llvmpipe: Special case complementary and identify blend factors in SoA.

One multiplication instead of two.

Also fix floating point random number generation and verification.

TODO: Do the same for AoS blending.

4 files changed, 79 insertions, 44 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
index b5924cbb7d..d1c9b88f9b 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
@@ -320,9 +320,6 @@ lp_build_blend_aos(LLVMBuilderRef builder,
    if(!blend->rt[rt].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);
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
index b9c7a6ceed..30d261e979 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
@@ -1,6 +1,6 @@
 /**************************************************************************
  *
- * Copyright 2009 VMware, Inc.
+ * Copyright 2009-2010 VMware, Inc.
  * All Rights Reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -195,6 +195,13 @@ lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld,
 }
 
 
+static boolean
+lp_build_blend_factor_complementary(unsigned src_factor, unsigned dst_factor)
+{
+   return dst_factor == (src_factor ^ 0x10);
+}
+
+
 /**
  * Generate blend code in SOA mode.
  * \param rt  render target index (to index the blend / colormask state)
@@ -243,8 +250,41 @@ lp_build_blend_soa(LLVMBuilderRef builder,
             unsigned func = i < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
             boolean func_commutative = lp_build_blend_func_commutative(func);
 
-            /* It makes no sense to blend unless values are normalized */
-            assert(type.norm);
+	    if (func == PIPE_BLEND_ADD &&
+		lp_build_blend_factor_complementary(src_factor, dst_factor) && 0) {
+               /*
+                * Special case linear interpolation, (i.e., complementary factors).
+                */
+
+	       LLVMValueRef weight;
+	       if (src_factor < dst_factor) {
+		  weight = lp_build_blend_soa_factor(&bld, src_factor, i);
+		  res[i] = lp_build_lerp(&bld.base, weight, dst[i], src[i]);
+	       } else {
+		  weight = lp_build_blend_soa_factor(&bld, dst_factor, i);
+		  res[i] = lp_build_lerp(&bld.base, weight, src[i], dst[i]);
+	       }
+	       continue;
+	    }
+
+	    if ((func == PIPE_BLEND_ADD ||
+                 func == PIPE_BLEND_SUBTRACT ||
+                 func == PIPE_BLEND_REVERSE_SUBTRACT) &&
+		src_factor == dst_factor &&
+                type.floating) {
+               /*
+                * Special common factor.
+                *
+                * XXX: Only for floating points for now, since saturation will
+                * cause different results.
+                */
+
+	       LLVMValueRef factor;
+               factor = lp_build_blend_soa_factor(&bld, src_factor, i);
+               res[i] = lp_build_blend_func(&bld.base, func, src[i], dst[i]);
+               res[i] = lp_build_mul(&bld.base, res[i], factor);
+	       continue;
+	    }
 
             /*
              * Compute src/dst factors.
diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c
index d0389f0cb0..8b6b5e1298 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c
@@ -243,19 +243,6 @@ add_blend_test(LLVMModuleRef module,
 }
 
 
-/** 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,
@@ -423,19 +410,19 @@ compute_blend_ref(const struct pipe_blend_state *blend,
     */
    switch (blend->rt[0].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 */
+      res[0] = src_term[0] + dst_term[0]; /* R */
+      res[1] = src_term[1] + dst_term[1]; /* G */
+      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 */
+      res[0] = src_term[0] - dst_term[0]; /* R */
+      res[1] = src_term[1] - dst_term[1]; /* G */
+      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 */
+      res[0] = dst_term[0] - src_term[0]; /* R */
+      res[1] = dst_term[1] - src_term[1]; /* G */
+      res[2] = dst_term[2] - src_term[2]; /* B */
       break;
    case PIPE_BLEND_MIN:
       res[0] = MIN2(src_term[0], dst_term[0]); /* R */
@@ -456,13 +443,13 @@ compute_blend_ref(const struct pipe_blend_state *blend,
     */
    switch (blend->rt[0].alpha_func) {
    case PIPE_BLEND_ADD:
-      ADD_SAT(res[3], src_term[3], dst_term[3]); /* A */
+      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 */
+      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 */
+      res[3] = dst_term[3] - src_term[3]; /* A */
       break;
    case PIPE_BLEND_MIN:
       res[3] = MIN2(src_term[3], dst_term[3]); /* A */
@@ -676,6 +663,8 @@ test_one(unsigned verbose,
                fprintf(stderr, "  Ref%c: ", channel);
                dump_vec(stderr, type, ref + j*stride);
                fprintf(stderr, "\n");
+
+               fprintf(stderr, "\n");
             }
          }
       }
@@ -773,7 +762,7 @@ blend_funcs[] = {
 
 const struct lp_type blend_types[] = {
    /* float, fixed,  sign,  norm, width, len */
-   {   TRUE, FALSE, FALSE,  TRUE,    32,   4 }, /* f32 x 4 */
+   {   TRUE, FALSE,  TRUE, FALSE,    32,   4 }, /* f32 x 4 */
    {  FALSE, FALSE, FALSE,  TRUE,     8,  16 }, /* u8n x 16 */
 };
 
diff --git a/src/gallium/drivers/llvmpipe/lp_test_main.c b/src/gallium/drivers/llvmpipe/lp_test_main.c
index 7bbbc61d4c..7a0d06ae2c 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_main.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_main.c
@@ -205,16 +205,19 @@ random_elem(struct lp_type type, void *dst, unsigned index)
    assert(index < type.length);
    value = (double)rand()/(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 & rand());
+      if (type.floating) {
+         value *= 2.0;
+      }
+      else {
+         unsigned long long mask;
+	 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 & rand());
+      }
    }
    if(!type.sign)
       if(rand() & 1)
@@ -261,12 +264,18 @@ boolean
 compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps)
 {
    unsigned i;
+   eps *= type.floating ? 8.0 : 2.0;
    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)
+      double delta = res_elem - ref_elem;
+      if (ref_elem < -1.0 || ref_elem > 1.0) {
+	 delta /= ref_elem;
+      }
+      delta = fabs(delta);
+      if (delta >= eps) {
          return FALSE;
+      }
    }
 
    return TRUE;