llvmpipe: Fix perspective divide interpolation.

Intuition != mathematics, so this time I actually worked out the right
formula for first order approximation of perspective interpolation.

Ironically, per quad divide actually makes things slower when compared
with per pixel divide -- probably because the divide hardware unit is
rarely used, whereas the multiply unit is typically already saturated
and the first order approximation imply more multiplications.

2 files changed, 76 insertions, 9 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
index 2cf6f38c4b8..2a374f8c390 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
@@ -75,6 +75,33 @@
  */
 
 
+/**
+ * Do one perspective divide per quad.
+ *
+ * For perspective interpolation, the final attribute value is given
+ *
+ *  a' = a/w = a * oow
+ *
+ * where
+ *
+ *  a = a0 + dadx*x + dady*y
+ *  w = w0 + dwdx*x + dwdy*y
+ *  oow = 1/w = 1/(w0 + dwdx*x + dwdy*y)
+ *
+ * Instead of computing the division per pixel, with this macro we compute the
+ * division on the upper left pixel of each quad, and use a linear
+ * approximation in the remaining pixels, given by:
+ *
+ *  da'dx = (dadx - dwdx*a)*oow
+ *  da'dy = (dady - dwdy*a)*oow
+ *
+ * Ironically, this actually makes things slower -- probably because the
+ * divide hardware unit is rarely used, whereas the multiply unit is typically
+ * already saturated.
+ */
+#define PERSPECTIVE_DIVIDE_PER_QUAD 0
+
+
 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
 
@@ -107,7 +134,6 @@ coeffs_init(struct lp_build_interp_soa_context *bld,
    LLVMValueRef i1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
    LLVMValueRef i2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
    LLVMValueRef i3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
-   LLVMValueRef oow = NULL;
    unsigned attrib;
    unsigned chan;
 
@@ -213,22 +239,22 @@ coeffs_init(struct lp_build_interp_soa_context *bld,
 
             a = LLVMBuildFAdd(builder, a, dadq2, "");
 
+#if PERSPECTIVE_DIVIDE_PER_QUAD
             /*
-             * a    *= 1 / w
-             * dadq *= 1 / w
+             * a *= 1 / w
              */
 
             if (interp == LP_INTERP_PERSPECTIVE) {
                LLVMValueRef w = bld->a[0][3];
                assert(attrib != 0);
                assert(bld->mask[0] & TGSI_WRITEMASK_W);
-               if (!oow) {
-                  oow = lp_build_rcp(coeff_bld, w);
-                  lp_build_name(oow, "oow");
+               if (!bld->oow) {
+                  bld->oow = lp_build_rcp(coeff_bld, w);
+                  lp_build_name(bld->oow, "oow");
                }
-               a = lp_build_mul(coeff_bld, a, oow);
-               dadq = lp_build_mul(coeff_bld, dadq, oow);
+               a = lp_build_mul(coeff_bld, a, bld->oow);
             }
+#endif
 
             attrib_name(a, attrib, chan, ".a");
             attrib_name(dadq, attrib, chan, ".dadq");
@@ -250,6 +276,7 @@ attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
 {
    struct lp_build_context *coeff_bld = &bld->coeff_bld;
    LLVMValueRef shuffle = lp_build_const_int_vec(coeff_bld->type, quad_index);
+   LLVMValueRef oow = NULL;
    unsigned attrib;
    unsigned chan;
 
@@ -270,6 +297,8 @@ attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
                a = bld->attribs[0][chan];
             }
             else {
+               LLVMValueRef dadq;
+
                a = bld->a[attrib][chan];
 
                /*
@@ -280,10 +309,46 @@ attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
                                           a, coeff_bld->undef, shuffle, "");
 
                /*
+                * Get the derivatives.
+                */
+
+               dadq = bld->dadq[attrib][chan];
+
+#if PERSPECTIVE_DIVIDE_PER_QUAD
+               if (interp == LP_INTERP_PERSPECTIVE) {
+                  LLVMValueRef dwdq = bld->dadq[0][3];
+
+                  if (oow == NULL) {
+                     assert(bld->oow);
+                     oow = LLVMBuildShuffleVector(coeff_bld->builder,
+                                                  bld->oow, coeff_bld->undef,
+                                                  shuffle, "");
+                  }
+
+                  dadq = lp_build_sub(coeff_bld,
+                                      dadq,
+                                      lp_build_mul(coeff_bld, a, dwdq));
+                  dadq = lp_build_mul(coeff_bld, dadq, oow);
+               }
+#endif
+
+               /*
                 * Add the derivatives
                 */
 
-               a = lp_build_add(coeff_bld, a, bld->dadq[attrib][chan]);
+               a = lp_build_add(coeff_bld, a, dadq);
+
+#if !PERSPECTIVE_DIVIDE_PER_QUAD
+               if (interp == LP_INTERP_PERSPECTIVE) {
+                  if (oow == NULL) {
+                     LLVMValueRef w = bld->attribs[0][3];
+                     assert(attrib != 0);
+                     assert(bld->mask[0] & TGSI_WRITEMASK_W);
+                     oow = lp_build_rcp(coeff_bld, w);
+                  }
+                  a = lp_build_mul(coeff_bld, a, oow);
+               }
+#endif
 
                attrib_name(a, attrib, chan, "");
             }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.h b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
index 29055133011..3054030f739 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
@@ -64,6 +64,8 @@ struct lp_build_interp_soa_context
    LLVMValueRef a   [1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
    LLVMValueRef dadq[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
 
+   LLVMValueRef oow;
+
    LLVMValueRef attribs[1 + PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
 
    /*