llvmpipe: slightly simplify build_mask

[mesa.git] / src / gallium / drivers / llvmpipe / lp_rast_tri.c

diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index 5b3ad6e0a784412aa8c3fadd8425bfd15f9adcf7..dbaa8e023a4a70702147bb189706404397cb7400 100644 (file)
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -128,11 +128,71 @@ build_mask_linear(int c, int dcdx, int dcdy)
   
    return mask;
 }
+
+
+static INLINE void
+build_masks(int c, 
+           int cdiff,
+           int dcdx,
+           int dcdy,
+           unsigned *outmask,
+           unsigned *partmask)
+{
+   *outmask |= build_mask_linear(c, dcdx, dcdy);
+   *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
+}
+
 #else
 #include <emmintrin.h>
 #include "util/u_sse.h"
 
 
+static INLINE void
+build_masks(int c, 
+           int cdiff,
+           int dcdx,
+           int dcdy,
+           unsigned *outmask,
+           unsigned *partmask)
+{
+   __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+   __m128i xdcdy = _mm_set1_epi32(dcdy);
+
+   /* Get values across the quad
+    */
+   __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
+   __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
+   __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
+
+   {
+      __m128i cstep01, cstep23, result;
+
+      cstep01 = _mm_packs_epi32(cstep0, cstep1);
+      cstep23 = _mm_packs_epi32(cstep2, cstep3);
+      result = _mm_packs_epi16(cstep01, cstep23);
+
+      *outmask |= _mm_movemask_epi8(result);
+   }
+
+
+   {
+      __m128i cio4 = _mm_set1_epi32(cdiff);
+      __m128i cstep01, cstep23, result;
+
+      cstep0 = _mm_add_epi32(cstep0, cio4);
+      cstep1 = _mm_add_epi32(cstep1, cio4);
+      cstep2 = _mm_add_epi32(cstep2, cio4);
+      cstep3 = _mm_add_epi32(cstep3, cio4);
+
+      cstep01 = _mm_packs_epi32(cstep0, cstep1);
+      cstep23 = _mm_packs_epi32(cstep2, cstep3);
+      result = _mm_packs_epi16(cstep01, cstep23);
+
+      *partmask |= _mm_movemask_epi8(result);
+   }
+}
+
+
 static INLINE unsigned
 build_mask_linear(int c, int dcdx, int dcdy)
 {
@@ -169,13 +229,9 @@ build_mask(int c, int dcdx, int dcdy)
     */
    __m128i cstep0 = _mm_add_epi32(c0, step);
 
-   /* Scale up step for moving between quads.  This should probably
-    * be an arithmetic shift left, but there doesn't seem to be
-    * such a thing in SSE.  It's unlikely that the step value is
-    * going to be large enough to overflow across 4 pixels, though
-    * if it is that big, rendering will be incorrect anyway.
+   /* Scale up step for moving between quads.
     */
-   __m128i step4 = _mm_slli_epi32(step, 1);
+   __m128i step4 = _mm_add_epi32(step, step);
 
    /* Get values for the remaining quads:
     */
@@ -263,11 +319,14 @@ lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
       {
         const int dcdx = -plane[j].dcdx * 4;
         const int dcdy = plane[j].dcdy * 4;
-        const int cox = c[j] + plane[j].eo * 4;
-        const int cio = c[j] + plane[j].ei * 4 - 1;
-
-        outmask |= build_mask_linear(cox, dcdx, dcdy);
-        partmask |= build_mask_linear(cio, dcdx, dcdy);
+        const int cox = plane[j].eo * 4;
+        const int cio = plane[j].ei * 4 - 1;
+
+        build_masks(c[j] + cox,
+                    cio - cox,
+                    dcdx, dcdy, 
+                    &outmask,   /* sign bits from c[i][0..15] + cox */
+                    &partmask); /* sign bits from c[i][0..15] + cio */
       }
    }