Cell: checkpoint: start to SIMD-ize texture sampling

5 files changed, 33 insertions, 8 deletions

diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
index ba4d180cc0..412661061a 100644
--- a/src/mesa/pipe/cell/spu/spu_main.c
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -263,6 +263,16 @@ cmd_state_texture(const struct cell_command_texture *texture)
              spu.init.id, texture->start, texture->width, texture->height);
 
    memcpy(&spu.texture, texture, sizeof(*texture));
+   spu.tex_size = VEC_LITERAL(vector float,
+                              spu.texture.width,
+                              spu.texture.height,
+                              0.0,
+                              0.0);
+   spu.tex_size_mask = VEC_LITERAL(vector unsigned int,
+                                   spu.texture.width - 1,
+                                   spu.texture.height - 1,
+                                   0,
+                                   0);
 }
 
 
diff --git a/src/mesa/pipe/cell/spu/spu_main.h b/src/mesa/pipe/cell/spu/spu_main.h
index 7a12715b0b..02b62ee5cd 100644
--- a/src/mesa/pipe/cell/spu/spu_main.h
+++ b/src/mesa/pipe/cell/spu/spu_main.h
@@ -110,6 +110,10 @@ struct spu_global
 
    /** for converting RGBA to PIPE_FORMAT_x colors */
    vector unsigned char color_shuffle;
+
+   vector float tex_size;
+   vector unsigned int tex_size_mask; /**< == int(size - 1) */
+
 } ALIGN16_ATTRIB;
 
 
diff --git a/src/mesa/pipe/cell/spu/spu_texture.c b/src/mesa/pipe/cell/spu/spu_texture.c
index c1dc6bfe90..1cf958806f 100644
--- a/src/mesa/pipe/cell/spu/spu_texture.c
+++ b/src/mesa/pipe/cell/spu/spu_texture.c
@@ -128,12 +128,23 @@ get_tex_tile(uint i, uint j)
  * XXX this is extremely primitive for now.
  */
 uint
-sample_texture(float4 texcoord)
+sample_texture(vector float texcoord)
 {
+#if 0
    /* wrap/repeat */
-   uint i = (uint) (texcoord.f[0] * spu.texture.width) % spu.texture.width;
-   uint j = (uint) (texcoord.f[1] * spu.texture.height) % spu.texture.height;
+   uint i = (uint) (spu_extract(texcoord, 0) * spu.texture.width) % spu.texture.width;
+   uint j = (uint) (spu_extract(texcoord, 1) * spu.texture.height) % spu.texture.height;
    uint pos = get_tex_tile(i, j);
    uint texel = tex_tiles[pos].ui[j % TILE_SIZE][i % TILE_SIZE];
    return texel;
+#else
+   vector float tc = spu_mul(texcoord, spu.tex_size);
+   vector unsigned int itc = spu_convtu(tc, 0);
+   itc = spu_and(itc, spu.tex_size_mask);
+   uint i = spu_extract(itc, 0);
+   uint j = spu_extract(itc, 1);
+   uint pos = get_tex_tile(i, j);
+   uint texel = tex_tiles[pos].ui[j % TILE_SIZE][i % TILE_SIZE];
+   return texel;
+#endif
 }
diff --git a/src/mesa/pipe/cell/spu/spu_texture.h b/src/mesa/pipe/cell/spu/spu_texture.h
index 938a42b549..5bc8e71879 100644
--- a/src/mesa/pipe/cell/spu/spu_texture.h
+++ b/src/mesa/pipe/cell/spu/spu_texture.h
@@ -37,7 +37,7 @@ invalidate_tex_cache(void);
 
 
 extern uint
-sample_texture(float4 texcoord);
+sample_texture(vector float texcoord);
 
 
 #endif /* SPU_TEXTURE_H */
diff --git a/src/mesa/pipe/cell/spu/spu_tri.c b/src/mesa/pipe/cell/spu/spu_tri.c
index 3f46e75d7c..c148c75dd6 100644
--- a/src/mesa/pipe/cell/spu/spu_tri.c
+++ b/src/mesa/pipe/cell/spu/spu_tri.c
@@ -309,13 +309,13 @@ emit_quad( int x, int y, mask_t mask )
          eval_coeff(2, (float) x, (float) y, texcoords);
 
          if (spu_extract(mask, 0))
-            spu.ctile.ui[iy][ix] = sample_texture(texcoords[0]);
+            spu.ctile.ui[iy][ix] = sample_texture(texcoords[0].v);
          if (spu_extract(mask, 1))
-            spu.ctile.ui[iy][ix+1] = sample_texture(texcoords[1]);
+            spu.ctile.ui[iy][ix+1] = sample_texture(texcoords[1].v);
          if (spu_extract(mask, 2))
-            spu.ctile.ui[iy+1][ix] = sample_texture(texcoords[2]);
+            spu.ctile.ui[iy+1][ix] = sample_texture(texcoords[2].v);
          if (spu_extract(mask, 3))
-            spu.ctile.ui[iy+1][ix+1] = sample_texture(texcoords[3]);
+            spu.ctile.ui[iy+1][ix+1] = sample_texture(texcoords[3].v);
       }
       else {
          /* simple shading */