cell: do texture sampling/filtering for four pixels at a time.

5 files changed, 161 insertions, 10 deletions

diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index 91a4c137e7c..c59be7defde 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -301,10 +301,14 @@ cmd_state_sampler(const struct cell_command_sampler *sampler)
    DEBUG_PRINTF("SAMPLER [%u]\n", sampler->unit);
 
    spu.sampler[sampler->unit] = sampler->state;
-   if (spu.sampler[sampler->unit].min_img_filter == PIPE_TEX_FILTER_LINEAR)
+   if (spu.sampler[sampler->unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) {
       spu.sample_texture[sampler->unit] = sample_texture_bilinear;
-   else
+      spu.sample_texture4[sampler->unit] = sample_texture4_bilinear;
+   }
+   else {
       spu.sample_texture[sampler->unit] = sample_texture_nearest;
+      spu.sample_texture4[sampler->unit] = sample_texture4_nearest;
+   }
 }
 
 
@@ -323,6 +327,9 @@ cmd_state_texture(const struct cell_command_texture *texture)
    spu.texture[unit].width = width;
    spu.texture[unit].height = height;
 
+   spu.texture[unit].width4 = spu_splats((float) width);
+   spu.texture[unit].height4 = spu_splats((float) height);
+
    spu.texture[unit].tiles_per_row = width / TILE_SIZE;
 
    spu.texture[unit].tex_size = (vector float) { width, height, 0.0, 0.0};
diff --git a/src/gallium/drivers/cell/spu/spu_funcs.c b/src/gallium/drivers/cell/spu/spu_funcs.c
index 7dd7fcd253a..13c234ea2eb 100644
--- a/src/gallium/drivers/cell/spu/spu_funcs.c
+++ b/src/gallium/drivers/cell/spu/spu_funcs.c
@@ -106,6 +106,7 @@ spu_txp(vector float s, vector float t, vector float r, vector float q,
 {
    //const uint unit = 0;
    struct vec_4x4 colors;
+#if 0
    vector float coords[4];
 
    coords[0] = s;
@@ -121,6 +122,9 @@ spu_txp(vector float s, vector float t, vector float r, vector float q,
    colors.v[3] = spu.sample_texture[unit](unit, coords[3]);
 
    _transpose_matrix4x4(colors.v, colors.v);
+#else
+   spu.sample_texture4[unit](s, t, r, q, unit, colors.v);
+#endif
    return colors;
 }
 
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index 82c9c69a3a8..5d14be51c25 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -67,6 +67,14 @@ typedef union {
 typedef vector float (*spu_sample_texture_func)(uint unit,
                                                 vector float texcoord);
 
+typedef void (*spu_sample_texture4_func)(vector float s,
+                                         vector float t,
+                                         vector float r,
+                                         vector float q,
+                                         uint unit,
+                                         vector float colors[4]);
+
+
 /** Function for performing per-fragment ops */
 typedef void (*spu_fragment_ops_func)(uint x, uint y,
                                       tile_t *colorTile,
@@ -107,10 +115,12 @@ struct spu_texture
    void *start;
    ushort width, height;
    ushort tiles_per_row;
-   vector float tex_size;
-   vector unsigned int tex_size_mask; /**< == int(size - 1) */
-   vector unsigned int tex_size_x_mask; /**< == int(size - 1) */
-   vector unsigned int tex_size_y_mask; /**< == int(size - 1) */
+   vector float tex_size; /**< == {width, height, 0, 0} */
+   vector float width4;   /**< == {width, width, width, width} */
+   vector float height4;  /**< == {height, height, height, height} */
+   vector unsigned int tex_size_mask; /**< == {width-1, height-1, 0, 0 } */
+   vector unsigned int tex_size_x_mask; /**< splat(width-1) */
+   vector unsigned int tex_size_y_mask; /**< splat(height-1) */
 } ALIGN16_ATTRIB;
 
 
@@ -159,6 +169,7 @@ struct spu_global
 
    /** Current texture sampler function */
    spu_sample_texture_func sample_texture[CELL_MAX_SAMPLERS];
+   spu_sample_texture4_func sample_texture4[CELL_MAX_SAMPLERS];
 
    /** Fragment program constants */
    vector float constants[4 * CELL_MAX_CONSTANTS];
diff --git a/src/gallium/drivers/cell/spu/spu_texture.c b/src/gallium/drivers/cell/spu/spu_texture.c
index 117b8a36f80..12e6ed1ba1e 100644
--- a/src/gallium/drivers/cell/spu/spu_texture.c
+++ b/src/gallium/drivers/cell/spu/spu_texture.c
@@ -26,6 +26,8 @@
  **************************************************************************/
 
 
+#include <transpose_matrix4x4.h>
+
 #include "pipe/p_compiler.h"
 #include "spu_main.h"
 #include "spu_texture.h"
@@ -91,10 +93,10 @@ static void
 get_four_texels(uint unit, vec_uint4 x, vec_uint4 y, vec_uint4 *texels)
 {
    const unsigned texture_ea = (uintptr_t) spu.texture[unit].start;
-   vec_uint4 tile_x = spu_rlmask(x, -5);
-   vec_uint4 tile_y = spu_rlmask(y, -5);
-   const qword offset_x = si_andi((qword) x, 0x1f);
-   const qword offset_y = si_andi((qword) y, 0x1f);
+   vec_uint4 tile_x = spu_rlmask(x, -5);  /* tile_x = x / 32 */
+   vec_uint4 tile_y = spu_rlmask(y, -5);  /* tile_y = y / 32 */
+   const qword offset_x = si_andi((qword) x, 0x1f); /* offset_x = x & 0x1f */
+   const qword offset_y = si_andi((qword) y, 0x1f); /* offset_y = y & 0x1f */
 
    const qword tiles_per_row = (qword) spu_splats(spu.texture[unit].tiles_per_row);
    const qword tile_size = (qword) spu_splats((unsigned) sizeof(tile_t));
@@ -132,6 +134,31 @@ sample_texture_nearest(uint unit, vector float texcoord)
 }
 
 
+/**
+ * \param colors  returned colors in SOA format (rrrr, gggg, bbbb, aaaa).
+ */
+void
+sample_texture4_nearest(vector float s, vector float t,
+                        vector float r, vector float q,
+                        uint unit, vector float colors[4])
+{
+   vector float ss = spu_mul(s, spu.texture[unit].width4);
+   vector float tt = spu_mul(t, spu.texture[unit].height4);
+   vector unsigned int is = spu_convtu(ss, 0);
+   vector unsigned int it = spu_convtu(tt, 0);
+   vec_uint4 texels[4];
+
+   /* GL_REPEAT wrap mode: */
+   is = spu_and(is, spu.texture[unit].tex_size_x_mask);
+   it = spu_and(it, spu.texture[unit].tex_size_y_mask);
+
+   get_four_texels(unit, is, it, texels);
+
+   /* convert four packed ARGBA pixels to float RRRR,GGGG,BBBB,AAAA */
+   spu_unpack_A8R8G8B8_transpose4(texels, colors);
+}
+
+
 vector float
 sample_texture_bilinear(uint unit, vector float texcoord)
 {
@@ -198,3 +225,93 @@ sample_texture_bilinear(uint unit, vector float texcoord)
 
    return texel_sum;
 }
+
+
+void
+sample_texture4_bilinear(vector float s, vector float t,
+                         vector float r, vector float q,
+                         uint unit, vector float colors[4])
+{
+   vector float ss = spu_madd(s, spu.texture[unit].width4,  spu_splats(-0.5f));
+   vector float tt = spu_madd(t, spu.texture[unit].height4, spu_splats(-0.5f));
+
+   vector unsigned int is0 = spu_convtu(ss, 0);
+   vector unsigned int it0 = spu_convtu(tt, 0);
+
+   /* is + 1, it + 1 */
+   vector unsigned int is1 = spu_add(is0, 1);
+   vector unsigned int it1 = spu_add(it0, 1);
+
+   /* PIPE_TEX_WRAP_REPEAT */
+   is0 = spu_and(is0, spu.texture[unit].tex_size_x_mask);
+   it0 = spu_and(it0, spu.texture[unit].tex_size_y_mask);
+   is1 = spu_and(is1, spu.texture[unit].tex_size_x_mask);
+   it1 = spu_and(it1, spu.texture[unit].tex_size_y_mask);
+
+   /* get packed int texels */
+   vector unsigned int texels[16];
+   get_four_texels(unit, is0, it0, texels + 0);  /* upper-left */
+   get_four_texels(unit, is1, it0, texels + 4);  /* upper-right */
+   get_four_texels(unit, is0, it1, texels + 8);  /* lower-left */
+   get_four_texels(unit, is1, it1, texels + 12); /* lower-right */
+
+   /* XXX possibly rework following code to compute the weighted sample
+    * colors with integer arithmetic for fewer int->float conversions.
+    */
+
+   /* convert packed int texels to float colors */
+   vector float ftexels[16];
+   spu_unpack_A8R8G8B8_transpose4(texels + 0, ftexels + 0);
+   spu_unpack_A8R8G8B8_transpose4(texels + 4, ftexels + 4);
+   spu_unpack_A8R8G8B8_transpose4(texels + 8, ftexels + 8);
+   spu_unpack_A8R8G8B8_transpose4(texels + 12, ftexels + 12);
+
+   /* Compute weighting factors in [0,1]
+    * Multiply texcoord by 1024, AND with 1023, convert back to float.
+    */
+   vector float ss1024 = spu_mul(ss, spu_splats(1024.0f));
+   vector signed int iss1024 = spu_convts(ss1024, 0);
+   iss1024 = spu_and(iss1024, 1023);
+   vector float sWeights0 = spu_convtf(iss1024, 10);
+
+   vector float tt1024 = spu_mul(tt, spu_splats(1024.0f));
+   vector signed int itt1024 = spu_convts(tt1024, 0);
+   itt1024 = spu_and(itt1024, 1023);
+   vector float tWeights0 = spu_convtf(itt1024, 10);
+
+   /* 1 - sWeight and 1 - tWeight */
+   vector float sWeights1 = spu_sub(spu_splats(1.0f), sWeights0);
+   vector float tWeights1 = spu_sub(spu_splats(1.0f), tWeights0);
+
+   /* reds, for four pixels */
+   ftexels[ 0] = spu_mul(ftexels[ 0], spu_mul(sWeights1, tWeights1)); /*ul*/
+   ftexels[ 4] = spu_mul(ftexels[ 4], spu_mul(sWeights0, tWeights1)); /*ur*/
+   ftexels[ 8] = spu_mul(ftexels[ 8], spu_mul(sWeights1, tWeights0)); /*ll*/
+   ftexels[12] = spu_mul(ftexels[12], spu_mul(sWeights0, tWeights0)); /*lr*/
+   colors[0] = spu_add(spu_add(ftexels[0], ftexels[4]),
+                       spu_add(ftexels[8], ftexels[12]));
+
+   /* greens, for four pixels */
+   ftexels[ 1] = spu_mul(ftexels[ 1], spu_mul(sWeights1, tWeights1)); /*ul*/
+   ftexels[ 5] = spu_mul(ftexels[ 5], spu_mul(sWeights0, tWeights1)); /*ur*/
+   ftexels[ 9] = spu_mul(ftexels[ 9], spu_mul(sWeights1, tWeights0)); /*ll*/
+   ftexels[13] = spu_mul(ftexels[13], spu_mul(sWeights0, tWeights0)); /*lr*/
+   colors[1] = spu_add(spu_add(ftexels[1], ftexels[5]),
+                       spu_add(ftexels[9], ftexels[13]));
+
+   /* blues, for four pixels */
+   ftexels[ 2] = spu_mul(ftexels[ 2], spu_mul(sWeights1, tWeights1)); /*ul*/
+   ftexels[ 6] = spu_mul(ftexels[ 6], spu_mul(sWeights0, tWeights1)); /*ur*/
+   ftexels[10] = spu_mul(ftexels[10], spu_mul(sWeights1, tWeights0)); /*ll*/
+   ftexels[14] = spu_mul(ftexels[14], spu_mul(sWeights0, tWeights0)); /*lr*/
+   colors[2] = spu_add(spu_add(ftexels[2], ftexels[6]),
+                       spu_add(ftexels[10], ftexels[14]));
+
+   /* alphas, for four pixels */
+   ftexels[ 3] = spu_mul(ftexels[ 3], spu_mul(sWeights1, tWeights1)); /*ul*/
+   ftexels[ 7] = spu_mul(ftexels[ 7], spu_mul(sWeights0, tWeights1)); /*ur*/
+   ftexels[11] = spu_mul(ftexels[11], spu_mul(sWeights1, tWeights0)); /*ll*/
+   ftexels[15] = spu_mul(ftexels[15], spu_mul(sWeights0, tWeights0)); /*lr*/
+   colors[3] = spu_add(spu_add(ftexels[3], ftexels[7]),
+                       spu_add(ftexels[11], ftexels[15]));
+}
diff --git a/src/gallium/drivers/cell/spu/spu_texture.h b/src/gallium/drivers/cell/spu/spu_texture.h
index f7c9738be88..f019e7d8eff 100644
--- a/src/gallium/drivers/cell/spu/spu_texture.h
+++ b/src/gallium/drivers/cell/spu/spu_texture.h
@@ -40,8 +40,20 @@ extern vector float
 sample_texture_nearest(uint unit, vector float texcoord);
 
 
+extern void
+sample_texture4_nearest(vector float s, vector float t,
+                        vector float r, vector float q,
+                        uint unit, vector float colors[4]);
+
+
 extern vector float
 sample_texture_bilinear(uint unit, vector float texcoord);
 
 
+extern void
+sample_texture4_bilinear(vector float s, vector float t,
+                         vector float r, vector float q,
+                         uint unit, vector float colors[4]);
+
+
 #endif /* SPU_TEXTURE_H */