diff options
Diffstat (limited to 'src/mesa/state_tracker/st_glsl_to_nir.cpp')
-rw-r--r-- | src/mesa/state_tracker/st_glsl_to_nir.cpp | 783 |
1 files changed, 352 insertions, 431 deletions
diff --git a/src/mesa/state_tracker/st_glsl_to_nir.cpp b/src/mesa/state_tracker/st_glsl_to_nir.cpp index b3594154aef..a341b3a7723 100644 --- a/src/mesa/state_tracker/st_glsl_to_nir.cpp +++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp @@ -30,7 +30,6 @@ #include "program/program.h" #include "program/prog_statevars.h" #include "program/prog_parameter.h" -#include "program/ir_to_mesa.h" #include "main/context.h" #include "main/mtypes.h" #include "main/errors.h" @@ -44,6 +43,7 @@ #include "st_shader_cache.h" #include "compiler/nir/nir.h" +#include "compiler/nir/nir_builder.h" #include "compiler/glsl_types.h" #include "compiler/glsl/glsl_to_nir.h" #include "compiler/glsl/gl_nir.h" @@ -51,12 +51,14 @@ #include "compiler/glsl/ir.h" #include "compiler/glsl/ir_optimization.h" #include "compiler/glsl/linker_util.h" +#include "compiler/glsl/program.h" +#include "compiler/glsl/shader_cache.h" #include "compiler/glsl/string_to_uint_map.h" static int type_size(const struct glsl_type *type) { - return type->count_attribute_slots(false); + return glsl_count_attribute_slots(type, false); } /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we @@ -74,7 +76,7 @@ st_nir_fixup_varying_slots(struct st_context *st, nir_shader *shader, assert(!st->allow_st_finalize_nir_twice); nir_foreach_variable_with_modes(var, shader, mode) { - if (var->data.location >= VARYING_SLOT_VAR0) { + if (var->data.location >= VARYING_SLOT_VAR0 && var->data.location < VARYING_SLOT_PATCH0) { var->data.location += 9; } else if (var->data.location == VARYING_SLOT_PNTC) { var->data.location = VARYING_SLOT_VAR8; @@ -85,19 +87,6 @@ st_nir_fixup_varying_slots(struct st_context *st, nir_shader *shader, } } -static void -st_shader_gather_info(nir_shader *nir, struct gl_program *prog) -{ - nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir)); - - /* Copy the info we just generated back into the gl_program */ - const char *prog_name = prog->info.name; - const char *prog_label = prog->info.label; - prog->info = nir->info; - prog->info.name = prog_name; - prog->info.label = prog_label; -} - /* input location assignment for VS inputs must be handled specially, so * that it is aligned w/ st's vbo state. * (This isn't the case with, for ex, FS inputs, which only need to agree @@ -134,7 +123,7 @@ st_nir_assign_vs_in_locations(struct nir_shader *nir) /* Re-lower global vars, to deal with any dead VS inputs. */ if (removed_inputs) - NIR_PASS_V(nir, nir_lower_global_vars_to_local); + NIR_PASS(_, nir, nir_lower_global_vars_to_local); } static int @@ -172,15 +161,6 @@ st_nir_lookup_parameter_index(struct gl_program *prog, nir_variable *var) * fails. In this case just find the first matching "color.*".. * * Note for arrays you could end up w/ color[n].f, for example. - * - * glsl_to_tgsi works slightly differently in this regard. It is - * emitting something more low level, so it just translates the - * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI, - * it just calculates the additional offset of struct field members - * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or - * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never - * needs to work backwards to get base var loc from the param-list - * which already has them separated out. */ if (!prog->sh.data->spirv) { int namelen = strlen(var->name); @@ -204,12 +184,13 @@ st_nir_assign_uniform_locations(struct gl_context *ctx, int shaderidx = 0; int imageidx = 0; - nir_foreach_uniform_variable(uniform, nir) { + nir_foreach_variable_with_modes(uniform, nir, nir_var_uniform | + nir_var_image) { int loc; const struct glsl_type *type = glsl_without_array(uniform->type); - if (!uniform->data.bindless && (type->is_sampler() || type->is_image())) { - if (type->is_sampler()) { + if (!uniform->data.bindless && (glsl_type_is_sampler(type) || glsl_type_is_image(type))) { + if (glsl_type_is_sampler(type)) { loc = shaderidx; shaderidx += type_size(uniform->type); } else { @@ -218,9 +199,6 @@ st_nir_assign_uniform_locations(struct gl_context *ctx, } } else if (uniform->state_slots) { const gl_state_index16 *const stateTokens = uniform->state_slots[0].tokens; - /* This state reference has already been setup by ir_to_mesa, but we'll - * get the same index back here. - */ unsigned comps; if (glsl_type_is_struct_or_ifc(type)) { @@ -252,188 +230,11 @@ st_nir_assign_uniform_locations(struct gl_context *ctx, } } -void -st_nir_opts(nir_shader *nir) -{ - bool progress; - - do { - progress = false; - - NIR_PASS_V(nir, nir_lower_vars_to_ssa); - - /* Linking deals with unused inputs/outputs, but here we can remove - * things local to the shader in the hopes that we can cleanup other - * things. This pass will also remove variables with only stores, so we - * might be able to make progress after it. - */ - NIR_PASS(progress, nir, nir_remove_dead_variables, - nir_var_function_temp | nir_var_shader_temp | - nir_var_mem_shared, - NULL); - - NIR_PASS(progress, nir, nir_opt_copy_prop_vars); - NIR_PASS(progress, nir, nir_opt_dead_write_vars); - - if (nir->options->lower_to_scalar) { - NIR_PASS_V(nir, nir_lower_alu_to_scalar, - nir->options->lower_to_scalar_filter, NULL); - NIR_PASS_V(nir, nir_lower_phis_to_scalar, false); - } - - NIR_PASS_V(nir, nir_lower_alu); - NIR_PASS_V(nir, nir_lower_pack); - NIR_PASS(progress, nir, nir_copy_prop); - NIR_PASS(progress, nir, nir_opt_remove_phis); - NIR_PASS(progress, nir, nir_opt_dce); - if (nir_opt_trivial_continues(nir)) { - progress = true; - NIR_PASS(progress, nir, nir_copy_prop); - NIR_PASS(progress, nir, nir_opt_dce); - } - NIR_PASS(progress, nir, nir_opt_if, false); - NIR_PASS(progress, nir, nir_opt_dead_cf); - NIR_PASS(progress, nir, nir_opt_cse); - NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true); - - NIR_PASS(progress, nir, nir_opt_phi_precision); - NIR_PASS(progress, nir, nir_opt_algebraic); - NIR_PASS(progress, nir, nir_opt_constant_folding); - - if (!nir->info.flrp_lowered) { - unsigned lower_flrp = - (nir->options->lower_flrp16 ? 16 : 0) | - (nir->options->lower_flrp32 ? 32 : 0) | - (nir->options->lower_flrp64 ? 64 : 0); - - if (lower_flrp) { - bool lower_flrp_progress = false; - - NIR_PASS(lower_flrp_progress, nir, nir_lower_flrp, - lower_flrp, - false /* always_precise */); - if (lower_flrp_progress) { - NIR_PASS(progress, nir, - nir_opt_constant_folding); - progress = true; - } - } - - /* Nothing should rematerialize any flrps, so we only need to do this - * lowering once. - */ - nir->info.flrp_lowered = true; - } - - NIR_PASS(progress, nir, nir_opt_undef); - NIR_PASS(progress, nir, nir_opt_conditional_discard); - if (nir->options->max_unroll_iterations) { - NIR_PASS(progress, nir, nir_opt_loop_unroll); - } - } while (progress); -} - -static void -shared_type_info(const struct glsl_type *type, unsigned *size, unsigned *align) -{ - assert(glsl_type_is_vector_or_scalar(type)); - - uint32_t comp_size = glsl_type_is_boolean(type) - ? 4 : glsl_get_bit_size(type) / 8; - unsigned length = glsl_get_vector_elements(type); - *size = comp_size * length, - *align = comp_size * (length == 3 ? 4 : length); -} - -/* First third of converting glsl_to_nir.. this leaves things in a pre- - * nir_lower_io state, so that shader variants can more easily insert/ - * replace variables, etc. - */ -static void -st_nir_preprocess(struct st_context *st, struct gl_program *prog, - struct gl_shader_program *shader_program, - gl_shader_stage stage) -{ - struct pipe_screen *screen = st->screen; - const nir_shader_compiler_options *options = - st->ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions; - assert(options); - nir_shader *nir = prog->nir; - - /* Set the next shader stage hint for VS and TES. */ - if (!nir->info.separate_shader && - (nir->info.stage == MESA_SHADER_VERTEX || - nir->info.stage == MESA_SHADER_TESS_EVAL)) { - - unsigned prev_stages = (1 << (prog->info.stage + 1)) - 1; - unsigned stages_mask = - ~prev_stages & shader_program->data->linked_stages; - - nir->info.next_stage = stages_mask ? - (gl_shader_stage) u_bit_scan(&stages_mask) : MESA_SHADER_FRAGMENT; - } else { - nir->info.next_stage = MESA_SHADER_FRAGMENT; - } - - nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir)); - if (!st->ctx->SoftFP64 && ((nir->info.bit_sizes_int | nir->info.bit_sizes_float) & 64) && - (options->lower_doubles_options & nir_lower_fp64_full_software) != 0) { - st->ctx->SoftFP64 = glsl_float64_funcs_to_nir(st->ctx, options); - } - - /* ES has strict SSO validation rules for shader IO matching so we can't - * remove dead IO until the resource list has been built. Here we skip - * removing them until later. This will potentially make the IO lowering - * calls below do a little extra work but should otherwise have no impact. - */ - if (!_mesa_is_gles(st->ctx) || !nir->info.separate_shader) { - nir_variable_mode mask = nir_var_shader_in | nir_var_shader_out; - nir_remove_dead_variables(nir, mask, NULL); - } - - if (options->lower_all_io_to_temps || - nir->info.stage == MESA_SHADER_VERTEX || - nir->info.stage == MESA_SHADER_GEOMETRY) { - NIR_PASS_V(nir, nir_lower_io_to_temporaries, - nir_shader_get_entrypoint(nir), - true, true); - } else if (nir->info.stage == MESA_SHADER_FRAGMENT || - !screen->get_param(screen, PIPE_CAP_TGSI_CAN_READ_OUTPUTS)) { - NIR_PASS_V(nir, nir_lower_io_to_temporaries, - nir_shader_get_entrypoint(nir), - true, false); - } - - NIR_PASS_V(nir, nir_lower_global_vars_to_local); - NIR_PASS_V(nir, nir_split_var_copies); - NIR_PASS_V(nir, nir_lower_var_copies); - - if (options->lower_to_scalar) { - NIR_PASS_V(nir, nir_lower_alu_to_scalar, - options->lower_to_scalar_filter, NULL); - } - - /* before buffers and vars_to_ssa */ - NIR_PASS_V(nir, gl_nir_lower_images, true); - - /* TODO: Change GLSL to not lower shared memory. */ - if (prog->nir->info.stage == MESA_SHADER_COMPUTE && - shader_program->data->spirv) { - NIR_PASS_V(prog->nir, nir_lower_vars_to_explicit_types, - nir_var_mem_shared, shared_type_info); - NIR_PASS_V(prog->nir, nir_lower_explicit_io, - nir_var_mem_shared, nir_address_format_32bit_offset); - } - - /* Do a round of constant folding to clean up address calculations */ - NIR_PASS_V(nir, nir_opt_constant_folding); -} - static bool -dest_is_64bit(nir_dest *dest, void *state) +def_is_64bit(nir_def *def, void *state) { bool *lower = (bool *)state; - if (dest && (nir_dest_bit_size(*dest) == 64)) { + if (def && (def->bit_size == 64)) { *lower = true; return false; } @@ -459,7 +260,7 @@ filter_64_bit_instr(const nir_instr *const_instr, UNUSED const void *data) * doesn't have const variants, so do the ugly const_cast here. */ nir_instr *instr = const_cast<nir_instr *>(const_instr); - nir_foreach_dest(instr, dest_is_64bit, &lower); + nir_foreach_def(instr, def_is_64bit, &lower); if (lower) return true; nir_foreach_src(instr, src_is_64bit, &lower); @@ -511,30 +312,21 @@ st_glsl_to_nir_post_opts(struct st_context *st, struct gl_program *prog, * storage is only associated with the original parameter list. * This should be enough for Bitmap and DrawPixels constants. */ - _mesa_ensure_and_associate_uniform_storage(st->ctx, shader_program, prog, 16); - - st_set_prog_affected_state_flags(prog); + _mesa_ensure_and_associate_uniform_storage(st->ctx, shader_program, prog, 28); /* None of the builtins being lowered here can be produced by SPIR-V. See * _mesa_builtin_uniform_desc. Also drivers that support packed uniform * storage don't need to lower builtins. */ if (!shader_program->data->spirv && - !st->ctx->Const.PackedDriverUniformStorage) { - /* at this point, array uniforms have been split into separate - * nir_variable structs where possible. this codepath can't handle dynamic - * array indexing, however, so all indirect uniform derefs - * must be eliminated beforehand to avoid trying to lower one of those builtins - */ - NIR_PASS_V(nir, nir_lower_indirect_builtin_uniform_derefs); - NIR_PASS_V(nir, st_nir_lower_builtin); - } + !st->ctx->Const.PackedDriverUniformStorage) + NIR_PASS(_, nir, st_nir_lower_builtin); if (!screen->get_param(screen, PIPE_CAP_NIR_ATOMICS_AS_DEREF)) - NIR_PASS_V(nir, gl_nir_lower_atomics, shader_program, true); + NIR_PASS(_, nir, gl_nir_lower_atomics, shader_program, true); - NIR_PASS_V(nir, nir_opt_intrinsics); - NIR_PASS_V(nir, nir_opt_fragdepth); + NIR_PASS(_, nir, nir_opt_intrinsics); + NIR_PASS(_, nir, nir_opt_fragdepth); /* Lower 64-bit ops. */ if (nir->options->lower_int64_options || @@ -542,40 +334,61 @@ st_glsl_to_nir_post_opts(struct st_context *st, struct gl_program *prog, bool lowered_64bit_ops = false; bool revectorize = false; - /* nir_lower_doubles is not prepared for vector ops, so if the backend doesn't - * request lower_alu_to_scalar until now, lower all 64 bit ops, and try to - * vectorize them afterwards again */ - if (!nir->options->lower_to_scalar) { - NIR_PASS(revectorize, nir, nir_lower_alu_to_scalar, filter_64_bit_instr, nullptr); - NIR_PASS(revectorize, nir, nir_lower_phis_to_scalar, false); - } - if (nir->options->lower_doubles_options) { + /* nir_lower_doubles is not prepared for vector ops, so if the backend doesn't + * request lower_alu_to_scalar until now, lower all 64 bit ops, and try to + * vectorize them afterwards again */ + if (!nir->options->lower_to_scalar) { + NIR_PASS(revectorize, nir, nir_lower_alu_to_scalar, filter_64_bit_instr, nullptr); + NIR_PASS(revectorize, nir, nir_lower_phis_to_scalar, false); + } + /* doubles lowering requires frexp to be lowered first if it will be, + * since the pass generates other 64-bit ops. Most backends lower + * frexp, and using doubles is rare, and using frexp is even more rare + * (no instances in shader-db), so we're not too worried about + * accidentally lowering a 32-bit frexp here. + */ + NIR_PASS(lowered_64bit_ops, nir, nir_lower_frexp); + NIR_PASS(lowered_64bit_ops, nir, nir_lower_doubles, st->ctx->SoftFP64, nir->options->lower_doubles_options); } if (nir->options->lower_int64_options) NIR_PASS(lowered_64bit_ops, nir, nir_lower_int64); - if (revectorize) - NIR_PASS_V(nir, nir_opt_vectorize, nullptr, nullptr); + if (revectorize && !nir->options->vectorize_vec2_16bit) + NIR_PASS(_, nir, nir_opt_vectorize, nullptr, nullptr); if (revectorize || lowered_64bit_ops) - st_nir_opts(nir); + gl_nir_opts(nir); } nir_variable_mode mask = nir_var_shader_in | nir_var_shader_out | nir_var_function_temp; nir_remove_dead_variables(nir, mask, NULL); - if (!st->has_hw_atomics && !screen->get_param(screen, PIPE_CAP_NIR_ATOMICS_AS_DEREF)) - NIR_PASS_V(nir, nir_lower_atomics_to_ssbo); + if (!st->has_hw_atomics && !screen->get_param(screen, PIPE_CAP_NIR_ATOMICS_AS_DEREF)) { + unsigned align_offset_state = 0; + if (st->ctx->Const.ShaderStorageBufferOffsetAlignment > 4) { + struct gl_program_parameter_list *params = prog->Parameters; + for (unsigned i = 0; i < shader_program->data->NumAtomicBuffers; i++) { + gl_state_index16 state[STATE_LENGTH] = { STATE_ATOMIC_COUNTER_OFFSET, (short)shader_program->data->AtomicBuffers[i].Binding }; + _mesa_add_state_reference(params, state); + } + align_offset_state = STATE_ATOMIC_COUNTER_OFFSET; + } + NIR_PASS(_, nir, nir_lower_atomics_to_ssbo, align_offset_state); + } + + st_set_prog_affected_state_flags(prog); st_finalize_nir_before_variants(nir); char *msg = NULL; - if (st->allow_st_finalize_nir_twice) - msg = st_finalize_nir(st, prog, shader_program, nir, true, true); + if (st->allow_st_finalize_nir_twice) { + st_serialize_base_nir(prog, nir); + msg = st_finalize_nir(st, prog, shader_program, nir, true, true, false); + } if (st->ctx->_Shader->Flags & GLSL_DUMP) { _mesa_log("\n"); @@ -592,9 +405,19 @@ st_glsl_to_nir_post_opts(struct st_context *st, struct gl_program *prog, static void st_nir_vectorize_io(nir_shader *producer, nir_shader *consumer) { - NIR_PASS_V(producer, nir_lower_io_to_vector, nir_var_shader_out); - NIR_PASS_V(producer, nir_opt_combine_stores, nir_var_shader_out); - NIR_PASS_V(consumer, nir_lower_io_to_vector, nir_var_shader_in); + if (consumer) + NIR_PASS(_, consumer, nir_lower_io_to_vector, nir_var_shader_in); + + if (!producer) + return; + + NIR_PASS(_, producer, nir_lower_io_to_vector, nir_var_shader_out); + + if (producer->info.stage == MESA_SHADER_TESS_CTRL && + producer->options->vectorize_tess_levels) + NIR_PASS(_, producer, nir_vectorize_tess_levels); + + NIR_PASS(_, producer, nir_opt_combine_stores, nir_var_shader_out); if ((producer)->info.stage != MESA_SHADER_TESS_CTRL) { /* Calling lower_io_to_vector creates output variable writes with @@ -603,90 +426,34 @@ st_nir_vectorize_io(nir_shader *producer, nir_shader *consumer) * them. This, in turn, creates temporary variables and extra * copy_deref intrinsics that we need to clean up. */ - NIR_PASS_V(producer, nir_lower_io_to_temporaries, + NIR_PASS(_, producer, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(producer), true, false); - NIR_PASS_V(producer, nir_lower_global_vars_to_local); - NIR_PASS_V(producer, nir_split_var_copies); - NIR_PASS_V(producer, nir_lower_var_copies); + NIR_PASS(_, producer, nir_lower_global_vars_to_local); + NIR_PASS(_, producer, nir_split_var_copies); + NIR_PASS(_, producer, nir_lower_var_copies); } /* Undef scalar store_deref intrinsics are not ignored by nir_lower_io, * so they must be removed before that. These passes remove them. */ - NIR_PASS_V(producer, nir_lower_vars_to_ssa); - NIR_PASS_V(producer, nir_opt_undef); - NIR_PASS_V(producer, nir_opt_dce); -} - -static void -st_nir_link_shaders(nir_shader *producer, nir_shader *consumer) -{ - if (producer->options->lower_to_scalar) { - NIR_PASS_V(producer, nir_lower_io_to_scalar_early, nir_var_shader_out); - NIR_PASS_V(consumer, nir_lower_io_to_scalar_early, nir_var_shader_in); - } - - nir_lower_io_arrays_to_elements(producer, consumer); - - st_nir_opts(producer); - st_nir_opts(consumer); - - if (nir_link_opt_varyings(producer, consumer)) - st_nir_opts(consumer); - - NIR_PASS_V(producer, nir_remove_dead_variables, nir_var_shader_out, NULL); - NIR_PASS_V(consumer, nir_remove_dead_variables, nir_var_shader_in, NULL); - - if (nir_remove_unused_varyings(producer, consumer)) { - NIR_PASS_V(producer, nir_lower_global_vars_to_local); - NIR_PASS_V(consumer, nir_lower_global_vars_to_local); - - st_nir_opts(producer); - st_nir_opts(consumer); - - /* Optimizations can cause varyings to become unused. - * nir_compact_varyings() depends on all dead varyings being removed so - * we need to call nir_remove_dead_variables() again here. - */ - NIR_PASS_V(producer, nir_remove_dead_variables, nir_var_shader_out, - NULL); - NIR_PASS_V(consumer, nir_remove_dead_variables, nir_var_shader_in, - NULL); - } - - nir_link_varying_precision(producer, consumer); -} - -static void -st_lower_patch_vertices_in(struct gl_shader_program *shader_prog) -{ - struct gl_linked_shader *linked_tcs = - shader_prog->_LinkedShaders[MESA_SHADER_TESS_CTRL]; - struct gl_linked_shader *linked_tes = - shader_prog->_LinkedShaders[MESA_SHADER_TESS_EVAL]; - - /* If we have a TCS and TES linked together, lower TES patch vertices. */ - if (linked_tcs && linked_tes) { - nir_shader *tcs_nir = linked_tcs->Program->nir; - nir_shader *tes_nir = linked_tes->Program->nir; - - /* The TES input vertex count is the TCS output vertex count, - * lower TES gl_PatchVerticesIn to a constant. - */ - uint32_t tes_patch_verts = tcs_nir->info.tess.tcs_vertices_out; - NIR_PASS_V(tes_nir, nir_lower_patch_vertices, tes_patch_verts, NULL); - } + NIR_PASS(_, producer, nir_lower_vars_to_ssa); + NIR_PASS(_, producer, nir_opt_undef); + NIR_PASS(_, producer, nir_opt_dce); } extern "C" { -void +bool st_nir_lower_wpos_ytransform(struct nir_shader *nir, struct gl_program *prog, struct pipe_screen *pscreen) { - if (nir->info.stage != MESA_SHADER_FRAGMENT) - return; + bool progress = false; + + if (nir->info.stage != MESA_SHADER_FRAGMENT) { + nir_shader_preserve_all_metadata(nir); + return progress; + } static const gl_state_index16 wposTransformState[STATE_LENGTH] = { STATE_FB_WPOS_Y_TRANSFORM @@ -697,20 +464,20 @@ st_nir_lower_wpos_ytransform(struct nir_shader *nir, sizeof(wpos_options.state_tokens)); wpos_options.fs_coord_origin_upper_left = pscreen->get_param(pscreen, - PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT); + PIPE_CAP_FS_COORD_ORIGIN_UPPER_LEFT); wpos_options.fs_coord_origin_lower_left = pscreen->get_param(pscreen, - PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT); + PIPE_CAP_FS_COORD_ORIGIN_LOWER_LEFT); wpos_options.fs_coord_pixel_center_integer = pscreen->get_param(pscreen, - PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER); + PIPE_CAP_FS_COORD_PIXEL_CENTER_INTEGER); wpos_options.fs_coord_pixel_center_half_integer = pscreen->get_param(pscreen, - PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER); + PIPE_CAP_FS_COORD_PIXEL_CENTER_HALF_INTEGER); if (nir_lower_wpos_ytransform(nir, &wpos_options)) { - nir_validate_shader(nir, "after nir_lower_wpos_ytransform"); _mesa_add_state_reference(prog->Parameters, wposTransformState); + progress = true; } static const gl_state_index16 pntcTransformState[STATE_LENGTH] = { @@ -719,17 +486,29 @@ st_nir_lower_wpos_ytransform(struct nir_shader *nir, if (nir_lower_pntc_ytransform(nir, &pntcTransformState)) { _mesa_add_state_reference(prog->Parameters, pntcTransformState); + progress = true; } + + return progress; } -bool -st_link_nir(struct gl_context *ctx, - struct gl_shader_program *shader_program) +static bool +st_link_glsl_to_nir(struct gl_context *ctx, + struct gl_shader_program *shader_program) { struct st_context *st = st_context(ctx); struct gl_linked_shader *linked_shader[MESA_SHADER_STAGES]; unsigned num_shaders = 0; + /* Return early if we are loading the shader from on-disk cache */ + if (st_load_nir_from_disk_cache(ctx, shader_program)) { + return GL_TRUE; + } + + MESA_TRACE_FUNC(); + + assert(shader_program->data->LinkStatus); + for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) { if (shader_program->_LinkedShaders[i]) linked_shader[num_shaders++] = shader_program->_LinkedShaders[i]; @@ -740,13 +519,12 @@ st_link_nir(struct gl_context *ctx, const nir_shader_compiler_options *options = st->ctx->Const.ShaderCompilerOptions[shader->Stage].NirOptions; struct gl_program *prog = shader->Program; - struct st_program *stp = (struct st_program *)prog; - _mesa_copy_linked_program_data(shader_program, shader); + shader->Program->info.separate_shader = shader_program->SeparateShader; assert(!prog->nir); - stp->shader_program = shader_program; - stp->state.type = PIPE_SHADER_IR_NIR; + prog->shader_program = shader_program; + prog->state.type = PIPE_SHADER_IR_NIR; /* Parameters will be filled during NIR linking. */ prog->Parameters = _mesa_new_parameter_list(); @@ -754,8 +532,6 @@ st_link_nir(struct gl_context *ctx, if (shader_program->data->spirv) { prog->nir = _mesa_spirv_to_nir(ctx, shader_program, shader->Stage, options); } else { - validate_ir_tree(shader->ir); - if (ctx->_Shader->Flags & GLSL_DUMP) { _mesa_log("\n"); _mesa_log("GLSL IR for linked %s program %d:\n", @@ -765,84 +541,97 @@ st_link_nir(struct gl_context *ctx, _mesa_log("\n\n"); } - prog->nir = glsl_to_nir(st->ctx, shader_program, shader->Stage, options); - st_nir_preprocess(st, prog, shader_program, shader->Stage); + prog->nir = glsl_to_nir(&st->ctx->Const, shader_program, shader->Stage, options); } - if (options->lower_to_scalar) { - NIR_PASS_V(shader->Program->nir, nir_lower_load_const_to_scalar); + memcpy(prog->nir->info.source_sha1, shader->linked_source_sha1, + SHA1_DIGEST_LENGTH); + + nir_shader_gather_info(prog->nir, nir_shader_get_entrypoint(prog->nir)); + if (!st->ctx->SoftFP64 && ((prog->nir->info.bit_sizes_int | prog->nir->info.bit_sizes_float) & 64) && + (options->lower_doubles_options & nir_lower_fp64_full_software) != 0) { + + /* It's not possible to use float64 on GLSL ES, so don't bother trying to + * build the support code. The support code depends on higher versions of + * desktop GLSL, so it will fail to compile (below) anyway. + */ + if (_mesa_is_desktop_gl(st->ctx) && st->ctx->Const.GLSLVersion >= 400) + st->ctx->SoftFP64 = glsl_float64_funcs_to_nir(st->ctx, options); } } - st_lower_patch_vertices_in(shader_program); - - /* For SPIR-V, we have to perform the NIR linking before applying - * st_nir_preprocess. - */ if (shader_program->data->spirv) { static const gl_nir_linker_options opts = { true /*fill_parameters */ }; - if (!gl_nir_link_spirv(ctx, shader_program, &opts)) + if (!gl_nir_link_spirv(&ctx->Const, &ctx->Extensions, shader_program, + &opts)) return GL_FALSE; - - nir_build_program_resource_list(ctx, shader_program, true); - - for (unsigned i = 0; i < num_shaders; i++) { - struct gl_linked_shader *shader = linked_shader[i]; - struct gl_program *prog = shader->Program; - - prog->ExternalSamplersUsed = gl_external_samplers(prog); - _mesa_update_shader_textures_used(shader_program, prog); - st_nir_preprocess(st, prog, shader_program, shader->Stage); - } - } - - /* Linking the stages in the opposite order (from fragment to vertex) - * ensures that inter-shader outputs written to in an earlier stage - * are eliminated if they are (transitively) not used in a later - * stage. - */ - for (int i = num_shaders - 2; i >= 0; i--) { - st_nir_link_shaders(linked_shader[i]->Program->nir, - linked_shader[i + 1]->Program->nir); - } - /* Linking shaders also optimizes them. Separate shaders, compute shaders - * and shaders with a fixed-func VS or FS that don't need linking are - * optimized here. - */ - if (num_shaders == 1) - st_nir_opts(linked_shader[0]->Program->nir); - - if (!shader_program->data->spirv) { + } else { if (!gl_nir_link_glsl(ctx, shader_program)) return GL_FALSE; + } - for (unsigned i = 0; i < num_shaders; i++) { - struct gl_program *prog = linked_shader[i]->Program; - prog->ExternalSamplersUsed = gl_external_samplers(prog); - _mesa_update_shader_textures_used(shader_program, prog); - } - - nir_build_program_resource_list(ctx, shader_program, false); + for (unsigned i = 0; i < num_shaders; i++) { + struct gl_program *prog = linked_shader[i]->Program; + prog->ExternalSamplersUsed = gl_external_samplers(prog); + _mesa_update_shader_textures_used(shader_program, prog); } + nir_build_program_resource_list(&ctx->Const, shader_program, + shader_program->data->spirv); + for (unsigned i = 0; i < num_shaders; i++) { struct gl_linked_shader *shader = linked_shader[i]; nir_shader *nir = shader->Program->nir; + gl_shader_stage stage = shader->Stage; + const struct gl_shader_compiler_options *options = + &ctx->Const.ShaderCompilerOptions[stage]; - /* don't infer ACCESS_NON_READABLE so that Program->sh.ImageAccess is - * correct: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3278 + if (nir->info.io_lowered) { + /* Since IO is lowered, we won't need the IO variables from now on. + * nir_build_program_resource_list was the last pass that needed them. + */ + NIR_PASS_V(nir, nir_remove_dead_variables, + nir_var_shader_in | nir_var_shader_out, NULL); + } + + /* If there are forms of indirect addressing that the driver + * cannot handle, perform the lowering pass. */ - nir_opt_access_options opt_access_options; - opt_access_options.is_vulkan = false; - opt_access_options.infer_non_readable = false; - NIR_PASS_V(nir, nir_opt_access, &opt_access_options); + if (options->EmitNoIndirectInput || options->EmitNoIndirectOutput || + options->EmitNoIndirectTemp || options->EmitNoIndirectUniform) { + nir_variable_mode mode = (nir_variable_mode)0; + + if (!nir->info.io_lowered) { + mode |= options->EmitNoIndirectInput ? + nir_var_shader_in : (nir_variable_mode)0; + mode |= options->EmitNoIndirectOutput ? + nir_var_shader_out : (nir_variable_mode)0; + } + mode |= options->EmitNoIndirectTemp ? + nir_var_function_temp : (nir_variable_mode)0; + mode |= options->EmitNoIndirectUniform ? + nir_var_uniform | nir_var_mem_ubo | nir_var_mem_ssbo : + (nir_variable_mode)0; + + if (mode) + nir_lower_indirect_derefs(nir, mode, UINT32_MAX); + } /* This needs to run after the initial pass of nir_lower_vars_to_ssa, so * that the buffer indices are constants in nir where they where * constants in GLSL. */ - NIR_PASS_V(nir, gl_nir_lower_buffers, shader_program); + NIR_PASS(_, nir, gl_nir_lower_buffers, shader_program); + + NIR_PASS(_, nir, st_nir_lower_wpos_ytransform, shader->Program, + st->screen); + + NIR_PASS(_, nir, nir_lower_system_values); + NIR_PASS(_, nir, nir_lower_compute_system_values, NULL); + + if (nir->info.io_lowered) + continue; /* the rest is for non-lowered IO only */ /* Remap the locations to slots so those requiring two slots will occupy * two locations. For instance, if we have in the IR code a dvec3 attr0 in @@ -852,25 +641,6 @@ st_link_nir(struct gl_context *ctx, if (nir->info.stage == MESA_SHADER_VERTEX && !shader_program->data->spirv) nir_remap_dual_slot_attributes(nir, &shader->Program->DualSlotInputs); - NIR_PASS_V(nir, st_nir_lower_wpos_ytransform, shader->Program, - st->screen); - - NIR_PASS_V(nir, nir_lower_system_values); - NIR_PASS_V(nir, nir_lower_compute_system_values, NULL); - - NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays); - - st_shader_gather_info(nir, shader->Program); - if (shader->Stage == MESA_SHADER_VERTEX) { - /* NIR expands dual-slot inputs out to two locations. We need to - * compact things back down GL-style single-slot inputs to avoid - * confusing the state tracker. - */ - shader->Program->info.inputs_read = - nir_get_single_slot_attribs_mask(nir->info.inputs_read, - shader->Program->DualSlotInputs); - } - if (i >= 1) { struct gl_program *prev_shader = linked_shader[i - 1]->Program; @@ -888,6 +658,23 @@ st_link_nir(struct gl_context *ctx, } } + /* If the program is a separate shader program check if we need to vectorise + * the first and last program interfaces too. + */ + if (shader_program->SeparateShader && num_shaders > 0) { + struct gl_linked_shader *first_shader = linked_shader[0]; + struct gl_linked_shader *last_shader = linked_shader[num_shaders - 1]; + if (first_shader->Stage != MESA_SHADER_COMPUTE) { + if (ctx->Const.ShaderCompilerOptions[first_shader->Stage].NirOptions->vectorize_io && + first_shader->Stage > MESA_SHADER_VERTEX) + st_nir_vectorize_io(NULL, first_shader->Program->nir); + + if (ctx->Const.ShaderCompilerOptions[last_shader->Stage].NirOptions->vectorize_io && + last_shader->Stage < MESA_SHADER_FRAGMENT) + st_nir_vectorize_io(last_shader->Program->nir, NULL); + } + } + struct shader_info *prev_info = NULL; for (unsigned i = 0; i < num_shaders; i++) { @@ -897,7 +684,7 @@ st_link_nir(struct gl_context *ctx, char *msg = st_glsl_to_nir_post_opts(st, shader->Program, shader_program); if (msg) { linker_error(shader_program, msg); - break; + return false; } if (prev_info && @@ -916,7 +703,6 @@ st_link_nir(struct gl_context *ctx, for (unsigned i = 0; i < num_shaders; i++) { struct gl_linked_shader *shader = linked_shader[i]; struct gl_program *prog = shader->Program; - struct st_program *stp = st_program(prog); /* Make sure that prog->info is in sync with nir->info, but st/mesa * expects some of the values to be from before lowering. @@ -928,12 +714,25 @@ st_link_nir(struct gl_context *ctx, prog->info.num_ssbos = old_info.num_ssbos; prog->info.num_ubos = old_info.num_ubos; prog->info.num_abos = old_info.num_abos; - if (prog->info.stage == MESA_SHADER_VERTEX) - prog->info.inputs_read = old_info.inputs_read; - /* Initialize st_vertex_program members. */ - if (shader->Stage == MESA_SHADER_VERTEX) - st_prepare_vertex_program(stp, NULL); + if (prog->info.stage == MESA_SHADER_VERTEX) { + if (prog->nir->info.io_lowered && + prog->nir->options->io_options & nir_io_glsl_opt_varyings) { + prog->info.inputs_read = prog->nir->info.inputs_read; + prog->DualSlotInputs = prog->nir->info.dual_slot_inputs; + } else { + /* NIR expands dual-slot inputs out to two locations. We need to + * compact things back down GL-style single-slot inputs to avoid + * confusing the state tracker. + */ + prog->info.inputs_read = + nir_get_single_slot_attribs_mask(prog->nir->info.inputs_read, + prog->DualSlotInputs); + } + + /* Initialize st_vertex_program members. */ + st_prepare_vertex_program(prog); + } /* Get pipe_stream_output_info. */ if (shader->Stage == MESA_SHADER_VERTEX || @@ -941,14 +740,29 @@ st_link_nir(struct gl_context *ctx, shader->Stage == MESA_SHADER_GEOMETRY) st_translate_stream_output_info(prog); - st_store_ir_in_disk_cache(st, prog, true); + st_store_nir_in_disk_cache(st, prog); - st_release_variants(st, stp); + st_release_variants(st, prog); st_finalize_program(st, prog); + } + + struct pipe_context *pctx = st_context(ctx)->pipe; + if (pctx->link_shader) { + void *driver_handles[PIPE_SHADER_TYPES]; + memset(driver_handles, 0, sizeof(driver_handles)); - /* The GLSL IR won't be needed anymore. */ - ralloc_free(shader->ir); - shader->ir = NULL; + for (uint32_t i = 0; i < MESA_SHADER_STAGES; ++i) { + struct gl_linked_shader *shader = shader_program->_LinkedShaders[i]; + if (shader) { + struct gl_program *p = shader->Program; + if (p && p->variants) { + enum pipe_shader_type type = pipe_shader_type_from_mesa(shader->Stage); + driver_handles[type] = p->variants->driver_shader; + } + } + } + + pctx->link_shader(pctx, driver_handles); } return true; @@ -957,6 +771,10 @@ st_link_nir(struct gl_context *ctx, void st_nir_assign_varying_locations(struct st_context *st, nir_shader *nir) { + /* Lowered IO don't have variables, so exit. */ + if (nir->info.io_lowered) + return; + if (nir->info.stage == MESA_SHADER_VERTEX) { nir_assign_io_var_locations(nir, nir_var_shader_out, &nir->num_outputs, @@ -995,14 +813,17 @@ st_nir_lower_samplers(struct pipe_screen *screen, nir_shader *nir, struct gl_program *prog) { if (screen->get_param(screen, PIPE_CAP_NIR_SAMPLERS_AS_DEREF)) - NIR_PASS_V(nir, gl_nir_lower_samplers_as_deref, shader_program); + NIR_PASS(_, nir, gl_nir_lower_samplers_as_deref, shader_program); else - NIR_PASS_V(nir, gl_nir_lower_samplers, shader_program); + NIR_PASS(_, nir, gl_nir_lower_samplers, shader_program); if (prog) { BITSET_COPY(prog->info.textures_used, nir->info.textures_used); BITSET_COPY(prog->info.textures_used_by_txf, nir->info.textures_used_by_txf); - prog->info.images_used = nir->info.images_used; + BITSET_COPY(prog->info.samplers_used, nir->info.samplers_used); + BITSET_COPY(prog->info.images_used, nir->info.images_used); + BITSET_COPY(prog->info.image_buffers, nir->info.image_buffers); + BITSET_COPY(prog->info.msaa_images, nir->info.msaa_images); } } @@ -1022,17 +843,17 @@ void st_nir_lower_uniforms(struct st_context *st, nir_shader *nir) { if (st->ctx->Const.PackedDriverUniformStorage) { - NIR_PASS_V(nir, nir_lower_io, nir_var_uniform, + NIR_PASS(_, nir, nir_lower_io, nir_var_uniform, st_packed_uniforms_type_size, (nir_lower_io_options)0); } else { - NIR_PASS_V(nir, nir_lower_io, nir_var_uniform, + NIR_PASS(_, nir, nir_lower_io, nir_var_uniform, st_unpacked_uniforms_type_size, (nir_lower_io_options)0); } if (nir->options->lower_uniforms_to_ubo) - NIR_PASS_V(nir, nir_lower_uniforms_to_ubo, + NIR_PASS(_, nir, nir_lower_uniforms_to_ubo, st->ctx->Const.PackedDriverUniformStorage, !st->ctx->Const.NativeIntegers); } @@ -1044,30 +865,48 @@ char * st_finalize_nir(struct st_context *st, struct gl_program *prog, struct gl_shader_program *shader_program, nir_shader *nir, bool finalize_by_driver, - bool is_before_variants) + bool is_before_variants, + bool is_draw_shader) { struct pipe_screen *screen = st->screen; - NIR_PASS_V(nir, nir_split_var_copies); - NIR_PASS_V(nir, nir_lower_var_copies); + MESA_TRACE_FUNC(); - if (st->lower_rect_tex) { - struct nir_lower_tex_options opts = { 0 }; + NIR_PASS(_, nir, nir_split_var_copies); + NIR_PASS(_, nir, nir_lower_var_copies); - opts.lower_rect = true; + const bool lower_tg4_offsets = + !is_draw_shader && !st->screen->get_param(screen, PIPE_CAP_TEXTURE_GATHER_OFFSETS); - NIR_PASS_V(nir, nir_lower_tex, &opts); + if (!is_draw_shader && (st->lower_rect_tex || lower_tg4_offsets)) { + struct nir_lower_tex_options opts = {0}; + opts.lower_rect = !!st->lower_rect_tex; + opts.lower_tg4_offsets = lower_tg4_offsets; + + NIR_PASS(_, nir, nir_lower_tex, &opts); } st_nir_assign_varying_locations(st, nir); st_nir_assign_uniform_locations(st->ctx, prog, nir); + /* Lower load_deref/store_deref of inputs and outputs. + * This depends on st_nir_assign_varying_locations. + * + * TODO: remove this once nir_io_glsl_opt_varyings is enabled by default. + */ + if (!is_draw_shader && nir->options->io_options & nir_io_glsl_lower_derefs && + !(nir->options->io_options & nir_io_glsl_opt_varyings)) { + nir_lower_io_passes(nir, false); + NIR_PASS(_, nir, nir_remove_dead_variables, + nir_var_shader_in | nir_var_shader_out, NULL); + } + /* Set num_uniforms in number of attribute slots (vec4s) */ nir->num_uniforms = DIV_ROUND_UP(prog->Parameters->NumParameterValues, 4); st_nir_lower_uniforms(st, nir); - if (is_before_variants && nir->options->lower_uniforms_to_ubo) { + if (!is_draw_shader && is_before_variants && nir->options->lower_uniforms_to_ubo) { /* This must be done after uniforms are lowered to UBO and all * nir_var_uniform variables are removed from NIR to prevent conflicts * between state parameter merging and shader variant generation. @@ -1076,14 +915,96 @@ st_finalize_nir(struct st_context *st, struct gl_program *prog, } st_nir_lower_samplers(screen, nir, shader_program, prog); - if (!screen->get_param(screen, PIPE_CAP_NIR_IMAGES_AS_DEREF)) - NIR_PASS_V(nir, gl_nir_lower_images, false); + if (!is_draw_shader && !screen->get_param(screen, PIPE_CAP_NIR_IMAGES_AS_DEREF)) + NIR_PASS(_, nir, gl_nir_lower_images, false); char *msg = NULL; - if (finalize_by_driver && screen->finalize_nir) + if (!is_draw_shader && finalize_by_driver && screen->finalize_nir) msg = screen->finalize_nir(screen, nir); return msg; } +/** + * Link a GLSL shader program. Called via glLinkProgram(). + */ +void +st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog) +{ + unsigned int i; + bool spirv = false; + + MESA_TRACE_FUNC(); + + _mesa_clear_shader_program_data(ctx, prog); + + prog->data = _mesa_create_shader_program_data(); + + prog->data->LinkStatus = LINKING_SUCCESS; + + for (i = 0; i < prog->NumShaders; i++) { + if (!prog->Shaders[i]->CompileStatus) { + linker_error(prog, "linking with uncompiled/unspecialized shader"); + } + + if (!i) { + spirv = (prog->Shaders[i]->spirv_data != NULL); + } else if (spirv && !prog->Shaders[i]->spirv_data) { + /* The GL_ARB_gl_spirv spec adds a new bullet point to the list of + * reasons LinkProgram can fail: + * + * "All the shader objects attached to <program> do not have the + * same value for the SPIR_V_BINARY_ARB state." + */ + linker_error(prog, + "not all attached shaders have the same " + "SPIR_V_BINARY_ARB state"); + } + } + prog->data->spirv = spirv; + + if (prog->data->LinkStatus) { + if (!spirv) + link_shaders(ctx, prog); + else + _mesa_spirv_link_shaders(ctx, prog); + } + + /* If LinkStatus is LINKING_SUCCESS, then reset sampler validated to true. + * Validation happens via the LinkShader call below. If LinkStatus is + * LINKING_SKIPPED, then SamplersValidated will have been restored from the + * shader cache. + */ + if (prog->data->LinkStatus == LINKING_SUCCESS) { + prog->SamplersValidated = GL_TRUE; + } + + if (prog->data->LinkStatus && !st_link_glsl_to_nir(ctx, prog)) { + prog->data->LinkStatus = LINKING_FAILURE; + } + + if (prog->data->LinkStatus != LINKING_FAILURE) + _mesa_create_program_resource_hash(prog); + + /* Return early if we are loading the shader from on-disk cache */ + if (prog->data->LinkStatus == LINKING_SKIPPED) + return; + + if (ctx->_Shader->Flags & GLSL_DUMP) { + if (!prog->data->LinkStatus) { + fprintf(stderr, "GLSL shader program %d failed to link\n", prog->Name); + } + + if (prog->data->InfoLog && prog->data->InfoLog[0] != 0) { + fprintf(stderr, "GLSL shader program %d info log:\n", prog->Name); + fprintf(stderr, "%s\n", prog->data->InfoLog); + } + } + +#ifdef ENABLE_SHADER_CACHE + if (prog->data->LinkStatus) + shader_cache_write_program_metadata(ctx, prog); +#endif +} + } /* extern "C" */ |