draw,llvmpipe,util: add depth bias calculation for arb_depth_buffer_float

With this patch, the llvmpipe and draw modules will calculate the depth bias
according to floating point depth buffer semantics described in the
arb_depth_buffer_float specification, when the driver has a z buffer bound
with a format type of UTIL_FORMAT_TYPE_FLOAT.

By default, the driver will use the existing UNORM calculation for depth bias.

A new function, draw_set_zs_format, was added to calculate the Minimum
Resolvable Depth value and floating point depth sense for the draw module.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

5 files changed, 86 insertions, 20 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h
index 106288a6ad3..9ec18e8429e 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_context.h
@@ -124,9 +124,10 @@ struct llvmpipe_context {
    /** A fake frontface output for unfilled primitives */
    int face_slot;
 
-   /**< minimum resolvable depth value, for polygon offset */   
-   double mrd;
-   
+   /** Depth format and bias settings. */
+   boolean floating_point_depth;
+   double mrd;   /**< minimum resolvable depth value, for polygon offset */
+
    /** The tiling engine */
    struct lp_setup_context *setup;
    struct lp_setup_variant setup_variant;
diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c
index 8aee92b1b96..47e413b776c 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_derived.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c
@@ -186,7 +186,8 @@ void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
       llvmpipe_update_fs( llvmpipe );
 
    if (llvmpipe->dirty & (LP_NEW_FS |
-			  LP_NEW_RASTERIZER))
+                          LP_NEW_FRAMEBUFFER |
+                          LP_NEW_RASTERIZER))
       llvmpipe_update_setup( llvmpipe );
 
    if (llvmpipe->dirty & LP_NEW_BLEND_COLOR)
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.c b/src/gallium/drivers/llvmpipe/lp_state_setup.c
index dab5096c61d..ccc9fb902f3 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.c
@@ -256,6 +256,7 @@ lp_do_offset_tri(struct gallivm_state *gallivm,
    LLVMBuilderRef b = gallivm->builder;
    struct lp_build_context bld;
    struct lp_build_context flt_scalar_bld;
+   struct lp_build_context int_scalar_bld;
    LLVMValueRef zoffset, mult;
    LLVMValueRef z0_new, z1_new, z2_new;
    LLVMValueRef dzdxdzdy, dzdx, dzdy, dzxyz20, dyzzx01, dyzzx01_dzxyz20, dzx01_dyz20;
@@ -267,6 +268,8 @@ lp_do_offset_tri(struct gallivm_state *gallivm,
    LLVMValueRef zeroi = lp_build_const_int32(gallivm, 0);
    LLVMValueRef twoi = lp_build_const_int32(gallivm, 2);
    LLVMValueRef threei  = lp_build_const_int32(gallivm, 3);
+   LLVMValueRef mantissa_bits, exp, bias;
+   LLVMValueRef maxz_value, maxz0z1_value;
 
    /* (res12) = cross(e,f).xy */
    shuffles[0] = twoi;
@@ -300,17 +303,56 @@ lp_do_offset_tri(struct gallivm_state *gallivm,
    dzdx = LLVMBuildExtractElement(b, dzdxdzdy, zeroi, "");
    dzdy = LLVMBuildExtractElement(b, dzdxdzdy, onei, "");
 
-   /* zoffset = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale */
+   /* mult = MAX2(dzdx, dzdy) * pgon_offset_scale */
    max = LLVMBuildFCmp(b, LLVMRealUGT, dzdx, dzdy, "");
    max_value = LLVMBuildSelect(b, max, dzdx, dzdy, "max"); 
 
    mult = LLVMBuildFMul(b, max_value,
                         lp_build_const_float(gallivm, key->pgon_offset_scale), "");
-   zoffset = LLVMBuildFAdd(b,
-                           lp_build_const_float(gallivm, key->pgon_offset_units),
-                           mult, "zoffset");
 
    lp_build_context_init(&flt_scalar_bld, gallivm, lp_type_float_vec(32, 32));
+
+   if (key->floating_point_depth) {
+      /*
+       * bias = pgon_offset_units * 2^(exponent(max(z0, z1, z2)) - mantissa_bits) +
+       *           MAX2(dzdx, dzdy) * pgon_offset_scale
+       *
+       * NOTE: Assumes IEEE float32.
+       */
+      lp_build_context_init(&int_scalar_bld, gallivm, lp_type_int_vec(32, 32));
+
+      mantissa_bits = lp_build_const_int32(gallivm, 23);
+
+      maxz0z1_value = lp_build_max(&flt_scalar_bld,
+                         LLVMBuildExtractElement(b, attribv[0], twoi, ""),
+                         LLVMBuildExtractElement(b, attribv[1], twoi, ""));
+
+      maxz_value = lp_build_max(&flt_scalar_bld,
+                      LLVMBuildExtractElement(b, attribv[2], twoi, ""),
+                      maxz0z1_value);
+
+      /**
+       * XXX: TODO optimize this to quickly resolve a pow2 number through
+       *      an exponent only operation.
+       */
+      exp = lp_build_extract_exponent(&flt_scalar_bld, maxz_value, 0);
+      exp = lp_build_sub(&int_scalar_bld, exp, mantissa_bits);
+      exp = lp_build_int_to_float(&flt_scalar_bld, exp);
+
+      bias = LLVMBuildFMul(b, lp_build_exp2(&flt_scalar_bld, exp),
+                           lp_build_const_float(gallivm, key->pgon_offset_units),
+                           "bias");
+
+      zoffset = LLVMBuildFAdd(b, bias, mult, "zoffset");
+   } else {
+      /*
+       * bias = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale
+       */
+      zoffset = LLVMBuildFAdd(b,
+                              lp_build_const_float(gallivm, key->pgon_offset_units),
+                              mult, "zoffset");
+   }
+
    if (key->pgon_offset_clamp > 0) {
       zoffset = lp_build_min(&flt_scalar_bld,
                              lp_build_const_float(gallivm, key->pgon_offset_clamp),
@@ -849,7 +891,20 @@ lp_make_setup_variant_key(struct llvmpipe_context *lp,
    assert(key->spec_slot   == lp->color_slot [1]);
    assert(key->bspec_slot  == lp->bcolor_slot[1]);
 
-   key->pgon_offset_units = (float) (lp->rasterizer->offset_units * lp->mrd);
+   /*
+    * If depth is floating point, depth bias is calculated with respect
+    * to the primitive's maximum Z value. Retain the original depth bias
+    * value until that stage.
+    */
+   key->floating_point_depth = lp->floating_point_depth;
+
+   if (key->floating_point_depth) {
+      key->pgon_offset_units = (float) lp->rasterizer->offset_units;
+   } else {
+      key->pgon_offset_units =
+         (float) (lp->rasterizer->offset_units * lp->mrd);
+   }
+
    key->pgon_offset_scale = lp->rasterizer->offset_scale;
    key->pgon_offset_clamp = lp->rasterizer->offset_clamp;
    key->pad = 0;
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.h b/src/gallium/drivers/llvmpipe/lp_state_setup.h
index f49e575aa1f..82af8350fb7 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_setup.h
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.h
@@ -25,7 +25,8 @@ struct lp_setup_variant_key {
    unsigned flatshade_first:1;
    unsigned pixel_center_half:1;
    unsigned twoside:1;
-   unsigned pad:5;
+   unsigned floating_point_depth:1;
+   unsigned pad:4;
 
    /* TODO: get those floats out of the key and use a jit_context for setup */
    float pgon_offset_units;
diff --git a/src/gallium/drivers/llvmpipe/lp_state_surface.c b/src/gallium/drivers/llvmpipe/lp_state_surface.c
index 89098410280..c228c63d139 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_surface.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_surface.c
@@ -57,24 +57,32 @@ llvmpipe_set_framebuffer_state(struct pipe_context *pipe,
    assert(fb->height <= LP_MAX_HEIGHT);
 
    if (changed) {
+      /*
+       * If no depth buffer is bound, send the utility function the default
+       * format for no bound depth (PIPE_FORMAT_NONE).
+       */ 
+      enum pipe_format depth_format = fb->zsbuf ?
+         fb->zsbuf->format : PIPE_FORMAT_NONE;
+      const struct util_format_description *depth_desc =
+         util_format_description(depth_format);
+
       util_copy_framebuffer_state(&lp->framebuffer, fb);
 
       if (LP_PERF & PERF_NO_DEPTH) {
 	 pipe_surface_reference(&lp->framebuffer.zsbuf, NULL);
       }
 
-      /* Tell draw module how deep the Z/depth buffer is.
-       *
-       * If no depth buffer is bound, send the utility function the default
-       * format for no bound depth (PIPE_FORMAT_NONE).
-       *
-       * FIXME: mrd constant isn't right should use a value derived from
-       * current primitive not a constant (for float depth buffers)
+      /*
+       * Calculate the floating point depth sense and Minimum Resolvable Depth
+       * value for the llvmpipe module. This is separate from the draw module.
        */
-      lp->mrd = util_get_depth_format_mrd((lp->framebuffer.zsbuf) ?
-                  lp->framebuffer.zsbuf->format : PIPE_FORMAT_NONE);
+      lp->floating_point_depth =
+         (util_get_depth_format_type(depth_desc) == UTIL_FORMAT_TYPE_FLOAT);
+ 
+      lp->mrd = util_get_depth_format_mrd(depth_desc);
 
-      draw_set_mrd(lp->draw, lp->mrd);
+      /* Tell the draw module how deep the Z/depth buffer is. */
+      draw_set_zs_format(lp->draw, depth_format);
 
       lp_setup_bind_framebuffer( lp->setup, &lp->framebuffer );