From 67b957781d8195b8f8867e994c03b68f8dc5c807 Mon Sep 17 00:00:00 2001
From: Keith Whitwell <keithw@vmware.com>
Date: Tue, 7 Sep 2010 07:55:28 +0100
Subject: [PATCH] llvmpipe: pass linear masks to fragment shader

Fragment shader can extract the correct bits for each quad.
---
 src/gallium/drivers/llvmpipe/lp_rast_tri.c    | 66 -------------------
 .../drivers/llvmpipe/lp_rast_tri_tmp.h        |  6 +-
 src/gallium/drivers/llvmpipe/lp_state_fs.c    | 24 +++++--
 3 files changed, 23 insertions(+), 73 deletions(-)

diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index dbaa8e023a4..278375652be 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -68,36 +68,6 @@ block_full_16(struct lp_rasterizer_task *task,
 }
 
 #if !defined(PIPE_ARCH_SSE)
-static INLINE unsigned
-build_mask(int c, int dcdx, int dcdy)
-{
-   int mask = 0;
-
-   int c0 = c;
-   int c1 = c0 + dcdx;
-   int c2 = c1 + dcdx;
-   int c3 = c2 + dcdx;
-
-   mask |= ((c0 + 0 * dcdy) >> 31) & (1 << 0);
-   mask |= ((c0 + 1 * dcdy) >> 31) & (1 << 2);
-   mask |= ((c0 + 2 * dcdy) >> 31) & (1 << 8);
-   mask |= ((c0 + 3 * dcdy) >> 31) & (1 << 10);
-   mask |= ((c1 + 0 * dcdy) >> 31) & (1 << 1);
-   mask |= ((c1 + 1 * dcdy) >> 31) & (1 << 3);
-   mask |= ((c1 + 2 * dcdy) >> 31) & (1 << 9);
-   mask |= ((c1 + 3 * dcdy) >> 31) & (1 << 11); 
-   mask |= ((c2 + 0 * dcdy) >> 31) & (1 << 4);
-   mask |= ((c2 + 1 * dcdy) >> 31) & (1 << 6);
-   mask |= ((c2 + 2 * dcdy) >> 31) & (1 << 12);
-   mask |= ((c2 + 3 * dcdy) >> 31) & (1 << 14);
-   mask |= ((c3 + 0 * dcdy) >> 31) & (1 << 5);
-   mask |= ((c3 + 1 * dcdy) >> 31) & (1 << 7);
-   mask |= ((c3 + 2 * dcdy) >> 31) & (1 << 13);
-   mask |= ((c3 + 3 * dcdy) >> 31) & (1 << 15);
-  
-   return mask;
-}
-
 
 static INLINE unsigned
 build_mask_linear(int c, int dcdx, int dcdy)
@@ -219,42 +189,6 @@ build_mask_linear(int c, int dcdx, int dcdy)
    return _mm_movemask_epi8(result);
 }
 
-static INLINE unsigned
-build_mask(int c, int dcdx, int dcdy)
-{
-   __m128i step = _mm_setr_epi32(0, dcdx, dcdy, dcdx + dcdy);
-   __m128i c0 = _mm_set1_epi32(c);
-
-   /* Get values across the quad
-    */
-   __m128i cstep0 = _mm_add_epi32(c0, step);
-
-   /* Scale up step for moving between quads.
-    */
-   __m128i step4 = _mm_add_epi32(step, step);
-
-   /* Get values for the remaining quads:
-    */
-   __m128i cstep1 = _mm_add_epi32(cstep0, 
-				  _mm_shuffle_epi32(step4, _MM_SHUFFLE(1,1,1,1)));
-   __m128i cstep2 = _mm_add_epi32(cstep0,
-				  _mm_shuffle_epi32(step4, _MM_SHUFFLE(2,2,2,2)));
-   __m128i cstep3 = _mm_add_epi32(cstep2,
-				  _mm_shuffle_epi32(step4, _MM_SHUFFLE(1,1,1,1)));
-
-   /* pack pairs of results into epi16
-    */
-   __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
-   __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
-
-   /* pack into epi8, preserving sign bits
-    */
-   __m128i result = _mm_packs_epi16(cstep01, cstep23);
-
-   /* extract sign bits to create mask
-    */
-   return _mm_movemask_epi8(result);
-}
 
 #endif
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
index 52b9012036d..a99c8ecfa78 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
@@ -50,9 +50,9 @@ TAG(do_block_4)(struct lp_rasterizer_task *task,
    int j;
 
    for (j = 0; j < NR_PLANES; j++) {
-      mask &= ~build_mask(c[j] - 1, 
-			  -plane[j].dcdx,
-			  plane[j].dcdy);
+      mask &= ~build_mask_linear(c[j] - 1, 
+				 -plane[j].dcdx,
+				 plane[j].dcdy);
    }
 
    /* Now pass to the shader:
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 1bcc16dd6dd..6053e825e17 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -186,6 +186,7 @@ generate_quad_mask(LLVMBuilderRef builder,
    LLVMTypeRef i32t = LLVMInt32Type();
    LLVMValueRef bits[4];
    LLVMValueRef mask;
+   int shift;
 
    /*
     * XXX: We'll need a different path for 16 x u8
@@ -197,10 +198,25 @@ generate_quad_mask(LLVMBuilderRef builder,
    /*
     * mask_input >>= (quad * 4)
     */
+   
+   switch (quad) {
+   case 0:
+      shift = 0;
+      break;
+   case 1:
+      shift = 2;
+      break;
+   case 2:
+      shift = 8;
+      break;
+   case 3:
+      shift = 10;
+      break;
+   }
 
    mask_input = LLVMBuildLShr(builder,
                               mask_input,
-                              LLVMConstInt(i32t, quad * 4, 0),
+                              LLVMConstInt(i32t, shift, 0),
                               "");
 
    /*
@@ -211,9 +227,9 @@ generate_quad_mask(LLVMBuilderRef builder,
 
    bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
    bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
-   bits[2] = LLVMConstInt(i32t, 1 << 2, 0);
-   bits[3] = LLVMConstInt(i32t, 1 << 3, 0);
-
+   bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
+   bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
+   
    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 
    /*
-- 
2.30.2