j); + dump_vec(stderr, src_type, src + j*src_stride); + fprintf(stderr, "\n"); + } + +#if 1 + fprintf(stderr, " Ref: "); + for(j = 0; j < src_type.length*num_srcs; ++j) + fprintf(stderr, " %f", fref[j]); + fprintf(stderr, "\n"); +#endif + + for(j = 0; j < num_dsts; ++j) { + fprintf(stderr, " Dst%u: ", j); + dump_vec(stderr, dst_type, dst + j*dst_stride); + fprintf(stderr, "\n"); + + fprintf(stderr, " Ref%u: ", j); + dump_vec(stderr, dst_type, ref + j*dst_stride); + fprintf(stderr, "\n"); + } + } + } + + /* + * Unfortunately the output of cycle counter is not very reliable as it comes + * -- sometimes we get outliers (due IRQs perhaps?) which are + * better removed to avoid random or biased data. + */ + { + double sum = 0.0, sum2 = 0.0; + double avg, std; + unsigned m; + + for(i = 0; i < n; ++i) { + sum += cycles[i]; + sum2 += cycles[i]*cycles[i]; + } + + avg = sum/n; + std = sqrtf((sum2 - n*avg*avg)/n); + + m = 0; + sum = 0.0; + for(i = 0; i < n; ++i) { + if(fabs(cycles[i] - avg) <= 4.0*std) { + sum += cycles[i]; + ++m; + } + } + + cycles_avg = sum/m; + + } + + if(fp) + write_tsv_row(fp, src_type, dst_type, cycles_avg, success); + + if (!success) { + static boolean firsttime = TRUE; + if(firsttime) { + if(verbose < 2) + LLVMDumpModule(module); + LLVMWriteBitcodeToFile(module, "conv.bc"); + fprintf(stderr, "conv.bc written\n"); + fprintf(stderr, "Invoke as \"llc -o - conv.bc\"\n"); + firsttime = FALSE; + //abort(); + } + } + + LLVMFreeMachineCodeForFunction(engine, func); + + LLVMDisposeExecutionEngine(engine); + if(pass) + LLVMDisposePassManager(pass); + + return success; +} + + +const union lp_type conv_types[] = { + /* float, fixed, sign, norm, width, len */ + + {{ TRUE, FALSE, TRUE, TRUE, 32, 4 }}, + {{ TRUE, FALSE, TRUE, FALSE, 32, 4 }}, + {{ TRUE, FALSE, FALSE, TRUE, 32, 4 }}, + {{ TRUE, FALSE, FALSE, FALSE, 32, 4 }}, + + /* TODO: test fixed formats too */ + + {{ FALSE, FALSE, TRUE, TRUE, 16, 8 }}, + {{ FALSE, FALSE, TRUE, FALSE, 16, 8 }}, + {{ FALSE, FALSE, FALSE, TRUE, 16, 8 }}, + {{ FALSE, FALSE, FALSE, FALSE, 16, 8 }}, + + {{ FALSE, FALSE, TRUE, TRUE, 32, 4 }}, + {{ FALSE, FALSE, TRUE, FALSE, 32, 4 }}, + {{ FALSE, FALSE, FALSE, TRUE, 32, 4 }}, + {{ FALSE, FALSE, FALSE, FALSE, 32, 4 }}, + + {{ FALSE, FALSE, TRUE, TRUE, 16, 8 }}, + {{ FALSE, FALSE, TRUE, FALSE, 16, 8 }}, + {{ FALSE, FALSE, FALSE, TRUE, 16, 8 }}, + {{ FALSE, FALSE, FALSE, FALSE, 16, 8 }}, + + {{ FALSE, FALSE, TRUE, TRUE, 8, 16 }}, + {{ FALSE, FALSE, TRUE, FALSE, 8, 16 }}, + {{ FALSE, FALSE, FALSE, TRUE, 8, 16 }}, + {{ FALSE, FALSE, FALSE, FALSE, 8, 16 }}, +}; + + +const unsigned num_types = sizeof(conv_types)/sizeof(conv_types[0]); + + +boolean +test_all(unsigned verbose, FILE *fp) +{ + const union lp_type *src_type; + const union lp_type *dst_type; + bool success = TRUE; + + for(src_type = conv_types; src_type < &conv_types[num_types]; ++src_type) { + for(dst_type = conv_types; dst_type < &conv_types[num_types]; ++dst_type) { + + if(src_type == dst_type) + continue; + + if(src_type->norm != dst_type->norm) + continue; + + if(!test_one(verbose, fp, *src_type, *dst_type)) + success = FALSE; + + } + } + + return success; +} + + +boolean +test_some(unsigned verbose, FILE *fp, unsigned long n) +{ + const union lp_type *src_type; + const union lp_type *dst_type; + unsigned long i; + bool success = TRUE; + + for(i = 0; i < n; ++i) { + src_type = &conv_types[random() % num_types]; + + do { + dst_type = &conv_types[random() % num_types]; + } while (src_type == dst_type || src_type->norm != dst_type->norm); + + if(!test_one(verbose, fp, *src_type, *dst_type)) + success = FALSE; + } + + return success; +} diff --git a/src/gallium/drivers/llvmpipe/lp_test_format.c b/src/gallium/drivers/llvmpipe/lp_test_format.c new file mode 100644 index 00000000000..1d192355eed --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test_format.c @@ -0,0 +1,272 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include <stdlib.h> +#include <stdio.h> + +#include <llvm-c/Core.h> +#include <llvm-c/Analysis.h> +#include <llvm-c/ExecutionEngine.h> +#include <llvm-c/Target.h> +#include <llvm-c/Transforms/Scalar.h> + +#include "util/u_format.h" + +#include "lp_bld_flow.h" +#include "lp_bld_format.h" + + +struct pixel_test_case +{ + enum pipe_format format; + uint32_t packed; + double unpacked[4]; +}; + + +struct pixel_test_case test_cases[] = +{ + {PIPE_FORMAT_R5G6B5_UNORM, 0x0000, {0.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_R5G6B5_UNORM, 0x001f, {0.0, 0.0, 1.0, 1.0}}, + {PIPE_FORMAT_R5G6B5_UNORM, 0x07e0, {0.0, 1.0, 0.0, 1.0}}, + {PIPE_FORMAT_R5G6B5_UNORM, 0xf800, {1.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_R5G6B5_UNORM, 0xffff, {1.0, 1.0, 1.0, 1.0}}, + + {PIPE_FORMAT_A1R5G5B5_UNORM, 0x0000, {0.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_A1R5G5B5_UNORM, 0x001f, {0.0, 0.0, 1.0, 0.0}}, + {PIPE_FORMAT_A1R5G5B5_UNORM, 0x03e0, {0.0, 1.0, 0.0, 0.0}}, + {PIPE_FORMAT_A1R5G5B5_UNORM, 0x7c00, {1.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_A1R5G5B5_UNORM, 0x8000, {0.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_A1R5G5B5_UNORM, 0xffff, {1.0, 1.0, 1.0, 1.0}}, + + {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_A8R8G8B8_UNORM, 0x000000ff, {0.0, 0.0, 1.0, 0.0}}, + {PIPE_FORMAT_A8R8G8B8_UNORM, 0x0000ff00, {0.0, 1.0, 0.0, 0.0}}, + {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00ff0000, {1.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_A8R8G8B8_UNORM, 0xff000000, {0.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_A8R8G8B8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}}, + +#if 0 + {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_R8G8B8A8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_R8G8B8A8_UNORM, 0x0000ff00, {0.0, 0.0, 1.0, 0.0}}, + {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}}, + {PIPE_FORMAT_R8G8B8A8_UNORM, 0xff000000, {1.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_R8G8B8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}}, +#endif + + {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_B8G8R8A8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}}, + {PIPE_FORMAT_B8G8R8A8_UNORM, 0x0000ff00, {1.0, 0.0, 0.0, 0.0}}, + {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}}, + {PIPE_FORMAT_B8G8R8A8_UNORM, 0xff000000, {0.0, 0.0, 1.0, 0.0}}, + {PIPE_FORMAT_B8G8R8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}}, +}; + + +typedef void (*load_ptr_t)(const void *, float *); + + +static LLVMValueRef +add_load_rgba_test(LLVMModuleRef module, + enum pipe_format format) +{ + LLVMTypeRef args[2]; + LLVMValueRef func; + LLVMValueRef ptr; + LLVMValueRef rgba_ptr; + LLVMBasicBlockRef block; + LLVMBuilderRef builder; + LLVMValueRef rgba; + struct lp_build_loop_state loop; + + args[0] = LLVMPointerType(LLVMInt8Type(), 0); + args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0); + + func = LLVMAddFunction(module, "load", LLVMFunctionType(LLVMVoidType(), args, 2, 0)); + LLVMSetFunctionCallConv(func, LLVMCCallConv); + ptr = LLVMGetParam(func, 0); + rgba_ptr = LLVMGetParam(func, 1); + + block = LLVMAppendBasicBlock(func, "entry"); + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, block); + + lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 1, 0), &loop); + + rgba = lp_build_load_rgba(builder, format, ptr); + LLVMBuildStore(builder, rgba, rgba_ptr); + + lp_build_loop_end(builder, LLVMConstInt(LLVMInt32Type(), 4, 0), NULL, &loop); + + LLVMBuildRetVoid(builder); + + LLVMDisposeBuilder(builder); + return func; +} + + +typedef void (*store_ptr_t)(void *, const float *); + + +static LLVMValueRef +add_store_rgba_test(LLVMModuleRef module, + enum pipe_format format) +{ + LLVMTypeRef args[2]; + LLVMValueRef func; + LLVMValueRef ptr; + LLVMValueRef rgba_ptr; + LLVMBasicBlockRef block; + LLVMBuilderRef builder; + LLVMValueRef rgba; + + args[0] = LLVMPointerType(LLVMInt8Type(), 0); + args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0); + + func = LLVMAddFunction(module, "store", LLVMFunctionType(LLVMVoidType(), args, 2, 0)); + LLVMSetFunctionCallConv(func, LLVMCCallConv); + ptr = LLVMGetParam(func, 0); + rgba_ptr = LLVMGetParam(func, 1); + + block = LLVMAppendBasicBlock(func, "entry"); + builder = LLVMCreateBuilder(); + LLVMPositionBuilderAtEnd(builder, block); + + rgba = LLVMBuildLoad(builder, rgba_ptr, ""); + + lp_build_store_rgba(builder, format, ptr, rgba); + + LLVMBuildRetVoid(builder); + + LLVMDisposeBuilder(builder); + return func; +} + + +static boolean +test_format(const struct pixel_test_case *test) +{ + LLVMModuleRef module = NULL; + LLVMValueRef load = NULL; + LLVMValueRef store = NULL; + LLVMExecutionEngineRef engine = NULL; + LLVMModuleProviderRef provider = NULL; + LLVMPassManagerRef pass = NULL; + char *error = NULL; + const struct util_format_description *desc; + load_ptr_t load_ptr; + store_ptr_t store_ptr; + float unpacked[4]; + unsigned packed; + boolean success; + unsigned i; + + desc = util_format_description(test->format); + fprintf(stderr, "%s\n", desc->name); + + module = LLVMModuleCreateWithName("test"); + + load = add_load_rgba_test(module, test->format); + store = add_store_rgba_test(module, test->format); + + if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) { + LLVMDumpModule(module); + abort(); + } + LLVMDisposeMessage(error); + + provider = LLVMCreateModuleProviderForExistingModule(module); + if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) { + fprintf(stderr, "%s\n", error); + LLVMDisposeMessage(error); + abort(); + } + +#if 0 + pass = LLVMCreatePassManager(); + LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass); + /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, + * but there are more on SVN. */ + LLVMAddConstantPropagationPass(pass); + LLVMAddInstructionCombiningPass(pass); + LLVMAddPromoteMemoryToRegisterPass(pass); + LLVMAddGVNPass(pass); + LLVMAddCFGSimplificationPass(pass); + LLVMRunPassManager(pass, module); +#else + (void)pass; +#endif + + load_ptr = (load_ptr_t) LLVMGetPointerToGlobal(engine, load); + store_ptr = (store_ptr_t)LLVMGetPointerToGlobal(engine, store); + + memset(unpacked, 0, sizeof unpacked); + packed = 0; + + load_ptr(&test->packed, unpacked); + store_ptr(&packed, unpacked); + + success = TRUE; + if(test->packed != packed) + success = FALSE; + for(i = 0; i < 4; ++i) + if(test->unpacked[i] != unpacked[i]) + success = FALSE; + + if (!success) { + printf("FAILED\n"); + printf(" Packed: %08x\n", test->packed); + printf(" %08x\n", packed); + printf(" Unpacked: %f %f %f %f\n", unpacked[0], unpacked[1], unpacked[2], unpacked[3]); + printf(" %f %f %f %f\n", test->unpacked[0], test->unpacked[1], test->unpacked[2], test->unpacked[3]); + LLVMDumpModule(module); + } + + LLVMFreeMachineCodeForFunction(engine, store); + LLVMFreeMachineCodeForFunction(engine, load); + + LLVMDisposeExecutionEngine(engine); + if(pass) + LLVMDisposePassManager(pass); + + return success; +} + + +int main(int argc, char **argv) +{ + unsigned i; + int ret; + + for (i = 0; i < sizeof(test_cases)/sizeof(test_cases[0]); ++i) + if(!test_format(&test_cases[i])) + ret = 1; + + return ret; +} diff --git a/src/gallium/drivers/llvmpipe/lp_test_main.c b/src/gallium/drivers/llvmpipe/lp_test_main.c new file mode 100644 index 00000000000..49213fb4f0b --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test_main.c @@ -0,0 +1,384 @@ +/************************************************************************** + * + * Copyright 2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * @file + * Shared testing code. + * + * @author Jose Fonseca <jfonseca@vmware.com> + */ + + +#include "lp_bld_const.h" +#include "lp_test.h" + + +void +dump_type(FILE *fp, + union lp_type type) +{ + fprintf(fp, "%s%s%u%sx%u", + type.sign ? (type.floating || type.fixed ? "" : "s") : "u", + type.floating ? "f" : (type.fixed ? "h" : "i"), + type.width, + type.norm ? "n" : "", + type.length); +} + + +double +read_elem(union lp_type type, const void *src, unsigned index) +{ + double scale = lp_const_scale(type); + double value; + assert(index < type.length); + if (type.floating) { + switch(type.width) { + case 32: + value = *((const float *)src + index); + break; + case 64: + value = *((const double *)src + index); + break; + default: + assert(0); + return 0.0; + } + } + else { + if(type.sign) { + switch(type.width) { + case 8: + value = *((const int8_t *)src + index); + break; + case 16: + value = *((const int16_t *)src + index); + break; + case 32: + value = *((const int32_t *)src + index); + break; + case 64: + value = *((const int64_t *)src + index); + break; + default: + assert(0); + return 0.0; + } + } + else { + switch(type.width) { + case 8: + value = *((const uint8_t *)src + index); + break; + case 16: + value = *((const uint16_t *)src + index); + break; + case 32: + value = *((const uint32_t *)src + index); + break; + case 64: + value = *((const uint64_t *)src + index); + break; + default: + assert(0); + return 0.0; + } + } + } + return value/scale; +} + + +void +write_elem(union lp_type type, void *dst, unsigned index, double value) +{ + assert(index < type.length); + if(!type.sign && value < 0.0) + value = 0.0; + if(type.norm && value < -1.0) + value = -1.0; + if(type.norm && value > 1.0) + value = 1.0; + if (type.floating) { + switch(type.width) { + case 32: + *((float *)dst + index) = (float)(value); + break; + case 64: + *((double *)dst + index) = value; + break; + default: + assert(0); + } + } + else { + double scale = lp_const_scale(type); + value = round(value*scale); + if(type.sign) { + long long lvalue = (long long)value; + lvalue = MIN2(lvalue, ((long long)1 << (type.width - 1)) - 1); + switch(type.width) { + case 8: + *((int8_t *)dst + index) = (int8_t)lvalue; + break; + case 16: + *((int16_t *)dst + index) = (int16_t)lvalue; + break; + case 32: + *((int32_t *)dst + index) = (int32_t)lvalue; + break; + case 64: + *((int64_t *)dst + index) = (int64_t)lvalue; + break; + default: + assert(0); + } + } + else { + unsigned long long lvalue = (long long)value; + lvalue = MIN2(lvalue, ((unsigned long long)1 << type.width) - 1); + switch(type.width) { + case 8: + *((uint8_t *)dst + index) = (uint8_t)lvalue; + break; + case 16: + *((uint16_t *)dst + index) = (uint16_t)lvalue; + break; + case 32: + *((uint32_t *)dst + index) = (uint32_t)lvalue; + break; + case 64: + *((uint64_t *)dst + index) = (uint64_t)lvalue; + break; + default: + assert(0); + } + } + } +} + + +void +random_elem(union lp_type type, void *dst, unsigned index) +{ + double value; + assert(index < type.length); + value = (double)random()/(double)RAND_MAX; + if(!type.norm) { + unsigned long long mask; + if (type.floating) + mask = ~(unsigned long long)0; + else if (type.fixed) + mask = ((unsigned long long)1 << (type.width / 2)) - 1; + else if (type.sign) + mask = ((unsigned long long)1 << (type.width - 1)) - 1; + else + mask = ((unsigned long long)1 << type.width) - 1; + value += (double)(mask & random()); + } + if(!type.sign) + if(random() & 1) + value = -value; + write_elem(type, dst, index, value); +} + + +void +read_vec(union lp_type type, const void *src, double *dst) +{ + unsigned i; + for (i = 0; i < type.length; ++i) + dst[i] = read_elem(type, src, i); +} + + +void +write_vec(union lp_type type, void *dst, const double *src) +{ + unsigned i; + for (i = 0; i < type.length; ++i) + write_elem(type, dst, i, src[i]); +} + + +float +random_float(void) +{ + return (float)((double)random()/(double)RAND_MAX); +} + + +void +random_vec(union lp_type type, void *dst) +{ + unsigned i; + for (i = 0; i < type.length; ++i) + random_elem(type, dst, i); +} + + +boolean +compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps) +{ + unsigned i; + for (i = 0; i < type.length; ++i) { + double res_elem = read_elem(type, res, i); + double ref_elem = read_elem(type, ref, i); + double delta = fabs(res_elem - ref_elem); + if(delta >= 2.0*eps) + return FALSE; + } + + return TRUE; +} + + +boolean +compare_vec(union lp_type type, const void *res, const void *ref) +{ + double eps = lp_const_eps(type); + return compare_vec_with_eps(type, res, ref, eps); +} + + +void +dump_vec(FILE *fp, union lp_type type, const void *src) +{ + unsigned i; + for (i = 0; i < type.length; ++i) { + if(i) + fprintf(fp, " "); + if (type.floating) { + double value; + switch(type.width) { + case 32: + value = *((const float *)src + i); + break; + case 64: + value = *((const double *)src + i); + break; + default: + assert(0); + value = 0.0; + } + fprintf(fp, "%f", value); + } + else { + if(type.sign && !type.norm) { + long long value; + const char *format; + switch(type.width) { + case 8: + value = *((const int8_t *)src + i); + format = "%3lli"; + break; + case 16: + value = *((const int16_t *)src + i); + format = "%5lli"; + break; + case 32: + value = *((const int32_t *)src + i); + format = "%10lli"; + break; + case 64: + value = *((const int64_t *)src + i); + format = "%20lli"; + break; + default: + assert(0); + value = 0.0; + format = "?"; + } + fprintf(fp, format, value); + } + else { + unsigned long long value; + const char *format; + switch(type.width) { + case 8: + value = *((const uint8_t *)src + i); + format = type.norm ? "%2x" : "%4llu"; + break; + case 16: + value = *((const uint16_t *)src + i); + format = type.norm ? "%4x" : "%6llx"; + break; + case 32: + value = *((const uint32_t *)src + i); + format = type.norm ? "%8x" : "%11llx"; + break; + case 64: + value = *((const uint64_t *)src + i); + format = type.norm ? "%16x" : "%21llx"; + break; + default: + assert(0); + value = 0.0; + format = "?"; + } + fprintf(fp, format, value); + } + } + } +} + + +int main(int argc, char **argv) +{ + unsigned verbose = 0; + FILE *fp = NULL; + unsigned long n = 1000; + unsigned i; + boolean success; + + for(i = 1; i < argc; ++i) { + if(strcmp(argv[i], "-v") == 0) + ++verbose; + else if(strcmp(argv[i], "-o") == 0) + fp = fopen(argv[++i], "wt"); + else + n = atoi(argv[i]); + } + + if(fp) { + /* Warm up the caches */ + test_some(0, NULL, 100); + + write_tsv_header(fp); + } + + if(n) + success = test_some(verbose, fp, n); + else + success = test_all(verbose, fp); + + if(fp) + fclose(fp); + + return success ? 0 : 1; +} diff --git a/src/gallium/drivers/llvmpipe/lp_tex_cache.c b/src/gallium/drivers/llvmpipe/lp_tex_cache.c new file mode 100644 index 00000000000..23a94b5b0d5 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tex_cache.c @@ -0,0 +1,304 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Texture tile caching. + * + * Author: + * Brian Paul + */ + +#include "pipe/p_inlines.h" +#include "util/u_memory.h" +#include "util/u_tile.h" +#include "util/u_format.h" +#include "lp_context.h" +#include "lp_surface.h" +#include "lp_texture.h" +#include "lp_tex_cache.h" + + + +/** + * Return the position in the cache for the tile that contains win pos (x,y). + * We currently use a direct mapped cache so this is like a hack key. + * At some point we should investige something more sophisticated, like + * a LRU replacement policy. + */ +#define CACHE_POS(x, y) \ + (((x) + (y) * 5) % NUM_ENTRIES) + + + +/** + * Is the tile at (x,y) in cleared state? + */ +static INLINE uint +is_clear_flag_set(const uint *bitvec, union tex_tile_address addr) +{ + int pos, bit; + pos = addr.bits.y * (MAX_TEX_WIDTH / TEX_TILE_SIZE) + addr.bits.x; + assert(pos / 32 < (MAX_TEX_WIDTH / TEX_TILE_SIZE) * (MAX_TEX_HEIGHT / TEX_TILE_SIZE) / 32); + bit = bitvec[pos / 32] & (1 << (pos & 31)); + return bit; +} + + +/** + * Mark the tile at (x,y) as not cleared. + */ +static INLINE void +clear_clear_flag(uint *bitvec, union tex_tile_address addr) +{ + int pos; + pos = addr.bits.y * (MAX_TEX_WIDTH / TEX_TILE_SIZE) + addr.bits.x; + assert(pos / 32 < (MAX_TEX_WIDTH / TEX_TILE_SIZE) * (MAX_TEX_HEIGHT / TEX_TILE_SIZE) / 32); + bitvec[pos / 32] &= ~(1 << (pos & 31)); +} + + +struct llvmpipe_tex_tile_cache * +lp_create_tex_tile_cache( struct pipe_screen *screen ) +{ + struct llvmpipe_tex_tile_cache *tc; + uint pos; + + tc = CALLOC_STRUCT( llvmpipe_tex_tile_cache ); + if (tc) { + tc->screen = screen; + for (pos = 0; pos < NUM_ENTRIES; pos++) { + tc->entries[pos].addr.bits.invalid = 1; + } + tc->last_tile = &tc->entries[0]; /* any tile */ + } + return tc; +} + + +void +lp_destroy_tex_tile_cache(struct llvmpipe_tex_tile_cache *tc) +{ + struct pipe_screen *screen; + uint pos; + + for (pos = 0; pos < NUM_ENTRIES; pos++) { + /*assert(tc->entries[pos].x < 0);*/ + } + if (tc->transfer) { + screen = tc->transfer->texture->screen; + screen->tex_transfer_destroy(tc->transfer); + } + if (tc->tex_trans) { + screen = tc->tex_trans->texture->screen; + screen->tex_transfer_destroy(tc->tex_trans); + } + + FREE( tc ); +} + + +void +lp_tex_tile_cache_map_transfers(struct llvmpipe_tex_tile_cache *tc) +{ + if (tc->transfer && !tc->transfer_map) + tc->transfer_map = tc->screen->transfer_map(tc->screen, tc->transfer); + + if (tc->tex_trans && !tc->tex_trans_map) + tc->tex_trans_map = tc->screen->transfer_map(tc->screen, tc->tex_trans); +} + + +void +lp_tex_tile_cache_unmap_transfers(struct llvmpipe_tex_tile_cache *tc) +{ + if (tc->transfer_map) { + tc->screen->transfer_unmap(tc->screen, tc->transfer); + tc->transfer_map = NULL; + } + + if (tc->tex_trans_map) { + tc->screen->transfer_unmap(tc->screen, tc->tex_trans); + tc->tex_trans_map = NULL; + } +} + +void +lp_tex_tile_cache_validate_texture(struct llvmpipe_tex_tile_cache *tc) +{ + if (tc->texture) { + struct llvmpipe_texture *lpt = llvmpipe_texture(tc->texture); + if (lpt->timestamp != tc->timestamp) { + /* texture was modified, invalidate all cached tiles */ + uint i; + _debug_printf("INV %d %d\n", tc->timestamp, lpt->timestamp); + for (i = 0; i < NUM_ENTRIES; i++) { + tc->entries[i].addr.bits.invalid = 1; + } + + tc->timestamp = lpt->timestamp; + } + } +} + +/** + * Specify the texture to cache. + */ +void +lp_tex_tile_cache_set_texture(struct llvmpipe_tex_tile_cache *tc, + struct pipe_texture *texture) +{ + uint i; + + assert(!tc->transfer); + + if (tc->texture != texture) { + pipe_texture_reference(&tc->texture, texture); + + if (tc->tex_trans) { + struct pipe_screen *screen = tc->tex_trans->texture->screen; + + if (tc->tex_trans_map) { + screen->transfer_unmap(screen, tc->tex_trans); + tc->tex_trans_map = NULL; + } + + screen->tex_transfer_destroy(tc->tex_trans); + tc->tex_trans = NULL; + } + + /* mark as entries as invalid/empty */ + /* XXX we should try to avoid this when the teximage hasn't changed */ + for (i = 0; i < NUM_ENTRIES; i++) { + tc->entries[i].addr.bits.invalid = 1; + } + + tc->tex_face = -1; /* any invalid value here */ + } +} + + +/** + * Given the texture face, level, zslice, x and y values, compute + * the cache entry position/index where we'd hope to find the + * cached texture tile. + * This is basically a direct-map cache. + * XXX There's probably lots of ways in which we can improve this. + */ +static INLINE uint +tex_cache_pos( union tex_tile_address addr ) +{ + uint entry = (addr.bits.x + + addr.bits.y * 9 + + addr.bits.z * 3 + + addr.bits.face + + addr.bits.level * 7); + + return entry % NUM_ENTRIES; +} + +/** + * Similar to lp_get_cached_tile() but for textures. + * Tiles are read-only and indexed with more params. + */ +const struct llvmpipe_cached_tex_tile * +lp_find_cached_tex_tile(struct llvmpipe_tex_tile_cache *tc, + union tex_tile_address addr ) +{ + struct pipe_screen *screen = tc->screen; + struct llvmpipe_cached_tex_tile *tile; + + tile = tc->entries + tex_cache_pos( addr ); + + if (addr.value != tile->addr.value) { + + /* cache miss. Most misses are because we've invaldiated the + * texture cache previously -- most commonly on binding a new + * texture. Currently we effectively flush the cache on texture + * bind. + */ +#if 0 + _debug_printf("miss at %u: x=%d y=%d z=%d face=%d level=%d\n" + " tile %u: x=%d y=%d z=%d face=%d level=%d\n", + pos, x/TEX_TILE_SIZE, y/TEX_TILE_SIZE, z, face, level, + pos, tile->addr.bits.x, tile->addr.bits.y, tile->z, tile->face, tile->level); +#endif + + /* check if we need to get a new transfer */ + if (!tc->tex_trans || + tc->tex_face != addr.bits.face || + tc->tex_level != addr.bits.level || + tc->tex_z != addr.bits.z) { + /* get new transfer (view into texture) */ + + if (tc->tex_trans) { + if (tc->tex_trans_map) { + tc->screen->transfer_unmap(tc->screen, tc->tex_trans); + tc->tex_trans_map = NULL; + } + + screen->tex_transfer_destroy(tc->tex_trans); + tc->tex_trans = NULL; + } + + tc->tex_trans = + screen->get_tex_transfer(screen, tc->texture, + addr.bits.face, + addr.bits.level, + addr.bits.z, + PIPE_TRANSFER_READ, 0, 0, + tc->texture->width[addr.bits.level], + tc->texture->height[addr.bits.level]); + + tc->tex_trans_map = screen->transfer_map(screen, tc->tex_trans); + + tc->tex_face = addr.bits.face; + tc->tex_level = addr.bits.level; + tc->tex_z = addr.bits.z; + } + + { + unsigned x = addr.bits.x * TEX_TILE_SIZE; + unsigned y = addr.bits.y * TEX_TILE_SIZE; + unsigned w = TEX_TILE_SIZE; + unsigned h = TEX_TILE_SIZE; + + if (pipe_clip_tile(x, y, &w, &h, tc->tex_trans)) { + assert(0); + } + + util_format_read_4ub(tc->tex_trans->format, + (uint8_t *)tile->color, sizeof tile->color[0], + tc->tex_trans_map, tc->tex_trans->stride, + x, y, w, h); + } + + tile->addr = addr; + } + + tc->last_tile = tile; + return tile; +} diff --git a/src/gallium/drivers/llvmpipe/lp_tex_cache.h b/src/gallium/drivers/llvmpipe/lp_tex_cache.h new file mode 100644 index 00000000000..9fa6c368125 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tex_cache.h @@ -0,0 +1,151 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_TEX_CACHE_H +#define LP_TEX_CACHE_H + + +#include "pipe/p_compiler.h" + + +struct llvmpipe_context; +struct llvmpipe_tex_tile_cache; + + +/** + * Cache tile size (width and height). This needs to be a power of two. + */ +#define TEX_TILE_SIZE 64 + + +/* If we need to support > 4096, just expand this to be a 64 bit + * union, or consider tiling in Z as well. + */ +union tex_tile_address { + struct { + unsigned x:6; /* 4096 / TEX_TILE_SIZE */ + unsigned y:6; /* 4096 / TEX_TILE_SIZE */ + unsigned z:12; /* 4096 -- z not tiled */ + unsigned face:3; + unsigned level:4; + unsigned invalid:1; + } bits; + unsigned value; +}; + + +struct llvmpipe_cached_tex_tile +{ + union tex_tile_address addr; + uint8_t color[TEX_TILE_SIZE][TEX_TILE_SIZE][4]; +}; + +#define NUM_ENTRIES 50 + + +/** XXX move these */ +#define MAX_TEX_WIDTH 2048 +#define MAX_TEX_HEIGHT 2048 + + +struct llvmpipe_tex_tile_cache +{ + struct pipe_screen *screen; + struct pipe_surface *surface; /**< the surface we're caching */ + struct pipe_transfer *transfer; + void *transfer_map; + + struct pipe_texture *texture; /**< if caching a texture */ + unsigned timestamp; + + struct llvmpipe_cached_tex_tile entries[NUM_ENTRIES]; + + struct pipe_transfer *tex_trans; + void *tex_trans_map; + int tex_face, tex_level, tex_z; + + struct llvmpipe_cached_tex_tile *last_tile; /**< most recently retrieved tile */ +}; + + +extern struct llvmpipe_tex_tile_cache * +lp_create_tex_tile_cache( struct pipe_screen *screen ); + +extern void +lp_destroy_tex_tile_cache(struct llvmpipe_tex_tile_cache *tc); + +extern void +lp_tex_tile_cache_map_transfers(struct llvmpipe_tex_tile_cache *tc); + +extern void +lp_tex_tile_cache_unmap_transfers(struct llvmpipe_tex_tile_cache *tc); + +extern void +lp_tex_tile_cache_set_texture(struct llvmpipe_tex_tile_cache *tc, + struct pipe_texture *texture); + +void +lp_tex_tile_cache_validate_texture(struct llvmpipe_tex_tile_cache *tc); + +extern const struct llvmpipe_cached_tex_tile * +lp_find_cached_tex_tile(struct llvmpipe_tex_tile_cache *tc, + union tex_tile_address addr ); + +static INLINE const union tex_tile_address +tex_tile_address( unsigned x, + unsigned y, + unsigned z, + unsigned face, + unsigned level ) +{ + union tex_tile_address addr; + + addr.value = 0; + addr.bits.x = x / TEX_TILE_SIZE; + addr.bits.y = y / TEX_TILE_SIZE; + addr.bits.z = z; + addr.bits.face = face; + addr.bits.level = level; + + return addr; +} + +/* Quickly retrieve tile if it matches last lookup. + */ +static INLINE const struct llvmpipe_cached_tex_tile * +lp_get_cached_tex_tile(struct llvmpipe_tex_tile_cache *tc, + union tex_tile_address addr ) +{ + if (tc->last_tile->addr.value == addr.value) + return tc->last_tile; + + return lp_find_cached_tex_tile( tc, addr ); +} + + +#endif /* LP_TEX_CACHE_H */ + diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.c b/src/gallium/drivers/llvmpipe/lp_tex_sample.c new file mode 100644 index 00000000000..94eb6dad5af --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tex_sample.c @@ -0,0 +1,1580 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * Copyright 2008 VMware, Inc. All rights reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Texture sampling + * + * Authors: + * Brian Paul + */ + +#include "lp_context.h" +#include "lp_quad.h" +#include "lp_surface.h" +#include "lp_texture.h" +#include "lp_tex_sample.h" +#include "lp_tex_cache.h" +#include "pipe/p_context.h" +#include "pipe/p_defines.h" +#include "pipe/p_shader_tokens.h" +#include "util/u_math.h" +#include "util/u_memory.h" + + + +/* + * Note, the FRAC macro has to work perfectly. Otherwise you'll sometimes + * see 1-pixel bands of improperly weighted linear-filtered textures. + * The tests/texwrap.c demo is a good test. + * Also note, FRAC(x) doesn't truly return the fractional part of x for x < 0. + * Instead, if x < 0 then FRAC(x) = 1 - true_frac(x). + */ +#define FRAC(f) ((f) - util_ifloor(f)) + + +/** + * Linear interpolation macro + */ +static INLINE float +lerp(float a, float v0, float v1) +{ + return v0 + a * (v1 - v0); +} + + +/** + * Do 2D/biliner interpolation of float values. + * v00, v10, v01 and v11 are typically four texture samples in a square/box. + * a and b are the horizontal and vertical interpolants. + * It's important that this function is inlined when compiled with + * optimization! If we find that's not true on some systems, convert + * to a macro. + */ +static INLINE float +lerp_2d(float a, float b, + float v00, float v10, float v01, float v11) +{ + const float temp0 = lerp(a, v00, v10); + const float temp1 = lerp(a, v01, v11); + return lerp(b, temp0, temp1); +} + + +/** + * As above, but 3D interpolation of 8 values. + */ +static INLINE float +lerp_3d(float a, float b, float c, + float v000, float v100, float v010, float v110, + float v001, float v101, float v011, float v111) +{ + const float temp0 = lerp_2d(a, b, v000, v100, v010, v110); + const float temp1 = lerp_2d(a, b, v001, v101, v011, v111); + return lerp(c, temp0, temp1); +} + + + +/** + * If A is a signed integer, A % B doesn't give the right value for A < 0 + * (in terms of texture repeat). Just casting to unsigned fixes that. + */ +#define REMAINDER(A, B) ((unsigned) (A) % (unsigned) (B)) + + +/** + * Apply texture coord wrapping mode and return integer texture indexes + * for a vector of four texcoords (S or T or P). + * \param wrapMode PIPE_TEX_WRAP_x + * \param s the incoming texcoords + * \param size the texture image size + * \param icoord returns the integer texcoords + * \return integer texture index + */ +static INLINE void +nearest_texcoord_4(unsigned wrapMode, const float s[4], unsigned size, + int icoord[4]) +{ + uint ch; + switch (wrapMode) { + case PIPE_TEX_WRAP_REPEAT: + /* s limited to [0,1) */ + /* i limited to [0,size-1] */ + for (ch = 0; ch < 4; ch++) { + int i = util_ifloor(s[ch] * size); + icoord[ch] = REMAINDER(i, size); + } + return; + case PIPE_TEX_WRAP_CLAMP: + /* s limited to [0,1] */ + /* i limited to [0,size-1] */ + for (ch = 0; ch < 4; ch++) { + if (s[ch] <= 0.0F) + icoord[ch] = 0; + else if (s[ch] >= 1.0F) + icoord[ch] = size - 1; + else + icoord[ch] = util_ifloor(s[ch] * size); + } + return; + case PIPE_TEX_WRAP_CLAMP_TO_EDGE: + { + /* s limited to [min,max] */ + /* i limited to [0, size-1] */ + const float min = 1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + if (s[ch] < min) + icoord[ch] = 0; + else if (s[ch] > max) + icoord[ch] = size - 1; + else + icoord[ch] = util_ifloor(s[ch] * size); + } + } + return; + case PIPE_TEX_WRAP_CLAMP_TO_BORDER: + { + /* s limited to [min,max] */ + /* i limited to [-1, size] */ + const float min = -1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + if (s[ch] <= min) + icoord[ch] = -1; + else if (s[ch] >= max) + icoord[ch] = size; + else + icoord[ch] = util_ifloor(s[ch] * size); + } + } + return; + case PIPE_TEX_WRAP_MIRROR_REPEAT: + { + const float min = 1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + const int flr = util_ifloor(s[ch]); + float u; + if (flr & 1) + u = 1.0F - (s[ch] - (float) flr); + else + u = s[ch] - (float) flr; + if (u < min) + icoord[ch] = 0; + else if (u > max) + icoord[ch] = size - 1; + else + icoord[ch] = util_ifloor(u * size); + } + } + return; + case PIPE_TEX_WRAP_MIRROR_CLAMP: + for (ch = 0; ch < 4; ch++) { + /* s limited to [0,1] */ + /* i limited to [0,size-1] */ + const float u = fabsf(s[ch]); + if (u <= 0.0F) + icoord[ch] = 0; + else if (u >= 1.0F) + icoord[ch] = size - 1; + else + icoord[ch] = util_ifloor(u * size); + } + return; + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: + { + /* s limited to [min,max] */ + /* i limited to [0, size-1] */ + const float min = 1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + const float u = fabsf(s[ch]); + if (u < min) + icoord[ch] = 0; + else if (u > max) + icoord[ch] = size - 1; + else + icoord[ch] = util_ifloor(u * size); + } + } + return; + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: + { + /* s limited to [min,max] */ + /* i limited to [0, size-1] */ + const float min = -1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + const float u = fabsf(s[ch]); + if (u < min) + icoord[ch] = -1; + else if (u > max) + icoord[ch] = size; + else + icoord[ch] = util_ifloor(u * size); + } + } + return; + default: + assert(0); + } +} + + +/** + * Used to compute texel locations for linear sampling for four texcoords. + * \param wrapMode PIPE_TEX_WRAP_x + * \param s the texcoords + * \param size the texture image size + * \param icoord0 returns first texture indexes + * \param icoord1 returns second texture indexes (usually icoord0 + 1) + * \param w returns blend factor/weight between texture indexes + * \param icoord returns the computed integer texture coords + */ +static INLINE void +linear_texcoord_4(unsigned wrapMode, const float s[4], unsigned size, + int icoord0[4], int icoord1[4], float w[4]) +{ + uint ch; + + switch (wrapMode) { + case PIPE_TEX_WRAP_REPEAT: + for (ch = 0; ch < 4; ch++) { + float u = s[ch] * size - 0.5F; + icoord0[ch] = REMAINDER(util_ifloor(u), size); + icoord1[ch] = REMAINDER(icoord0[ch] + 1, size); + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_CLAMP: + for (ch = 0; ch < 4; ch++) { + float u = CLAMP(s[ch], 0.0F, 1.0F); + u = u * size - 0.5f; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_CLAMP_TO_EDGE: + for (ch = 0; ch < 4; ch++) { + float u = CLAMP(s[ch], 0.0F, 1.0F); + u = u * size - 0.5f; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + if (icoord0[ch] < 0) + icoord0[ch] = 0; + if (icoord1[ch] >= (int) size) + icoord1[ch] = size - 1; + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_CLAMP_TO_BORDER: + { + const float min = -1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + float u = CLAMP(s[ch], min, max); + u = u * size - 0.5f; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + w[ch] = FRAC(u); + } + } + break;; + case PIPE_TEX_WRAP_MIRROR_REPEAT: + for (ch = 0; ch < 4; ch++) { + const int flr = util_ifloor(s[ch]); + float u; + if (flr & 1) + u = 1.0F - (s[ch] - (float) flr); + else + u = s[ch] - (float) flr; + u = u * size - 0.5F; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + if (icoord0[ch] < 0) + icoord0[ch] = 0; + if (icoord1[ch] >= (int) size) + icoord1[ch] = size - 1; + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_MIRROR_CLAMP: + for (ch = 0; ch < 4; ch++) { + float u = fabsf(s[ch]); + if (u >= 1.0F) + u = (float) size; + else + u *= size; + u -= 0.5F; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: + for (ch = 0; ch < 4; ch++) { + float u = fabsf(s[ch]); + if (u >= 1.0F) + u = (float) size; + else + u *= size; + u -= 0.5F; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + if (icoord0[ch] < 0) + icoord0[ch] = 0; + if (icoord1[ch] >= (int) size) + icoord1[ch] = size - 1; + w[ch] = FRAC(u); + } + break;; + case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: + { + const float min = -1.0F / (2.0F * size); + const float max = 1.0F - min; + for (ch = 0; ch < 4; ch++) { + float u = fabsf(s[ch]); + if (u <= min) + u = min * size; + else if (u >= max) + u = max * size; + else + u *= size; + u -= 0.5F; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + w[ch] = FRAC(u); + } + } + break;; + default: + assert(0); + } +} + + +/** + * For RECT textures / unnormalized texcoords + * Only a subset of wrap modes supported. + */ +static INLINE void +nearest_texcoord_unnorm_4(unsigned wrapMode, const float s[4], unsigned size, + int icoord[4]) +{ + uint ch; + switch (wrapMode) { + case PIPE_TEX_WRAP_CLAMP: + for (ch = 0; ch < 4; ch++) { + int i = util_ifloor(s[ch]); + icoord[ch]= CLAMP(i, 0, (int) size-1); + } + return; + case PIPE_TEX_WRAP_CLAMP_TO_EDGE: + /* fall-through */ + case PIPE_TEX_WRAP_CLAMP_TO_BORDER: + for (ch = 0; ch < 4; ch++) { + icoord[ch]= util_ifloor( CLAMP(s[ch], 0.5F, (float) size - 0.5F) ); + } + return; + default: + assert(0); + } +} + + +/** + * For RECT textures / unnormalized texcoords. + * Only a subset of wrap modes supported. + */ +static INLINE void +linear_texcoord_unnorm_4(unsigned wrapMode, const float s[4], unsigned size, + int icoord0[4], int icoord1[4], float w[4]) +{ + uint ch; + switch (wrapMode) { + case PIPE_TEX_WRAP_CLAMP: + for (ch = 0; ch < 4; ch++) { + /* Not exactly what the spec says, but it matches NVIDIA output */ + float u = CLAMP(s[ch] - 0.5F, 0.0f, (float) size - 1.0f); + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + w[ch] = FRAC(u); + } + return; + case PIPE_TEX_WRAP_CLAMP_TO_EDGE: + /* fall-through */ + case PIPE_TEX_WRAP_CLAMP_TO_BORDER: + for (ch = 0; ch < 4; ch++) { + float u = CLAMP(s[ch], 0.5F, (float) size - 0.5F); + u -= 0.5F; + icoord0[ch] = util_ifloor(u); + icoord1[ch] = icoord0[ch] + 1; + if (icoord1[ch] > (int) size - 1) + icoord1[ch] = size - 1; + w[ch] = FRAC(u); + } + break; + default: + assert(0); + } +} + + +static unsigned +choose_cube_face(float rx, float ry, float rz, float *newS, float *newT) +{ + /* + major axis + direction target sc tc ma + ---------- ------------------------------- --- --- --- + +rx TEXTURE_CUBE_MAP_POSITIVE_X_EXT -rz -ry rx + -rx TEXTURE_CUBE_MAP_NEGATIVE_X_EXT +rz -ry rx + +ry TEXTURE_CUBE_MAP_POSITIVE_Y_EXT +rx +rz ry + -ry TEXTURE_CUBE_MAP_NEGATIVE_Y_EXT +rx -rz ry + +rz TEXTURE_CUBE_MAP_POSITIVE_Z_EXT +rx -ry rz + -rz TEXTURE_CUBE_MAP_NEGATIVE_Z_EXT -rx -ry rz + */ + const float arx = fabsf(rx), ary = fabsf(ry), arz = fabsf(rz); + unsigned face; + float sc, tc, ma; + + if (arx > ary && arx > arz) { + if (rx >= 0.0F) { + face = PIPE_TEX_FACE_POS_X; + sc = -rz; + tc = -ry; + ma = arx; + } + else { + face = PIPE_TEX_FACE_NEG_X; + sc = rz; + tc = -ry; + ma = arx; + } + } + else if (ary > arx && ary > arz) { + if (ry >= 0.0F) { + face = PIPE_TEX_FACE_POS_Y; + sc = rx; + tc = rz; + ma = ary; + } + else { + face = PIPE_TEX_FACE_NEG_Y; + sc = rx; + tc = -rz; + ma = ary; + } + } + else { + if (rz > 0.0F) { + face = PIPE_TEX_FACE_POS_Z; + sc = rx; + tc = -ry; + ma = arz; + } + else { + face = PIPE_TEX_FACE_NEG_Z; + sc = -rx; + tc = -ry; + ma = arz; + } + } + + *newS = ( sc / ma + 1.0F ) * 0.5F; + *newT = ( tc / ma + 1.0F ) * 0.5F; + + return face; +} + + +/** + * Examine the quad's texture coordinates to compute the partial + * derivatives w.r.t X and Y, then compute lambda (level of detail). + * + * This is only done for fragment shaders, not vertex shaders. + */ +static float +compute_lambda(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + float rho, lambda; + + if (samp->processor == TGSI_PROCESSOR_VERTEX) + return lodbias; + + assert(sampler->normalized_coords); + + assert(s); + { + float dsdx = s[QUAD_BOTTOM_RIGHT] - s[QUAD_BOTTOM_LEFT]; + float dsdy = s[QUAD_TOP_LEFT] - s[QUAD_BOTTOM_LEFT]; + dsdx = fabsf(dsdx); + dsdy = fabsf(dsdy); + rho = MAX2(dsdx, dsdy) * texture->width[0]; + } + if (t) { + float dtdx = t[QUAD_BOTTOM_RIGHT] - t[QUAD_BOTTOM_LEFT]; + float dtdy = t[QUAD_TOP_LEFT] - t[QUAD_BOTTOM_LEFT]; + float max; + dtdx = fabsf(dtdx); + dtdy = fabsf(dtdy); + max = MAX2(dtdx, dtdy) * texture->height[0]; + rho = MAX2(rho, max); + } + if (p) { + float dpdx = p[QUAD_BOTTOM_RIGHT] - p[QUAD_BOTTOM_LEFT]; + float dpdy = p[QUAD_TOP_LEFT] - p[QUAD_BOTTOM_LEFT]; + float max; + dpdx = fabsf(dpdx); + dpdy = fabsf(dpdy); + max = MAX2(dpdx, dpdy) * texture->depth[0]; + rho = MAX2(rho, max); + } + + lambda = util_fast_log2(rho); + lambda += lodbias + sampler->lod_bias; + lambda = CLAMP(lambda, sampler->min_lod, sampler->max_lod); + + return lambda; +} + + +/** + * Do several things here: + * 1. Compute lambda from the texcoords, if needed + * 2. Determine if we're minifying or magnifying + * 3. If minifying, choose mipmap levels + * 4. Return image filter to use within mipmap images + * \param level0 Returns first mipmap level to sample from + * \param level1 Returns second mipmap level to sample from + * \param levelBlend Returns blend factor between levels, in [0,1] + * \param imgFilter Returns either the min or mag filter, depending on lambda + */ +static void +choose_mipmap_levels(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + unsigned *level0, unsigned *level1, float *levelBlend, + unsigned *imgFilter) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + + if (sampler->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) { + /* no mipmap selection needed */ + *level0 = *level1 = CLAMP((int) sampler->min_lod, + 0, (int) texture->last_level); + + if (sampler->min_img_filter != sampler->mag_img_filter) { + /* non-mipmapped texture, but still need to determine if doing + * minification or magnification. + */ + float lambda = compute_lambda(tgsi_sampler, s, t, p, lodbias); + if (lambda <= 0.0) { + *imgFilter = sampler->mag_img_filter; + } + else { + *imgFilter = sampler->min_img_filter; + } + } + else { + *imgFilter = sampler->mag_img_filter; + } + } + else { + float lambda = compute_lambda(tgsi_sampler, s, t, p, lodbias); + + if (lambda <= 0.0) { /* XXX threshold depends on the filter */ + /* magnifying */ + *imgFilter = sampler->mag_img_filter; + *level0 = *level1 = 0; + } + else { + /* minifying */ + *imgFilter = sampler->min_img_filter; + + /* choose mipmap level(s) and compute the blend factor between them */ + if (sampler->min_mip_filter == PIPE_TEX_MIPFILTER_NEAREST) { + /* Nearest mipmap level */ + const int lvl = (int) (lambda + 0.5); + *level0 = + *level1 = CLAMP(lvl, 0, (int) texture->last_level); + } + else { + /* Linear interpolation between mipmap levels */ + const int lvl = (int) lambda; + *level0 = CLAMP(lvl, 0, (int) texture->last_level); + *level1 = CLAMP(lvl + 1, 0, (int) texture->last_level); + *levelBlend = FRAC(lambda); /* blending weight between levels */ + } + } + } +} + + +/** + * Get a texel from a texture, using the texture tile cache. + * + * \param face the cube face in 0..5 + * \param level the mipmap level + * \param x the x coord of texel within 2D image + * \param y the y coord of texel within 2D image + * \param z which slice of a 3D texture + * \param rgba the quad to put the texel/color into + * \param j which element of the rgba quad to write to + * + * XXX maybe move this into lp_tile_cache.c and merge with the + * lp_get_cached_tile_tex() function. Also, get 4 texels instead of 1... + */ +static void +get_texel_quad_2d(const struct tgsi_sampler *tgsi_sampler, + unsigned face, unsigned level, int x, int y, + const uint8_t *out[4]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + + const struct llvmpipe_cached_tex_tile *tile + = lp_get_cached_tex_tile(samp->cache, + tex_tile_address(x, y, 0, face, level)); + + y %= TEX_TILE_SIZE; + x %= TEX_TILE_SIZE; + + out[0] = &tile->color[y ][x ][0]; + out[1] = &tile->color[y ][x+1][0]; + out[2] = &tile->color[y+1][x ][0]; + out[3] = &tile->color[y+1][x+1][0]; +} + +static INLINE const uint8_t * +get_texel_2d_ptr(const struct tgsi_sampler *tgsi_sampler, + unsigned face, unsigned level, int x, int y) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + + const struct llvmpipe_cached_tex_tile *tile + = lp_get_cached_tex_tile(samp->cache, + tex_tile_address(x, y, 0, face, level)); + + y %= TEX_TILE_SIZE; + x %= TEX_TILE_SIZE; + + return &tile->color[y][x][0]; +} + + +static void +get_texel_quad_2d_mt(const struct tgsi_sampler *tgsi_sampler, + unsigned face, unsigned level, + int x0, int y0, + int x1, int y1, + const uint8_t *out[4]) +{ + unsigned i; + + for (i = 0; i < 4; i++) { + unsigned tx = (i & 1) ? x1 : x0; + unsigned ty = (i >> 1) ? y1 : y0; + + out[i] = get_texel_2d_ptr( tgsi_sampler, face, level, tx, ty ); + } +} + +static void +get_texel(const struct tgsi_sampler *tgsi_sampler, + unsigned face, unsigned level, int x, int y, int z, + float rgba[NUM_CHANNELS][QUAD_SIZE], unsigned j) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + + if (x < 0 || x >= (int) texture->width[level] || + y < 0 || y >= (int) texture->height[level] || + z < 0 || z >= (int) texture->depth[level]) { + rgba[0][j] = sampler->border_color[0]; + rgba[1][j] = sampler->border_color[1]; + rgba[2][j] = sampler->border_color[2]; + rgba[3][j] = sampler->border_color[3]; + } + else { + const unsigned tx = x % TEX_TILE_SIZE; + const unsigned ty = y % TEX_TILE_SIZE; + const struct llvmpipe_cached_tex_tile *tile; + + tile = lp_get_cached_tex_tile(samp->cache, + tex_tile_address(x, y, z, face, level)); + + rgba[0][j] = ubyte_to_float(tile->color[ty][tx][0]); + rgba[1][j] = ubyte_to_float(tile->color[ty][tx][1]); + rgba[2][j] = ubyte_to_float(tile->color[ty][tx][2]); + rgba[3][j] = ubyte_to_float(tile->color[ty][tx][3]); + if (0) + { + debug_printf("Get texel %f %f %f %f from %s\n", + rgba[0][j], rgba[1][j], rgba[2][j], rgba[3][j], + pf_name(texture->format)); + } + } +} + + +/** + * Compare texcoord 'p' (aka R) against texture value 'rgba[0]' + * When we sampled the depth texture, the depth value was put into all + * RGBA channels. We look at the red channel here. + * \param rgba quad of (depth) texel values + * \param p texture 'P' components for four pixels in quad + * \param j which pixel in the quad to test [0..3] + */ +static INLINE void +shadow_compare(const struct pipe_sampler_state *sampler, + float rgba[NUM_CHANNELS][QUAD_SIZE], + const float p[QUAD_SIZE], + uint j) +{ + int k; + switch (sampler->compare_func) { + case PIPE_FUNC_LESS: + k = p[j] < rgba[0][j]; + break; + case PIPE_FUNC_LEQUAL: + k = p[j] <= rgba[0][j]; + break; + case PIPE_FUNC_GREATER: + k = p[j] > rgba[0][j]; + break; + case PIPE_FUNC_GEQUAL: + k = p[j] >= rgba[0][j]; + break; + case PIPE_FUNC_EQUAL: + k = p[j] == rgba[0][j]; + break; + case PIPE_FUNC_NOTEQUAL: + k = p[j] != rgba[0][j]; + break; + case PIPE_FUNC_ALWAYS: + k = 1; + break; + case PIPE_FUNC_NEVER: + k = 0; + break; + default: + k = 0; + assert(0); + break; + } + + /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */ + rgba[0][j] = rgba[1][j] = rgba[2][j] = (float) k; + rgba[3][j] = 1.0F; +} + + +/** + * As above, but do four z/texture comparisons. + */ +static INLINE void +shadow_compare4(const struct pipe_sampler_state *sampler, + float rgba[NUM_CHANNELS][QUAD_SIZE], + const float p[QUAD_SIZE]) +{ + int j, k0, k1, k2, k3; + float val; + + /* compare four texcoords vs. four texture samples */ + switch (sampler->compare_func) { + case PIPE_FUNC_LESS: + k0 = p[0] < rgba[0][0]; + k1 = p[1] < rgba[0][1]; + k2 = p[2] < rgba[0][2]; + k3 = p[3] < rgba[0][3]; + break; + case PIPE_FUNC_LEQUAL: + k0 = p[0] <= rgba[0][0]; + k1 = p[1] <= rgba[0][1]; + k2 = p[2] <= rgba[0][2]; + k3 = p[3] <= rgba[0][3]; + break; + case PIPE_FUNC_GREATER: + k0 = p[0] > rgba[0][0]; + k1 = p[1] > rgba[0][1]; + k2 = p[2] > rgba[0][2]; + k3 = p[3] > rgba[0][3]; + break; + case PIPE_FUNC_GEQUAL: + k0 = p[0] >= rgba[0][0]; + k1 = p[1] >= rgba[0][1]; + k2 = p[2] >= rgba[0][2]; + k3 = p[3] >= rgba[0][3]; + break; + case PIPE_FUNC_EQUAL: + k0 = p[0] == rgba[0][0]; + k1 = p[1] == rgba[0][1]; + k2 = p[2] == rgba[0][2]; + k3 = p[3] == rgba[0][3]; + break; + case PIPE_FUNC_NOTEQUAL: + k0 = p[0] != rgba[0][0]; + k1 = p[1] != rgba[0][1]; + k2 = p[2] != rgba[0][2]; + k3 = p[3] != rgba[0][3]; + break; + case PIPE_FUNC_ALWAYS: + k0 = k1 = k2 = k3 = 1; + break; + case PIPE_FUNC_NEVER: + k0 = k1 = k2 = k3 = 0; + break; + default: + k0 = k1 = k2 = k3 = 0; + assert(0); + break; + } + + /* convert four pass/fail values to an intensity in [0,1] */ + val = 0.25F * (k0 + k1 + k2 + k3); + + /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */ + for (j = 0; j < 4; j++) { + rgba[0][j] = rgba[1][j] = rgba[2][j] = val; + rgba[3][j] = 1.0F; + } +} + + + +static void +lp_get_samples_2d_linear_repeat_POT(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + unsigned j; + unsigned level = samp->level; + unsigned xpot = 1 << (samp->xpot - level); + unsigned ypot = 1 << (samp->ypot - level); + unsigned xmax = (xpot - 1) & (TEX_TILE_SIZE - 1); /* MIN2(TEX_TILE_SIZE, xpot) - 1; */ + unsigned ymax = (ypot - 1) & (TEX_TILE_SIZE - 1); /* MIN2(TEX_TILE_SIZE, ypot) - 1; */ + + for (j = 0; j < QUAD_SIZE; j++) { + int c; + + float u = s[j] * xpot - 0.5F; + float v = t[j] * ypot - 0.5F; + + int uflr = util_ifloor(u); + int vflr = util_ifloor(v); + + float xw = u - (float)uflr; + float yw = v - (float)vflr; + + int x0 = uflr & (xpot - 1); + int y0 = vflr & (ypot - 1); + + const uint8_t *tx[4]; + + + /* Can we fetch all four at once: + */ + if (x0 < xmax && y0 < ymax) + { + get_texel_quad_2d(tgsi_sampler, 0, level, x0, y0, tx); + } + else + { + unsigned x1 = (x0 + 1) & (xpot - 1); + unsigned y1 = (y0 + 1) & (ypot - 1); + get_texel_quad_2d_mt(tgsi_sampler, 0, level, + x0, y0, x1, y1, tx); + } + + + /* interpolate R, G, B, A */ + for (c = 0; c < 4; c++) { + rgba[c][j] = lerp_2d(xw, yw, + ubyte_to_float(tx[0][c]), ubyte_to_float(tx[1][c]), + ubyte_to_float(tx[2][c]), ubyte_to_float(tx[3][c])); + } + } +} + + +static void +lp_get_samples_2d_nearest_repeat_POT(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + unsigned j; + unsigned level = samp->level; + unsigned xpot = 1 << (samp->xpot - level); + unsigned ypot = 1 << (samp->ypot - level); + + for (j = 0; j < QUAD_SIZE; j++) { + int c; + + float u = s[j] * xpot; + float v = t[j] * ypot; + + int uflr = util_ifloor(u); + int vflr = util_ifloor(v); + + int x0 = uflr & (xpot - 1); + int y0 = vflr & (ypot - 1); + + const uint8_t *out = get_texel_2d_ptr(tgsi_sampler, 0, level, x0, y0); + + for (c = 0; c < 4; c++) { + rgba[c][j] = ubyte_to_float(out[c]); + } + } +} + + +static void +lp_get_samples_2d_nearest_clamp_POT(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + unsigned j; + unsigned level = samp->level; + unsigned xpot = 1 << (samp->xpot - level); + unsigned ypot = 1 << (samp->ypot - level); + + for (j = 0; j < QUAD_SIZE; j++) { + int c; + + float u = s[j] * xpot; + float v = t[j] * ypot; + + int x0, y0; + const uint8_t *out; + + x0 = util_ifloor(u); + if (x0 < 0) + x0 = 0; + else if (x0 > xpot - 1) + x0 = xpot - 1; + + y0 = util_ifloor(v); + if (y0 < 0) + y0 = 0; + else if (y0 > ypot - 1) + y0 = ypot - 1; + + out = get_texel_2d_ptr(tgsi_sampler, 0, level, x0, y0); + + for (c = 0; c < 4; c++) { + rgba[c][j] = ubyte_to_float(out[c]); + } + } +} + + +static void +lp_get_samples_2d_linear_mip_linear_repeat_POT(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + int level0; + float lambda; + + lambda = compute_lambda(tgsi_sampler, s, t, p, lodbias); + level0 = (int)lambda; + + if (lambda < 0.0) { + samp->level = 0; + lp_get_samples_2d_linear_repeat_POT( tgsi_sampler, + s, t, p, 0, rgba ); + } + else if (level0 >= texture->last_level) { + samp->level = texture->last_level; + lp_get_samples_2d_linear_repeat_POT( tgsi_sampler, + s, t, p, 0, rgba ); + } + else { + float levelBlend = lambda - level0; + float rgba0[4][4]; + float rgba1[4][4]; + int c,j; + + samp->level = level0; + lp_get_samples_2d_linear_repeat_POT( tgsi_sampler, + s, t, p, 0, rgba0 ); + + samp->level = level0+1; + lp_get_samples_2d_linear_repeat_POT( tgsi_sampler, + s, t, p, 0, rgba1 ); + + for (j = 0; j < QUAD_SIZE; j++) { + for (c = 0; c < 4; c++) { + rgba[c][j] = lerp(levelBlend, rgba0[c][j], rgba1[c][j]); + } + } + } +} + +/** + * Common code for sampling 1D/2D/cube textures. + * Could probably extend for 3D... + */ +static void +lp_get_samples_2d_common(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE], + const unsigned faces[4]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + unsigned level0, level1, j, imgFilter; + int width, height; + float levelBlend; + + choose_mipmap_levels(tgsi_sampler, s, t, p, + lodbias, + &level0, &level1, &levelBlend, &imgFilter); + + assert(sampler->normalized_coords); + + width = texture->width[level0]; + height = texture->height[level0]; + + assert(width > 0); + + switch (imgFilter) { + case PIPE_TEX_FILTER_NEAREST: + { + int x[4], y[4]; + nearest_texcoord_4(sampler->wrap_s, s, width, x); + nearest_texcoord_4(sampler->wrap_t, t, height, y); + + for (j = 0; j < QUAD_SIZE; j++) { + get_texel(tgsi_sampler, faces[j], level0, x[j], y[j], 0, rgba, j); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) { + shadow_compare(sampler, rgba, p, j); + } + + if (level0 != level1) { + /* get texels from second mipmap level and blend */ + float rgba2[4][4]; + unsigned c; + x[j] /= 2; + y[j] /= 2; + get_texel(tgsi_sampler, faces[j], level1, x[j], y[j], 0, + rgba2, j); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE){ + shadow_compare(sampler, rgba2, p, j); + } + + for (c = 0; c < NUM_CHANNELS; c++) { + rgba[c][j] = lerp(levelBlend, rgba[c][j], rgba2[c][j]); + } + } + } + } + break; + case PIPE_TEX_FILTER_LINEAR: + case PIPE_TEX_FILTER_ANISO: + { + int x0[4], y0[4], x1[4], y1[4]; + float xw[4], yw[4]; /* weights */ + + linear_texcoord_4(sampler->wrap_s, s, width, x0, x1, xw); + linear_texcoord_4(sampler->wrap_t, t, height, y0, y1, yw); + + for (j = 0; j < QUAD_SIZE; j++) { + float tx[4][4]; /* texels */ + int c; + get_texel(tgsi_sampler, faces[j], level0, x0[j], y0[j], 0, tx, 0); + get_texel(tgsi_sampler, faces[j], level0, x1[j], y0[j], 0, tx, 1); + get_texel(tgsi_sampler, faces[j], level0, x0[j], y1[j], 0, tx, 2); + get_texel(tgsi_sampler, faces[j], level0, x1[j], y1[j], 0, tx, 3); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) { + shadow_compare4(sampler, tx, p); + } + + /* interpolate R, G, B, A */ + for (c = 0; c < 4; c++) { + rgba[c][j] = lerp_2d(xw[j], yw[j], + tx[c][0], tx[c][1], + tx[c][2], tx[c][3]); + } + + if (level0 != level1) { + /* get texels from second mipmap level and blend */ + float rgba2[4][4]; + + /* XXX: This is incorrect -- will often end up with (x0 + * == x1 && y0 == y1), meaning that we fetch the same + * texel four times and linearly interpolate between + * identical values. The correct approach would be to + * call linear_texcoord again for the second level. + */ + x0[j] /= 2; + y0[j] /= 2; + x1[j] /= 2; + y1[j] /= 2; + get_texel(tgsi_sampler, faces[j], level1, x0[j], y0[j], 0, tx, 0); + get_texel(tgsi_sampler, faces[j], level1, x1[j], y0[j], 0, tx, 1); + get_texel(tgsi_sampler, faces[j], level1, x0[j], y1[j], 0, tx, 2); + get_texel(tgsi_sampler, faces[j], level1, x1[j], y1[j], 0, tx, 3); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE){ + shadow_compare4(sampler, tx, p); + } + + /* interpolate R, G, B, A */ + for (c = 0; c < 4; c++) { + rgba2[c][j] = lerp_2d(xw[j], yw[j], + tx[c][0], tx[c][1], tx[c][2], tx[c][3]); + } + + for (c = 0; c < NUM_CHANNELS; c++) { + rgba[c][j] = lerp(levelBlend, rgba[c][j], rgba2[c][j]); + } + } + } + } + break; + default: + assert(0); + } +} + + +static INLINE void +lp_get_samples_1d(struct tgsi_sampler *sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + static const unsigned faces[4] = {0, 0, 0, 0}; + static const float tzero[4] = {0, 0, 0, 0}; + lp_get_samples_2d_common(sampler, s, tzero, NULL, + lodbias, rgba, faces); +} + + +static INLINE void +lp_get_samples_2d(struct tgsi_sampler *sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + static const unsigned faces[4] = {0, 0, 0, 0}; + lp_get_samples_2d_common(sampler, s, t, p, + lodbias, rgba, faces); +} + + +static INLINE void +lp_get_samples_3d(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + /* get/map pipe_surfaces corresponding to 3D tex slices */ + unsigned level0, level1, j, imgFilter; + int width, height, depth; + float levelBlend; + const uint face = 0; + + choose_mipmap_levels(tgsi_sampler, s, t, p, + lodbias, + &level0, &level1, &levelBlend, &imgFilter); + + assert(sampler->normalized_coords); + + width = texture->width[level0]; + height = texture->height[level0]; + depth = texture->depth[level0]; + + assert(width > 0); + assert(height > 0); + assert(depth > 0); + + switch (imgFilter) { + case PIPE_TEX_FILTER_NEAREST: + { + int x[4], y[4], z[4]; + nearest_texcoord_4(sampler->wrap_s, s, width, x); + nearest_texcoord_4(sampler->wrap_t, t, height, y); + nearest_texcoord_4(sampler->wrap_r, p, depth, z); + for (j = 0; j < QUAD_SIZE; j++) { + get_texel(tgsi_sampler, face, level0, x[j], y[j], z[j], rgba, j); + if (level0 != level1) { + /* get texels from second mipmap level and blend */ + float rgba2[4][4]; + unsigned c; + x[j] /= 2; + y[j] /= 2; + z[j] /= 2; + get_texel(tgsi_sampler, face, level1, x[j], y[j], z[j], rgba2, j); + for (c = 0; c < NUM_CHANNELS; c++) { + rgba[c][j] = lerp(levelBlend, rgba2[c][j], rgba[c][j]); + } + } + } + } + break; + case PIPE_TEX_FILTER_LINEAR: + case PIPE_TEX_FILTER_ANISO: + { + int x0[4], x1[4], y0[4], y1[4], z0[4], z1[4]; + float xw[4], yw[4], zw[4]; /* interpolation weights */ + linear_texcoord_4(sampler->wrap_s, s, width, x0, x1, xw); + linear_texcoord_4(sampler->wrap_t, t, height, y0, y1, yw); + linear_texcoord_4(sampler->wrap_r, p, depth, z0, z1, zw); + + for (j = 0; j < QUAD_SIZE; j++) { + int c; + float tx0[4][4], tx1[4][4]; + get_texel(tgsi_sampler, face, level0, x0[j], y0[j], z0[j], tx0, 0); + get_texel(tgsi_sampler, face, level0, x1[j], y0[j], z0[j], tx0, 1); + get_texel(tgsi_sampler, face, level0, x0[j], y1[j], z0[j], tx0, 2); + get_texel(tgsi_sampler, face, level0, x1[j], y1[j], z0[j], tx0, 3); + get_texel(tgsi_sampler, face, level0, x0[j], y0[j], z1[j], tx1, 0); + get_texel(tgsi_sampler, face, level0, x1[j], y0[j], z1[j], tx1, 1); + get_texel(tgsi_sampler, face, level0, x0[j], y1[j], z1[j], tx1, 2); + get_texel(tgsi_sampler, face, level0, x1[j], y1[j], z1[j], tx1, 3); + + /* interpolate R, G, B, A */ + for (c = 0; c < 4; c++) { + rgba[c][j] = lerp_3d(xw[j], yw[j], zw[j], + tx0[c][0], tx0[c][1], + tx0[c][2], tx0[c][3], + tx1[c][0], tx1[c][1], + tx1[c][2], tx1[c][3]); + } + + if (level0 != level1) { + /* get texels from second mipmap level and blend */ + float rgba2[4][4]; + x0[j] /= 2; + y0[j] /= 2; + z0[j] /= 2; + x1[j] /= 2; + y1[j] /= 2; + z1[j] /= 2; + get_texel(tgsi_sampler, face, level1, x0[j], y0[j], z0[j], tx0, 0); + get_texel(tgsi_sampler, face, level1, x1[j], y0[j], z0[j], tx0, 1); + get_texel(tgsi_sampler, face, level1, x0[j], y1[j], z0[j], tx0, 2); + get_texel(tgsi_sampler, face, level1, x1[j], y1[j], z0[j], tx0, 3); + get_texel(tgsi_sampler, face, level1, x0[j], y0[j], z1[j], tx1, 0); + get_texel(tgsi_sampler, face, level1, x1[j], y0[j], z1[j], tx1, 1); + get_texel(tgsi_sampler, face, level1, x0[j], y1[j], z1[j], tx1, 2); + get_texel(tgsi_sampler, face, level1, x1[j], y1[j], z1[j], tx1, 3); + + /* interpolate R, G, B, A */ + for (c = 0; c < 4; c++) { + rgba2[c][j] = lerp_3d(xw[j], yw[j], zw[j], + tx0[c][0], tx0[c][1], + tx0[c][2], tx0[c][3], + tx1[c][0], tx1[c][1], + tx1[c][2], tx1[c][3]); + } + + /* blend mipmap levels */ + for (c = 0; c < NUM_CHANNELS; c++) { + rgba[c][j] = lerp(levelBlend, rgba[c][j], rgba2[c][j]); + } + } + } + } + break; + default: + assert(0); + } +} + + +static void +lp_get_samples_cube(struct tgsi_sampler *sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + unsigned faces[QUAD_SIZE], j; + float ssss[4], tttt[4]; + for (j = 0; j < QUAD_SIZE; j++) { + faces[j] = choose_cube_face(s[j], t[j], p[j], ssss + j, tttt + j); + } + lp_get_samples_2d_common(sampler, ssss, tttt, NULL, + lodbias, rgba, faces); +} + + +static void +lp_get_samples_rect(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + const struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + const uint face = 0; + unsigned level0, level1, j, imgFilter; + int width, height; + float levelBlend; + + choose_mipmap_levels(tgsi_sampler, s, t, p, + lodbias, + &level0, &level1, &levelBlend, &imgFilter); + + /* texture RECTS cannot be mipmapped */ + assert(level0 == level1); + + width = texture->width[level0]; + height = texture->height[level0]; + + assert(width > 0); + + switch (imgFilter) { + case PIPE_TEX_FILTER_NEAREST: + { + int x[4], y[4]; + nearest_texcoord_unnorm_4(sampler->wrap_s, s, width, x); + nearest_texcoord_unnorm_4(sampler->wrap_t, t, height, y); + for (j = 0; j < QUAD_SIZE; j++) { + get_texel(tgsi_sampler, face, level0, x[j], y[j], 0, rgba, j); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) { + shadow_compare(sampler, rgba, p, j); + } + } + } + break; + case PIPE_TEX_FILTER_LINEAR: + case PIPE_TEX_FILTER_ANISO: + { + int x0[4], y0[4], x1[4], y1[4]; + float xw[4], yw[4]; /* weights */ + linear_texcoord_unnorm_4(sampler->wrap_s, s, width, x0, x1, xw); + linear_texcoord_unnorm_4(sampler->wrap_t, t, height, y0, y1, yw); + for (j = 0; j < QUAD_SIZE; j++) { + float tx[4][4]; /* texels */ + int c; + get_texel(tgsi_sampler, face, level0, x0[j], y0[j], 0, tx, 0); + get_texel(tgsi_sampler, face, level0, x1[j], y0[j], 0, tx, 1); + get_texel(tgsi_sampler, face, level0, x0[j], y1[j], 0, tx, 2); + get_texel(tgsi_sampler, face, level0, x1[j], y1[j], 0, tx, 3); + if (sampler->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) { + shadow_compare4(sampler, tx, p); + } + for (c = 0; c < 4; c++) { + rgba[c][j] = lerp_2d(xw[j], yw[j], + tx[c][0], tx[c][1], tx[c][2], tx[c][3]); + } + } + } + break; + default: + assert(0); + } +} + + +/** + * Error condition handler + */ +static INLINE void +lp_get_samples_null(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + int i,j; + + for (i = 0; i < 4; i++) + for (j = 0; j < 4; j++) + rgba[i][j] = 1.0; +} + +/** + * Called via tgsi_sampler::get_samples() when using a sampler for the + * first time. Determine the actual sampler function, link it in and + * call it. + */ +void +lp_get_samples(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]) +{ + struct lp_shader_sampler *samp = lp_shader_sampler(tgsi_sampler); + const struct pipe_texture *texture = samp->texture; + const struct pipe_sampler_state *sampler = samp->sampler; + + /* Default to the 'undefined' case: + */ + tgsi_sampler->get_samples = lp_get_samples_null; + + if (!texture) { + assert(0); /* is this legal?? */ + goto out; + } + + if (!sampler->normalized_coords) { + assert (texture->target == PIPE_TEXTURE_2D); + tgsi_sampler->get_samples = lp_get_samples_rect; + goto out; + } + + switch (texture->target) { + case PIPE_TEXTURE_1D: + tgsi_sampler->get_samples = lp_get_samples_1d; + break; + case PIPE_TEXTURE_2D: + tgsi_sampler->get_samples = lp_get_samples_2d; + break; + case PIPE_TEXTURE_3D: + tgsi_sampler->get_samples = lp_get_samples_3d; + break; + case PIPE_TEXTURE_CUBE: + tgsi_sampler->get_samples = lp_get_samples_cube; + break; + default: + assert(0); + break; + } + + /* Do this elsewhere: + */ + samp->xpot = util_unsigned_logbase2( samp->texture->width[0] ); + samp->ypot = util_unsigned_logbase2( samp->texture->height[0] ); + + /* Try to hook in a faster sampler. Ultimately we'll have to + * code-generate these. Luckily most of this looks like it is + * orthogonal state within the sampler. + */ + if (texture->target == PIPE_TEXTURE_2D && + sampler->min_img_filter == sampler->mag_img_filter && + sampler->wrap_s == sampler->wrap_t && + sampler->compare_mode == FALSE && + sampler->normalized_coords) + { + if (sampler->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) { + samp->level = CLAMP((int) sampler->min_lod, + 0, (int) texture->last_level); + + if (sampler->wrap_s == PIPE_TEX_WRAP_REPEAT) { + switch (sampler->min_img_filter) { + case PIPE_TEX_FILTER_NEAREST: + tgsi_sampler->get_samples = lp_get_samples_2d_nearest_repeat_POT; + break; + case PIPE_TEX_FILTER_LINEAR: + tgsi_sampler->get_samples = lp_get_samples_2d_linear_repeat_POT; + break; + default: + break; + } + } + else if (sampler->wrap_s == PIPE_TEX_WRAP_CLAMP) { + switch (sampler->min_img_filter) { + case PIPE_TEX_FILTER_NEAREST: + tgsi_sampler->get_samples = lp_get_samples_2d_nearest_clamp_POT; + break; + default: + break; + } + } + } + else if (sampler->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + if (sampler->wrap_s == PIPE_TEX_WRAP_REPEAT) { + switch (sampler->min_img_filter) { + case PIPE_TEX_FILTER_LINEAR: + tgsi_sampler->get_samples = lp_get_samples_2d_linear_mip_linear_repeat_POT; + break; + default: + break; + } + } + } + } + else if (0) { + _debug_printf("target %d/%d min_mip %d/%d min_img %d/%d wrap %d/%d compare %d/%d norm %d/%d\n", + texture->target, PIPE_TEXTURE_2D, + sampler->min_mip_filter, PIPE_TEX_MIPFILTER_NONE, + sampler->min_img_filter, sampler->mag_img_filter, + sampler->wrap_s, sampler->wrap_t, + sampler->compare_mode, FALSE, + sampler->normalized_coords, TRUE); + } + +out: + tgsi_sampler->get_samples( tgsi_sampler, s, t, p, lodbias, rgba ); +} + diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.h b/src/gallium/drivers/llvmpipe/lp_tex_sample.h new file mode 100644 index 00000000000..628ec3f1efd --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tex_sample.h @@ -0,0 +1,78 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_TEX_SAMPLE_H +#define LP_TEX_SAMPLE_H + + +#include "tgsi/tgsi_exec.h" + + +struct llvmpipe_tex_tile_cache; + + +/** + * Subclass of tgsi_sampler + */ +struct lp_shader_sampler +{ + struct tgsi_sampler base; /**< base class */ + + unsigned processor; + + /* For lp_get_samples_2d_linear_POT: + */ + unsigned xpot; + unsigned ypot; + unsigned level; + + const struct pipe_texture *texture; + const struct pipe_sampler_state *sampler; + + struct llvmpipe_tex_tile_cache *cache; +}; + + + +static INLINE struct lp_shader_sampler * +lp_shader_sampler(const struct tgsi_sampler *sampler) +{ + return (struct lp_shader_sampler *) sampler; +} + + + +extern void +lp_get_samples(struct tgsi_sampler *tgsi_sampler, + const float s[QUAD_SIZE], + const float t[QUAD_SIZE], + const float p[QUAD_SIZE], + float lodbias, + float rgba[NUM_CHANNELS][QUAD_SIZE]); + + +#endif /* LP_TEX_SAMPLE_H */ diff --git a/src/gallium/drivers/llvmpipe/lp_texture.c b/src/gallium/drivers/llvmpipe/lp_texture.c new file mode 100644 index 00000000000..724d4378336 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_texture.c @@ -0,0 +1,429 @@ +/************************************************************************** + * + * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + /* + * Authors: + * Keith Whitwell <keith@tungstengraphics.com> + * Michel Dänzer <michel@tungstengraphics.com> + */ + +#include "pipe/p_context.h" +#include "pipe/p_defines.h" +#include "pipe/p_inlines.h" +#include "pipe/internal/p_winsys_screen.h" +#include "util/u_math.h" +#include "util/u_memory.h" + +#include "lp_context.h" +#include "lp_state.h" +#include "lp_texture.h" +#include "lp_tex_cache.h" +#include "lp_screen.h" +#include "lp_winsys.h" + + +/* Simple, maximally packed layout. + */ + + +/* Conventional allocation path for non-display textures: + */ +static boolean +llvmpipe_texture_layout(struct llvmpipe_screen *screen, + struct llvmpipe_texture * lpt) +{ + struct pipe_texture *pt = &lpt->base; + unsigned level; + unsigned width = pt->width[0]; + unsigned height = pt->height[0]; + unsigned depth = pt->depth[0]; + + unsigned buffer_size = 0; + + pf_get_block(lpt->base.format, &lpt->base.block); + + for (level = 0; level <= pt->last_level; level++) { + pt->width[level] = width; + pt->height[level] = height; + pt->depth[level] = depth; + pt->nblocksx[level] = pf_get_nblocksx(&pt->block, width); + pt->nblocksy[level] = pf_get_nblocksy(&pt->block, height); + lpt->stride[level] = align(pt->nblocksx[level]*pt->block.size, 16); + + lpt->level_offset[level] = buffer_size; + + buffer_size += (pt->nblocksy[level] * + ((pt->target == PIPE_TEXTURE_CUBE) ? 6 : depth) * + lpt->stride[level]); + + width = minify(width); + height = minify(height); + depth = minify(depth); + } + + lpt->data = align_malloc(buffer_size, 16); + + return lpt->data != NULL; +} + +static boolean +llvmpipe_displaytarget_layout(struct llvmpipe_screen *screen, + struct llvmpipe_texture * lpt) +{ + struct llvmpipe_winsys *winsys = screen->winsys; + + pf_get_block(lpt->base.format, &lpt->base.block); + lpt->base.nblocksx[0] = pf_get_nblocksx(&lpt->base.block, lpt->base.width[0]); + lpt->base.nblocksy[0] = pf_get_nblocksy(&lpt->base.block, lpt->base.height[0]); + + lpt->dt = winsys->displaytarget_create(winsys, + lpt->base.format, + lpt->base.width[0], + lpt->base.height[0], + 16, + &lpt->stride[0] ); + + return lpt->dt != NULL; +} + + + + + +static struct pipe_texture * +llvmpipe_texture_create(struct pipe_screen *_screen, + const struct pipe_texture *templat) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + struct llvmpipe_texture *lpt = CALLOC_STRUCT(llvmpipe_texture); + if (!lpt) + return NULL; + + lpt->base = *templat; + pipe_reference_init(&lpt->base.reference, 1); + lpt->base.screen = &screen->base; + + /* XXX: The xlib state tracker is brain-dead and will request + * PIPE_FORMAT_Z16_UNORM no matter how much we tell it we don't support it. + */ + if(lpt->base.format == PIPE_FORMAT_Z16_UNORM) + lpt->base.format = PIPE_FORMAT_Z32_UNORM; + + if (lpt->base.tex_usage & (PIPE_TEXTURE_USAGE_DISPLAY_TARGET | + PIPE_TEXTURE_USAGE_PRIMARY)) { + if (!llvmpipe_displaytarget_layout(screen, lpt)) + goto fail; + } + else { + if (!llvmpipe_texture_layout(screen, lpt)) + goto fail; + } + + return &lpt->base; + + fail: + FREE(lpt); + return NULL; +} + + +static struct pipe_texture * +llvmpipe_texture_blanket(struct pipe_screen * screen, + const struct pipe_texture *base, + const unsigned *stride, + struct pipe_buffer *buffer) +{ + /* FIXME */ +#if 0 + struct llvmpipe_texture *lpt; + assert(screen); + + /* Only supports one type */ + if (base->target != PIPE_TEXTURE_2D || + base->last_level != 0 || + base->depth[0] != 1) { + return NULL; + } + + lpt = CALLOC_STRUCT(llvmpipe_texture); + if (!lpt) + return NULL; + + lpt->base = *base; + pipe_reference_init(&lpt->base.reference, 1); + lpt->base.screen = screen; + lpt->base.nblocksx[0] = pf_get_nblocksx(&lpt->base.block, lpt->base.width[0]); + lpt->base.nblocksy[0] = pf_get_nblocksy(&lpt->base.block, lpt->base.height[0]); + lpt->stride[0] = stride[0]; + + pipe_buffer_reference(&lpt->buffer, buffer); + + return &lpt->base; +#else + return NULL; +#endif +} + + +static void +llvmpipe_texture_destroy(struct pipe_texture *pt) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(pt->screen); + struct llvmpipe_texture *lpt = llvmpipe_texture(pt); + + if(lpt->dt) { + struct llvmpipe_winsys *winsys = screen->winsys; + winsys->displaytarget_destroy(winsys, lpt->dt); + } + else + align_free(lpt->data); + + FREE(lpt); +} + + +static struct pipe_surface * +llvmpipe_get_tex_surface(struct pipe_screen *screen, + struct pipe_texture *pt, + unsigned face, unsigned level, unsigned zslice, + unsigned usage) +{ + struct llvmpipe_texture *lpt = llvmpipe_texture(pt); + struct pipe_surface *ps; + + assert(level <= pt->last_level); + + ps = CALLOC_STRUCT(pipe_surface); + if (ps) { + pipe_reference_init(&ps->reference, 1); + pipe_texture_reference(&ps->texture, pt); + ps->format = pt->format; + ps->width = pt->width[level]; + ps->height = pt->height[level]; + ps->offset = lpt->level_offset[level]; + ps->usage = usage; + + /* Because we are llvmpipe, anything that the state tracker + * thought was going to be done with the GPU will actually get + * done with the CPU. Let's adjust the flags to take that into + * account. + */ + if (ps->usage & PIPE_BUFFER_USAGE_GPU_WRITE) { + /* GPU_WRITE means "render" and that can involve reads (blending) */ + ps->usage |= PIPE_BUFFER_USAGE_CPU_WRITE | PIPE_BUFFER_USAGE_CPU_READ; + } + + if (ps->usage & PIPE_BUFFER_USAGE_GPU_READ) + ps->usage |= PIPE_BUFFER_USAGE_CPU_READ; + + if (ps->usage & (PIPE_BUFFER_USAGE_CPU_WRITE | + PIPE_BUFFER_USAGE_GPU_WRITE)) { + /* Mark the surface as dirty. The tile cache will look for this. */ + lpt->timestamp++; + llvmpipe_screen(screen)->timestamp++; + } + + ps->face = face; + ps->level = level; + ps->zslice = zslice; + + if (pt->target == PIPE_TEXTURE_CUBE) { + ps->offset += face * pt->nblocksy[level] * lpt->stride[level]; + } + else if (pt->target == PIPE_TEXTURE_3D) { + ps->offset += zslice * pt->nblocksy[level] * lpt->stride[level]; + } + else { + assert(face == 0); + assert(zslice == 0); + } + } + return ps; +} + + +static void +llvmpipe_tex_surface_destroy(struct pipe_surface *surf) +{ + /* Effectively do the texture_update work here - if texture images + * needed post-processing to put them into hardware layout, this is + * where it would happen. For llvmpipe, nothing to do. + */ + assert(surf->texture); + pipe_texture_reference(&surf->texture, NULL); + FREE(surf); +} + + +static struct pipe_transfer * +llvmpipe_get_tex_transfer(struct pipe_screen *screen, + struct pipe_texture *texture, + unsigned face, unsigned level, unsigned zslice, + enum pipe_transfer_usage usage, + unsigned x, unsigned y, unsigned w, unsigned h) +{ + struct llvmpipe_texture *lptex = llvmpipe_texture(texture); + struct llvmpipe_transfer *lpt; + + assert(texture); + assert(level <= texture->last_level); + + lpt = CALLOC_STRUCT(llvmpipe_transfer); + if (lpt) { + struct pipe_transfer *pt = &lpt->base; + pipe_texture_reference(&pt->texture, texture); + pt->format = texture->format; + pt->block = texture->block; + pt->x = x; + pt->y = y; + pt->width = w; + pt->height = h; + pt->nblocksx = texture->nblocksx[level]; + pt->nblocksy = texture->nblocksy[level]; + pt->stride = lptex->stride[level]; + pt->usage = usage; + pt->face = face; + pt->level = level; + pt->zslice = zslice; + + lpt->offset = lptex->level_offset[level]; + + if (texture->target == PIPE_TEXTURE_CUBE) { + lpt->offset += face * pt->nblocksy * pt->stride; + } + else if (texture->target == PIPE_TEXTURE_3D) { + lpt->offset += zslice * pt->nblocksy * pt->stride; + } + else { + assert(face == 0); + assert(zslice == 0); + } + return pt; + } + return NULL; +} + + +static void +llvmpipe_tex_transfer_destroy(struct pipe_transfer *transfer) +{ + /* Effectively do the texture_update work here - if texture images + * needed post-processing to put them into hardware layout, this is + * where it would happen. For llvmpipe, nothing to do. + */ + assert (transfer->texture); + pipe_texture_reference(&transfer->texture, NULL); + FREE(transfer); +} + + +static void * +llvmpipe_transfer_map( struct pipe_screen *_screen, + struct pipe_transfer *transfer ) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + ubyte *map, *xfer_map; + struct llvmpipe_texture *lpt; + + assert(transfer->texture); + lpt = llvmpipe_texture(transfer->texture); + + if(lpt->dt) { + struct llvmpipe_winsys *winsys = screen->winsys; + unsigned flags = 0; + + if (transfer->usage != PIPE_TRANSFER_READ) { + flags |= PIPE_BUFFER_USAGE_CPU_WRITE; + } + + if (transfer->usage != PIPE_TRANSFER_WRITE) { + flags |= PIPE_BUFFER_USAGE_CPU_READ; + } + + map = winsys->displaytarget_map(winsys, lpt->dt, flags); + if (map == NULL) + return NULL; + } + else + map = lpt->data; + + /* May want to different things here depending on read/write nature + * of the map: + */ + if (transfer->texture && transfer->usage != PIPE_TRANSFER_READ) + { + /* Do something to notify sharing contexts of a texture change. + * In llvmpipe, that would mean flushing the texture cache. + */ + screen->timestamp++; + } + + xfer_map = map + llvmpipe_transfer(transfer)->offset + + transfer->y / transfer->block.height * transfer->stride + + transfer->x / transfer->block.width * transfer->block.size; + /*printf("map = %p xfer map = %p\n", map, xfer_map);*/ + return xfer_map; +} + + +static void +llvmpipe_transfer_unmap(struct pipe_screen *_screen, + struct pipe_transfer *transfer) +{ + struct llvmpipe_screen *screen = llvmpipe_screen(_screen); + struct llvmpipe_texture *lpt; + + assert(transfer->texture); + lpt = llvmpipe_texture(transfer->texture); + + if(lpt->dt) { + struct llvmpipe_winsys *winsys = screen->winsys; + winsys->displaytarget_unmap(winsys, lpt->dt); + } +} + + +void +llvmpipe_init_texture_funcs(struct llvmpipe_context *lp) +{ +} + + +void +llvmpipe_init_screen_texture_funcs(struct pipe_screen *screen) +{ + screen->texture_create = llvmpipe_texture_create; + screen->texture_blanket = llvmpipe_texture_blanket; + screen->texture_destroy = llvmpipe_texture_destroy; + + screen->get_tex_surface = llvmpipe_get_tex_surface; + screen->tex_surface_destroy = llvmpipe_tex_surface_destroy; + + screen->get_tex_transfer = llvmpipe_get_tex_transfer; + screen->tex_transfer_destroy = llvmpipe_tex_transfer_destroy; + screen->transfer_map = llvmpipe_transfer_map; + screen->transfer_unmap = llvmpipe_transfer_unmap; +} diff --git a/src/gallium/drivers/llvmpipe/lp_texture.h b/src/gallium/drivers/llvmpipe/lp_texture.h new file mode 100644 index 00000000000..00a20763e43 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_texture.h @@ -0,0 +1,90 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_TEXTURE_H +#define LP_TEXTURE_H + + +#include "pipe/p_state.h" + + +struct pipe_context; +struct pipe_screen; +struct llvmpipe_context; +struct llvmpipe_displaytarget; + +struct llvmpipe_texture +{ + struct pipe_texture base; + + unsigned long level_offset[PIPE_MAX_TEXTURE_LEVELS]; + unsigned stride[PIPE_MAX_TEXTURE_LEVELS]; + + /** + * Display target, for textures with the PIPE_TEXTURE_USAGE_DISPLAY_TARGET + * usage. + */ + struct llvmpipe_displaytarget *dt; + + /** + * Malloc'ed data for regular textures, or a mapping to dt above. + */ + void *data; + + unsigned timestamp; +}; + +struct llvmpipe_transfer +{ + struct pipe_transfer base; + + unsigned long offset; +}; + + +/** cast wrappers */ +static INLINE struct llvmpipe_texture * +llvmpipe_texture(struct pipe_texture *pt) +{ + return (struct llvmpipe_texture *) pt; +} + +static INLINE struct llvmpipe_transfer * +llvmpipe_transfer(struct pipe_transfer *pt) +{ + return (struct llvmpipe_transfer *) pt; +} + + +extern void +llvmpipe_init_texture_funcs( struct llvmpipe_context *llvmpipe ); + +extern void +llvmpipe_init_screen_texture_funcs(struct pipe_screen *screen); + + +#endif /* LP_TEXTURE */ diff --git a/src/gallium/drivers/llvmpipe/lp_tile_cache.c b/src/gallium/drivers/llvmpipe/lp_tile_cache.c new file mode 100644 index 00000000000..143afec3d35 --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tile_cache.c @@ -0,0 +1,299 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * Texture tile caching. + * + * Author: + * Brian Paul + */ + +#include "pipe/p_inlines.h" +#include "util/u_memory.h" +#include "util/u_math.h" +#include "util/u_tile.h" +#include "util/u_rect.h" +#include "lp_context.h" +#include "lp_surface.h" +#include "lp_texture.h" +#include "lp_tile_soa.h" +#include "lp_tile_cache.h" + + +struct llvmpipe_tile_cache * +lp_create_tile_cache( struct pipe_screen *screen ) +{ + struct llvmpipe_tile_cache *tc; + + tc = CALLOC_STRUCT( llvmpipe_tile_cache ); + if(!tc) + return NULL; + + tc->screen = screen; + + return tc; +} + + +void +lp_destroy_tile_cache(struct llvmpipe_tile_cache *tc) +{ + struct pipe_screen *screen; + unsigned x, y; + + for (y = 0; y < MAX_HEIGHT; y += TILE_SIZE) { + for (x = 0; x < MAX_WIDTH; x += TILE_SIZE) { + struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; + + if(tile->color) + align_free(tile->color); + } + } + + if (tc->transfer) { + screen = tc->transfer->texture->screen; + screen->tex_transfer_destroy(tc->transfer); + } + + FREE( tc ); +} + + +/** + * Specify the surface to cache. + */ +void +lp_tile_cache_set_surface(struct llvmpipe_tile_cache *tc, + struct pipe_surface *ps) +{ + if (tc->transfer) { + struct pipe_screen *screen = tc->transfer->texture->screen; + + if (ps == tc->surface) + return; + + if (tc->transfer_map) { + screen->transfer_unmap(screen, tc->transfer); + tc->transfer_map = NULL; + } + + screen->tex_transfer_destroy(tc->transfer); + tc->transfer = NULL; + } + + tc->surface = ps; + + if (ps) { + struct pipe_screen *screen = ps->texture->screen; + unsigned x, y; + + tc->transfer = screen->get_tex_transfer(screen, ps->texture, ps->face, + ps->level, ps->zslice, + PIPE_TRANSFER_READ_WRITE, + 0, 0, ps->width, ps->height); + + for (y = 0; y < ps->height; y += TILE_SIZE) { + for (x = 0; x < ps->width; x += TILE_SIZE) { + struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; + + tile->status = LP_TILE_STATUS_UNDEFINED; + + if(!tile->color) + tile->color = align_malloc( TILE_SIZE*TILE_SIZE*NUM_CHANNELS, 16 ); + } + } + } +} + + +/** + * Return the transfer being cached. + */ +struct pipe_surface * +lp_tile_cache_get_surface(struct llvmpipe_tile_cache *tc) +{ + return tc->surface; +} + + +void +lp_tile_cache_map_transfers(struct llvmpipe_tile_cache *tc) +{ + if (tc->transfer && !tc->transfer_map) + tc->transfer_map = tc->screen->transfer_map(tc->screen, tc->transfer); +} + + +void +lp_tile_cache_unmap_transfers(struct llvmpipe_tile_cache *tc) +{ + if (tc->transfer_map) { + tc->screen->transfer_unmap(tc->screen, tc->transfer); + tc->transfer_map = NULL; + } +} + + +/** + * Set a tile to a solid color. + */ +static void +clear_tile(struct llvmpipe_cached_tile *tile, + uint8_t clear_color[4]) +{ + if (clear_color[0] == clear_color[1] && + clear_color[1] == clear_color[2] && + clear_color[2] == clear_color[3]) { + memset(tile->color, clear_color[0], TILE_SIZE * TILE_SIZE * 4); + } + else { + uint x, y, chan; + for (y = 0; y < TILE_SIZE; y++) + for (x = 0; x < TILE_SIZE; x++) + for (chan = 0; chan < 4; ++chan) + TILE_PIXEL(tile->color, x, y, chan) = clear_color[chan]; + } +} + + +/** + * Flush the tile cache: write all dirty tiles back to the transfer. + * any tiles "flagged" as cleared will be "really" cleared. + */ +void +lp_flush_tile_cache(struct llvmpipe_tile_cache *tc) +{ + struct pipe_transfer *pt = tc->transfer; + unsigned x, y; + + if(!pt) + return; + + /* push the tile to all positions marked as clear */ + for (y = 0; y < pt->height; y += TILE_SIZE) { + for (x = 0; x < pt->width; x += TILE_SIZE) { + struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; + + switch(tile->status) { + case LP_TILE_STATUS_UNDEFINED: + break; + + case LP_TILE_STATUS_CLEAR: { + /** + * Actually clear the tiles which were flagged as being in a clear state. + */ + + struct pipe_screen *screen = pt->texture->screen; + unsigned tw = TILE_SIZE; + unsigned th = TILE_SIZE; + void *dst; + + if (pipe_clip_tile(x, y, &tw, &th, pt)) + continue; + + dst = screen->transfer_map(screen, pt); + assert(dst); + if(!dst) + continue; + + util_fill_rect(dst, &pt->block, pt->stride, + x, y, tw, th, + tc->clear_val); + + screen->transfer_unmap(screen, pt); + break; + } + + case LP_TILE_STATUS_DEFINED: + lp_put_tile_rgba_soa(pt, x, y, tile->color); + break; + } + } + } +} + + +/** + * Get a tile from the cache. + * \param x, y position of tile, in pixels + */ +void * +lp_get_cached_tile(struct llvmpipe_tile_cache *tc, + unsigned x, unsigned y ) +{ + struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; + struct pipe_transfer *pt = tc->transfer; + + switch(tile->status) { + case LP_TILE_STATUS_CLEAR: + /* don't get tile from framebuffer, just clear it */ + clear_tile(tile, tc->clear_color); + tile->status = LP_TILE_STATUS_DEFINED; + break; + + case LP_TILE_STATUS_UNDEFINED: + /* get new tile data from transfer */ + lp_get_tile_rgba_soa(pt, x, y, tile->color); + tile->status = LP_TILE_STATUS_DEFINED; + break; + + case LP_TILE_STATUS_DEFINED: + /* nothing to do */ + break; + } + + return tile->color; +} + + +/** + * When a whole surface is being cleared to a value we can avoid + * fetching tiles above. + * Save the color and set a 'clearflag' for each tile of the screen. + */ +void +lp_tile_cache_clear(struct llvmpipe_tile_cache *tc, const float *rgba, + uint clearValue) +{ + struct pipe_transfer *pt = tc->transfer; + const unsigned w = pt->width; + const unsigned h = pt->height; + unsigned x, y, chan; + + for(chan = 0; chan < 4; ++chan) + tc->clear_color[chan] = float_to_ubyte(rgba[chan]); + + tc->clear_val = clearValue; + + /* push the tile to all positions marked as clear */ + for (y = 0; y < h; y += TILE_SIZE) { + for (x = 0; x < w; x += TILE_SIZE) { + struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE]; + tile->status = LP_TILE_STATUS_CLEAR; + } + } +} diff --git a/src/gallium/drivers/llvmpipe/lp_tile_cache.h b/src/gallium/drivers/llvmpipe/lp_tile_cache.h new file mode 100644 index 00000000000..6d8ba5ece7a --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tile_cache.h @@ -0,0 +1,106 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_TILE_CACHE_H +#define LP_TILE_CACHE_H + + +#include "pipe/p_compiler.h" +#include "lp_tile_soa.h" + + +enum llvmpipe_tile_status +{ + LP_TILE_STATUS_UNDEFINED = 0, + LP_TILE_STATUS_CLEAR = 1, + LP_TILE_STATUS_DEFINED = 2 +}; + + +struct llvmpipe_cached_tile +{ + enum llvmpipe_tile_status status; + + /** color in SOA format */ + uint8_t *color; +}; + + +/** XXX move these */ +#define MAX_WIDTH 2048 +#define MAX_HEIGHT 2048 + + +struct llvmpipe_tile_cache +{ + struct pipe_screen *screen; + struct pipe_surface *surface; /**< the surface we're caching */ + struct pipe_transfer *transfer; + void *transfer_map; + + struct llvmpipe_cached_tile entries[MAX_WIDTH/TILE_SIZE][MAX_HEIGHT/TILE_SIZE]; + + uint8_t clear_color[4]; /**< for color bufs */ + uint clear_val; /**< for z+stencil, or packed color clear value */ + + struct llvmpipe_cached_tile *last_tile; /**< most recently retrieved tile */ +}; + + +extern struct llvmpipe_tile_cache * +lp_create_tile_cache( struct pipe_screen *screen ); + +extern void +lp_destroy_tile_cache(struct llvmpipe_tile_cache *tc); + +extern void +lp_tile_cache_set_surface(struct llvmpipe_tile_cache *tc, + struct pipe_surface *lps); + +extern struct pipe_surface * +lp_tile_cache_get_surface(struct llvmpipe_tile_cache *tc); + +extern void +lp_tile_cache_map_transfers(struct llvmpipe_tile_cache *tc); + +extern void +lp_tile_cache_unmap_transfers(struct llvmpipe_tile_cache *tc); + +extern void +lp_flush_tile_cache(struct llvmpipe_tile_cache *tc); + +extern void +lp_tile_cache_clear(struct llvmpipe_tile_cache *tc, const float *rgba, + uint clearValue); + +extern void * +lp_get_cached_tile(struct llvmpipe_tile_cache *tc, + unsigned x, unsigned y ); + + +#endif /* LP_TILE_CACHE_H */ + diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.c b/src/gallium/drivers/llvmpipe/lp_tile_soa.c new file mode 100644 index 00000000000..4e4ccb31ccd --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tile_soa.c @@ -0,0 +1,931 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * RGBA/float tile get/put functions. + * Usable both by drivers and state trackers. + */ + + +#include "pipe/p_defines.h" +#include "pipe/p_inlines.h" + +#include "util/u_math.h" +#include "util/u_memory.h" +#include "util/u_rect.h" +#include "util/u_tile.h" +#include "lp_tile_cache.h" +#include "lp_tile_soa.h" + + +const unsigned char +tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = { + { 0, 1, 4, 5, 8, 9, 12, 13}, + { 2, 3, 6, 7, 10, 11, 14, 15} +}; + + + +/*** PIPE_FORMAT_A8R8G8B8_UNORM ***/ + +static void +a8r8g8b8_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const unsigned pixel = *src++; + TILE_PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff; + TILE_PIXEL(p, j, i, 1) = (pixel >> 8) & 0xff; + TILE_PIXEL(p, j, i, 2) = (pixel >> 0) & 0xff; + TILE_PIXEL(p, j, i, 3) = (pixel >> 24) & 0xff; + } + } +} + + +static void +a8r8g8b8_put_tile_rgba(unsigned *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, g, b, a; + r = TILE_PIXEL(p, j, i, 0); + g = TILE_PIXEL(p, j, i, 1); + b = TILE_PIXEL(p, j, i, 2); + a = TILE_PIXEL(p, j, i, 3); + *dst++ = (a << 24) | (r << 16) | (g << 8) | b; + } + } +} + + +/*** PIPE_FORMAT_A8R8G8B8_UNORM ***/ + +static void +x8r8g8b8_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const unsigned pixel = *src++; + TILE_PIXEL(p, j, i, 0) = (pixel >> 16) & 0xff; + TILE_PIXEL(p, j, i, 1) = (pixel >> 8) & 0xff; + TILE_PIXEL(p, j, i, 2) = (pixel >> 0) & 0xff; + TILE_PIXEL(p, j, i, 3) = 0xff; + } + } +} + + +static void +x8r8g8b8_put_tile_rgba(unsigned *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, g, b; + r = TILE_PIXEL(p, j, i, 0); + g = TILE_PIXEL(p, j, i, 1); + b = TILE_PIXEL(p, j, i, 2); + *dst++ = (0xff << 24) | (r << 16) | (g << 8) | b; + } + } +} + + +/*** PIPE_FORMAT_B8G8R8A8_UNORM ***/ + +static void +b8g8r8a8_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const unsigned pixel = *src++; + TILE_PIXEL(p, j, i, 0) = (pixel >> 8) & 0xff; + TILE_PIXEL(p, j, i, 1) = (pixel >> 16) & 0xff; + TILE_PIXEL(p, j, i, 2) = (pixel >> 24) & 0xff; + TILE_PIXEL(p, j, i, 3) = (pixel >> 0) & 0xff; + } + } +} + + +static void +b8g8r8a8_put_tile_rgba(unsigned *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, g, b, a; + r = TILE_PIXEL(p, j, i, 0); + g = TILE_PIXEL(p, j, i, 1); + b = TILE_PIXEL(p, j, i, 2); + a = TILE_PIXEL(p, j, i, 3); + *dst++ = (b << 24) | (g << 16) | (r << 8) | a; + } + } +} + + +/*** PIPE_FORMAT_A1R5G5B5_UNORM ***/ + +static void +a1r5g5b5_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const ushort pixel = *src++; + TILE_PIXEL(p, j, i, 0) = ((pixel >> 10) & 0x1f) * 255 / 31; + TILE_PIXEL(p, j, i, 1) = ((pixel >> 5) & 0x1f) * 255 / 31; + TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0x1f) * 255 / 31; + TILE_PIXEL(p, j, i, 3) = ((pixel >> 15) ) * 255; + } + } +} + + +static void +a1r5g5b5_put_tile_rgba(ushort *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, g, b, a; + r = TILE_PIXEL(p, j, i, 0); + g = TILE_PIXEL(p, j, i, 1); + b = TILE_PIXEL(p, j, i, 2); + a = TILE_PIXEL(p, j, i, 3); + r = r >> 3; /* 5 bits */ + g = g >> 3; /* 5 bits */ + b = b >> 3; /* 5 bits */ + a = a >> 7; /* 1 bit */ + *dst++ = (a << 15) | (r << 10) | (g << 5) | b; + } + } +} + + +/*** PIPE_FORMAT_A4R4G4B4_UNORM ***/ + +static void +a4r4g4b4_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const ushort pixel = *src++; + TILE_PIXEL(p, j, i, 0) = ((pixel >> 8) & 0xf) * 255 / 15; + TILE_PIXEL(p, j, i, 1) = ((pixel >> 4) & 0xf) * 255 / 15; + TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0xf) * 255 / 15; + TILE_PIXEL(p, j, i, 3) = ((pixel >> 12) ) * 255 / 15; + } + } +} + + +static void +a4r4g4b4_put_tile_rgba(ushort *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, g, b, a; + r = TILE_PIXEL(p, j, i, 0); + g = TILE_PIXEL(p, j, i, 1); + b = TILE_PIXEL(p, j, i, 2); + a = TILE_PIXEL(p, j, i, 3); + r >>= 4; + g >>= 4; + b >>= 4; + a >>= 4; + *dst++ = (a << 12) | (r << 16) | (g << 4) | b; + } + } +} + + +/*** PIPE_FORMAT_R5G6B5_UNORM ***/ + +static void +r5g6b5_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + const ushort pixel = *src++; + TILE_PIXEL(p, j, i, 0) = ((pixel >> 11) & 0x1f) * 255 / 31; + TILE_PIXEL(p, j, i, 1) = ((pixel >> 5) & 0x3f) * 255 / 63; + TILE_PIXEL(p, j, i, 2) = ((pixel ) & 0x1f) * 255 / 31; + TILE_PIXEL(p, j, i, 3) = 255; + } + } +} + + +static void +r5g6b5_put_tile_rgba(ushort *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + uint r = (uint) TILE_PIXEL(p, j, i, 0) * 31 / 255; + uint g = (uint) TILE_PIXEL(p, j, i, 1) * 63 / 255; + uint b = (uint) TILE_PIXEL(p, j, i, 2) * 31 / 255; + *dst++ = (r << 11) | (g << 5) | (b); + } + } +} + + + +/*** PIPE_FORMAT_Z16_UNORM ***/ + +/** + * Return each Z value as four floats in [0,1]. + */ +static void +z16_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + const float scale = 1.0f / 65535.0f; + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = *src++ * scale; + } + } +} + + + + +/*** PIPE_FORMAT_L8_UNORM ***/ + +static void +l8_get_tile_rgba(const ubyte *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, src++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = *src; + TILE_PIXEL(p, j, i, 3) = 255; + } + } +} + + +static void +l8_put_tile_rgba(ubyte *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r; + r = TILE_PIXEL(p, j, i, 0); + *dst++ = (ubyte) r; + } + } +} + + + +/*** PIPE_FORMAT_A8_UNORM ***/ + +static void +a8_get_tile_rgba(const ubyte *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, src++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = 0; + TILE_PIXEL(p, j, i, 3) = *src; + } + } +} + + +static void +a8_put_tile_rgba(ubyte *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned a; + a = TILE_PIXEL(p, j, i, 3); + *dst++ = (ubyte) a; + } + } +} + + + +/*** PIPE_FORMAT_R16_SNORM ***/ + +static void +r16_get_tile_rgba(const short *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, src++) { + TILE_PIXEL(p, j, i, 0) = MAX2(src[0] >> 7, 0); + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = 0; + TILE_PIXEL(p, j, i, 3) = 255; + } + } +} + + +static void +r16_put_tile_rgba(short *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, dst++) { + dst[0] = TILE_PIXEL(p, j, i, 0) << 7; + } + } +} + + +/*** PIPE_FORMAT_R16G16B16A16_SNORM ***/ + +static void +r16g16b16a16_get_tile_rgba(const short *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, src += 4) { + TILE_PIXEL(p, j, i, 0) = src[0] >> 8; + TILE_PIXEL(p, j, i, 1) = src[1] >> 8; + TILE_PIXEL(p, j, i, 2) = src[2] >> 8; + TILE_PIXEL(p, j, i, 3) = src[3] >> 8; + } + } +} + + +static void +r16g16b16a16_put_tile_rgba(short *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, dst += 4) { + dst[0] = TILE_PIXEL(p, j, i, 0) << 8; + dst[1] = TILE_PIXEL(p, j, i, 1) << 8; + dst[2] = TILE_PIXEL(p, j, i, 2) << 8; + dst[3] = TILE_PIXEL(p, j, i, 3) << 8; + } + } +} + + + +/*** PIPE_FORMAT_I8_UNORM ***/ + +static void +i8_get_tile_rgba(const ubyte *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++, src++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = *src; + } + } +} + + +static void +i8_put_tile_rgba(ubyte *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r; + r = TILE_PIXEL(p, j, i, 0); + *dst++ = (ubyte) r; + } + } +} + + +/*** PIPE_FORMAT_A8L8_UNORM ***/ + +static void +a8l8_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + ushort ra = *src++; + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = ra & 0xff; + TILE_PIXEL(p, j, i, 3) = ra >> 8; + } + } +} + + +static void +a8l8_put_tile_rgba(ushort *dst, + unsigned w, unsigned h, + const uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + unsigned r, a; + r = TILE_PIXEL(p, j, i, 0); + a = TILE_PIXEL(p, j, i, 3); + *dst++ = (a << 8) | r; + } + } +} + + + + +/*** PIPE_FORMAT_Z32_UNORM ***/ + +/** + * Return each Z value as four floats in [0,1]. + */ +static void +z32_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + const double scale = 1.0 / (double) 0xffffffff; + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = (float) (*src++ * scale); + } + } +} + + +/*** PIPE_FORMAT_S8Z24_UNORM ***/ + +/** + * Return Z component as four float in [0,1]. Stencil part ignored. + */ +static void +s8z24_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + const double scale = 1.0 / ((1 << 24) - 1); + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = (float) (scale * (*src++ & 0xffffff)); + } + } +} + + +/*** PIPE_FORMAT_Z24S8_UNORM ***/ + +/** + * Return Z component as four float in [0,1]. Stencil part ignored. + */ +static void +z24s8_get_tile_rgba(const unsigned *src, + unsigned w, unsigned h, + uint8_t *p) +{ + const double scale = 1.0 / ((1 << 24) - 1); + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = (float) (scale * (*src++ >> 8)); + } + } +} + + +/*** PIPE_FORMAT_Z32_FLOAT ***/ + +/** + * Return each Z value as four floats in [0,1]. + */ +static void +z32f_get_tile_rgba(const float *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = *src++; + } + } +} + + +/*** PIPE_FORMAT_YCBCR / PIPE_FORMAT_YCBCR_REV ***/ + +/** + * Convert YCbCr (or YCrCb) to RGBA. + */ +static void +ycbcr_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p, + boolean rev) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + /* do two texels at a time */ + for (j = 0; j < (w & ~1); j += 2, src += 2) { + const ushort t0 = src[0]; + const ushort t1 = src[1]; + const ubyte y0 = (t0 >> 8) & 0xff; /* luminance */ + const ubyte y1 = (t1 >> 8) & 0xff; /* luminance */ + ubyte cb, cr; + float r, g, b; + + if (rev) { + cb = t1 & 0xff; /* chroma U */ + cr = t0 & 0xff; /* chroma V */ + } + else { + cb = t0 & 0xff; /* chroma U */ + cr = t1 & 0xff; /* chroma V */ + } + + /* even pixel: y0,cr,cb */ + r = 1.164f * (y0-16) + 1.596f * (cr-128); + g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128); + b = 1.164f * (y0-16) + 2.018f * (cb-128); + TILE_PIXEL(p, j, i, 0) = r; + TILE_PIXEL(p, j, i, 1) = g; + TILE_PIXEL(p, j, i, 2) = b; + TILE_PIXEL(p, j, i, 3) = 255; + + /* odd pixel: use y1,cr,cb */ + r = 1.164f * (y1-16) + 1.596f * (cr-128); + g = 1.164f * (y1-16) - 0.813f * (cr-128) - 0.391f * (cb-128); + b = 1.164f * (y1-16) + 2.018f * (cb-128); + TILE_PIXEL(p, j + 1, i, 0) = r; + TILE_PIXEL(p, j + 1, i, 1) = g; + TILE_PIXEL(p, j + 1, i, 2) = b; + TILE_PIXEL(p, j + 1, i, 3) = 255; + } + /* do the last texel */ + if (w & 1) { + const ushort t0 = src[0]; + const ushort t1 = src[1]; + const ubyte y0 = (t0 >> 8) & 0xff; /* luminance */ + ubyte cb, cr; + float r, g, b; + + if (rev) { + cb = t1 & 0xff; /* chroma U */ + cr = t0 & 0xff; /* chroma V */ + } + else { + cb = t0 & 0xff; /* chroma U */ + cr = t1 & 0xff; /* chroma V */ + } + + /* even pixel: y0,cr,cb */ + r = 1.164f * (y0-16) + 1.596f * (cr-128); + g = 1.164f * (y0-16) - 0.813f * (cr-128) - 0.391f * (cb-128); + b = 1.164f * (y0-16) + 2.018f * (cb-128); + TILE_PIXEL(p, j, i, 0) = r; + TILE_PIXEL(p, j, i, 1) = g; + TILE_PIXEL(p, j, i, 2) = b; + TILE_PIXEL(p, j, i, 3) = 255; + } + } +} + + +static void +fake_get_tile_rgba(const ushort *src, + unsigned w, unsigned h, + uint8_t *p) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + for (j = 0; j < w; j++) { + TILE_PIXEL(p, j, i, 0) = + TILE_PIXEL(p, j, i, 1) = + TILE_PIXEL(p, j, i, 2) = + TILE_PIXEL(p, j, i, 3) = (i ^ j) & 1 ? 255 : 0; + } + } +} + + +static void +lp_tile_raw_to_rgba_soa(enum pipe_format format, + void *src, + uint w, uint h, + uint8_t *p) +{ + switch (format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + a8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_X8R8G8B8_UNORM: + x8r8g8b8_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_B8G8R8A8_UNORM: + b8g8r8a8_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_A1R5G5B5_UNORM: + a1r5g5b5_get_tile_rgba((ushort *) src, w, h, p); + break; + case PIPE_FORMAT_A4R4G4B4_UNORM: + a4r4g4b4_get_tile_rgba((ushort *) src, w, h, p); + break; + case PIPE_FORMAT_R5G6B5_UNORM: + r5g6b5_get_tile_rgba((ushort *) src, w, h, p); + break; + case PIPE_FORMAT_L8_UNORM: + l8_get_tile_rgba((ubyte *) src, w, h, p); + break; + case PIPE_FORMAT_A8_UNORM: + a8_get_tile_rgba((ubyte *) src, w, h, p); + break; + case PIPE_FORMAT_I8_UNORM: + i8_get_tile_rgba((ubyte *) src, w, h, p); + break; + case PIPE_FORMAT_A8L8_UNORM: + a8l8_get_tile_rgba((ushort *) src, w, h, p); + break; + case PIPE_FORMAT_R16_SNORM: + r16_get_tile_rgba((short *) src, w, h, p); + break; + case PIPE_FORMAT_R16G16B16A16_SNORM: + r16g16b16a16_get_tile_rgba((short *) src, w, h, p); + break; + case PIPE_FORMAT_Z16_UNORM: + z16_get_tile_rgba((ushort *) src, w, h, p); + break; + case PIPE_FORMAT_Z32_UNORM: + z32_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: + s8z24_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_Z24S8_UNORM: + case PIPE_FORMAT_Z24X8_UNORM: + z24s8_get_tile_rgba((unsigned *) src, w, h, p); + break; + case PIPE_FORMAT_Z32_FLOAT: + z32f_get_tile_rgba((float *) src, w, h, p); + break; + case PIPE_FORMAT_YCBCR: + ycbcr_get_tile_rgba((ushort *) src, w, h, p, FALSE); + break; + case PIPE_FORMAT_YCBCR_REV: + ycbcr_get_tile_rgba((ushort *) src, w, h, p, TRUE); + break; + default: + debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(format)); + fake_get_tile_rgba(src, w, h, p); + } +} + + +void +lp_get_tile_rgba_soa(struct pipe_transfer *pt, + uint x, uint y, + uint8_t *p) +{ + uint w = TILE_SIZE, h = TILE_SIZE; + void *packed; + + if (pipe_clip_tile(x, y, &w, &h, pt)) + return; + + packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size); + + if (!packed) + return; + + if(pt->format == PIPE_FORMAT_YCBCR || pt->format == PIPE_FORMAT_YCBCR_REV) + assert((x & 1) == 0); + + pipe_get_tile_raw(pt, x, y, w, h, packed, 0); + + lp_tile_raw_to_rgba_soa(pt->format, packed, w, h, p); + + FREE(packed); +} + + +void +lp_put_tile_rgba_soa(struct pipe_transfer *pt, + uint x, uint y, + const uint8_t *p) +{ + uint w = TILE_SIZE, h = TILE_SIZE; + void *packed; + + if (pipe_clip_tile(x, y, &w, &h, pt)) + return; + + packed = MALLOC(pf_get_nblocks(&pt->block, w, h) * pt->block.size); + + if (!packed) + return; + + switch (pt->format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + a8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p); + break; + case PIPE_FORMAT_X8R8G8B8_UNORM: + x8r8g8b8_put_tile_rgba((unsigned *) packed, w, h, p); + break; + case PIPE_FORMAT_B8G8R8A8_UNORM: + b8g8r8a8_put_tile_rgba((unsigned *) packed, w, h, p); + break; + case PIPE_FORMAT_A1R5G5B5_UNORM: + a1r5g5b5_put_tile_rgba((ushort *) packed, w, h, p); + break; + case PIPE_FORMAT_R5G6B5_UNORM: + r5g6b5_put_tile_rgba((ushort *) packed, w, h, p); + break; + case PIPE_FORMAT_R8G8B8A8_UNORM: + assert(0); + break; + case PIPE_FORMAT_A4R4G4B4_UNORM: + a4r4g4b4_put_tile_rgba((ushort *) packed, w, h, p); + break; + case PIPE_FORMAT_L8_UNORM: + l8_put_tile_rgba((ubyte *) packed, w, h, p); + break; + case PIPE_FORMAT_A8_UNORM: + a8_put_tile_rgba((ubyte *) packed, w, h, p); + break; + case PIPE_FORMAT_I8_UNORM: + i8_put_tile_rgba((ubyte *) packed, w, h, p); + break; + case PIPE_FORMAT_A8L8_UNORM: + a8l8_put_tile_rgba((ushort *) packed, w, h, p); + break; + case PIPE_FORMAT_R16_SNORM: + r16_put_tile_rgba((short *) packed, w, h, p); + break; + case PIPE_FORMAT_R16G16B16A16_SNORM: + r16g16b16a16_put_tile_rgba((short *) packed, w, h, p); + break; + case PIPE_FORMAT_Z16_UNORM: + /*z16_put_tile_rgba((ushort *) packed, w, h, p);*/ + break; + case PIPE_FORMAT_Z32_UNORM: + /*z32_put_tile_rgba((unsigned *) packed, w, h, p);*/ + break; + case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: + /*s8z24_put_tile_rgba((unsigned *) packed, w, h, p);*/ + break; + case PIPE_FORMAT_Z24S8_UNORM: + case PIPE_FORMAT_Z24X8_UNORM: + /*z24s8_put_tile_rgba((unsigned *) packed, w, h, p);*/ + break; + default: + debug_printf("%s: unsupported format %s\n", __FUNCTION__, pf_name(pt->format)); + } + + pipe_put_tile_raw(pt, x, y, w, h, packed, 0); + + FREE(packed); +} + + diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.h b/src/gallium/drivers/llvmpipe/lp_tile_soa.h new file mode 100644 index 00000000000..3d8c703b73d --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_tile_soa.h @@ -0,0 +1,81 @@ +/************************************************************************** + * + * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#ifndef LP_TILE_SOA_H +#define LP_TILE_SOA_H + +#include "pipe/p_compiler.h" +#include "tgsi/tgsi_exec.h" // for NUM_CHANNELS + + +#ifdef __cplusplus +extern "C" { +#endif + + +struct pipe_transfer; + + +/** + * Cache tile size (width and height). This needs to be a power of two. + */ +#define TILE_SIZE 64 + + +#define TILE_VECTOR_HEIGHT 2 +#define TILE_VECTOR_WIDTH 8 + +extern const unsigned char +tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH]; + +#define TILE_C_STRIDE (TILE_VECTOR_HEIGHT*TILE_VECTOR_WIDTH) +#define TILE_X_STRIDE (NUM_CHANNELS*TILE_C_STRIDE) +#define TILE_Y_STRIDE (TILE_VECTOR_HEIGHT*TILE_SIZE*NUM_CHANNELS) + +#define TILE_PIXEL(_p, _x, _y, _c) \ + ((_p)[((_y)/TILE_VECTOR_HEIGHT)*TILE_Y_STRIDE + \ + ((_x)/TILE_VECTOR_WIDTH)*TILE_X_STRIDE + \ + (_c)*TILE_C_STRIDE + \ + tile_offset[(_y) % TILE_VECTOR_HEIGHT][(_x) % TILE_VECTOR_WIDTH]]) + + +void +lp_get_tile_rgba_soa(struct pipe_transfer *pt, + uint x, uint y, + uint8_t *p); + +void +lp_put_tile_rgba_soa(struct pipe_transfer *pt, + uint x, uint y, + const uint8_t *p); + + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/src/gallium/drivers/llvmpipe/lp_winsys.h b/src/gallium/drivers/llvmpipe/lp_winsys.h new file mode 100644 index 00000000000..595481c2cbc --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_winsys.h @@ -0,0 +1,128 @@ +/************************************************************************** + * + * Copyright 2007-2009 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +/** + * @file + * llvmpipe public interface. + */ + + +#ifndef LP_WINSYS_H +#define LP_WINSYS_H + + +#include "pipe/p_compiler.h" // for boolean +#include "pipe/p_format.h" + + +#ifdef __cplusplus +extern "C" { +#endif + + +struct pipe_screen; +struct pipe_context; + + +/** + * Opaque pointer. + */ +struct llvmpipe_displaytarget; + + +/** + * This is the interface that llvmpipe expects any window system + * hosting it to implement. + * + * llvmpipe is for the most part a self sufficient driver. The only thing it + * does not know is how to display a surface. + */ +struct llvmpipe_winsys +{ + void + (*destroy)( struct llvmpipe_winsys *ws ); + + boolean + (*is_displaytarget_format_supported)( struct llvmpipe_winsys *ws, + enum pipe_format format ); + + /** + * Allocate storage for a render target. + * + * Often surfaces which are meant to be blitted to the front screen (i.e., + * display targets) must be allocated with special characteristics, memory + * pools, or obtained directly from the windowing system. + * + * This callback is invoked by the pipe_screen when creating a texture marked + * with the PIPE_TEXTURE_USAGE_DISPLAY_TARGET flag to get the underlying + * storage. + */ + struct llvmpipe_displaytarget * + (*displaytarget_create)( struct llvmpipe_winsys *ws, + enum pipe_format format, + unsigned width, unsigned height, + unsigned alignment, + unsigned *stride ); + + void * + (*displaytarget_map)( struct llvmpipe_winsys *ws, + struct llvmpipe_displaytarget *dt, + unsigned flags ); + + void + (*displaytarget_unmap)( struct llvmpipe_winsys *ws, + struct llvmpipe_displaytarget *dt ); + + /** + * @sa pipe_screen:flush_frontbuffer. + * + * This call will likely become asynchronous eventually. + */ + void + (*displaytarget_display)( struct llvmpipe_winsys *ws, + struct llvmpipe_displaytarget *dt, + void *context_private ); + + void + (*displaytarget_destroy)( struct llvmpipe_winsys *ws, + struct llvmpipe_displaytarget *dt ); +}; + + +struct pipe_context * +llvmpipe_create( struct pipe_screen * ); + + +struct pipe_screen * +llvmpipe_create_screen( struct llvmpipe_winsys * ); + + +#ifdef __cplusplus +} +#endif + +#endif /* LP_WINSYS_H */ diff --git a/src/gallium/drivers/llvmpipe/sp2lp.sh b/src/gallium/drivers/llvmpipe/sp2lp.sh new file mode 100755 index 00000000000..c45a81ce3cf --- /dev/null +++ b/src/gallium/drivers/llvmpipe/sp2lp.sh @@ -0,0 +1,34 @@ +#!/bin/sh +# +# Port changes from softpipe to llvmpipe. Invoke as +# +# sp2lp.sh <commit> +# +# Note that this will only affect llvmpipe -- you still need to actually +# cherry-pick/merge the softpipe changes themselves if they affect directories +# outside src/gallium/drivers/softpipe + +git format-patch \ + --keep-subject \ + --relative=src/gallium/drivers/softpipe \ + --src-prefix=a/src/gallium/drivers/llvmpipe/ \ + --dst-prefix=b/src/gallium/drivers/llvmpipe/ \ + --stdout "$1^1..$1" \ +| sed \ + -e 's/\<softpipe\>/llvmpipe/g' \ + -e 's/\<sp\>/lp/g' \ + -e 's/\<softpipe_/llvmpipe_/g' \ + -e 's/\<sp_/lp_/g' \ + -e 's/\<SP_/LP_/g' \ + -e 's/\<SOFTPIPE_/LLVMPIPE_/g' \ + -e 's/\<spt\>/lpt/g' \ + -e 's/\<sps\>/lps/g' \ + -e 's/\<spfs\>/lpfs/g' \ + -e 's/\<sptex\>/lptex/g' \ + -e 's/\<setup_\(point\|line\|tri\)\>/llvmpipe_\0/g' \ + -e 's/\<llvmpipe_cached_tile\>/llvmpipe_cached_tex_tile/g' \ + -e 's/_get_cached_tile_tex\>/_get_cached_tex_tile/g' \ + -e 's/\<TILE_SIZE\>/TEX_TILE_SIZE/g' \ + -e 's/\<tile_address\>/tex_tile_address/g' \ + -e 's/\<tile->data\.color\>/tile->color/g' \ +| patch -p1 |