Fetch routines convert and transpose all 4 vertices at once.

2 files changed, 71 insertions, 83 deletions

diff --git a/src/mesa/pipe/cell/spu/spu_vertex_fetch.c b/src/mesa/pipe/cell/spu/spu_vertex_fetch.c
index cbd389435e6..3bbf9b7be4f 100644
--- a/src/mesa/pipe/cell/spu/spu_vertex_fetch.c
+++ b/src/mesa/pipe/cell/spu/spu_vertex_fetch.c
@@ -100,7 +100,7 @@ fetch_unaligned(qword *dst, unsigned ea, unsigned size)
 }
 
 
-#define CVT_32_FLOAT(q)    (*q)
+#define CVT_32_FLOAT(q)    (*(q))
 
 static INLINE qword
 CVT_64_FLOAT(const qword *qw)
@@ -242,85 +242,90 @@ CVT_32_SNORM(const qword *qw)
  * This is probably needed/dupliocated elsewhere, eg format
  * conversion, texture sampling etc.
  */
-#define FETCH_ATTRIB( NAME, SZ, CVT )			\
-static qword						\
-fetch_##NAME(const qword *qw)				\
-{							\
-   qword expanded = CVT(qw);				\
-   return si_selb(expanded, (qword) defaults, SZ);	\
+#define FETCH_ATTRIB( NAME, SZ, CVT, N )			\
+static void							\
+fetch_##NAME(qword *out, const qword *in)			\
+{								\
+   qword tmp[4];						\
+								\
+   tmp[0] = si_selb(CVT(in + (0 * N)), (qword) defaults, SZ);	\
+   tmp[1] = si_selb(CVT(in + (1 * N)), (qword) defaults, SZ);	\
+   tmp[2] = si_selb(CVT(in + (2 * N)), (qword) defaults, SZ);	\
+   tmp[3] = si_selb(CVT(in + (3 * N)), (qword) defaults, SZ);	\
+   _transpose_matrix4x4((vec_float4 *) out, (vec_float4 *) tmp);	\
 }
 
-FETCH_ATTRIB( R64G64B64A64_FLOAT,   SZ_4, CVT_64_FLOAT )
-FETCH_ATTRIB( R64G64B64_FLOAT,      SZ_3, CVT_64_FLOAT )
-FETCH_ATTRIB( R64G64_FLOAT,         SZ_2, CVT_64_FLOAT )
-FETCH_ATTRIB( R64_FLOAT,            SZ_1, CVT_64_FLOAT )
+FETCH_ATTRIB( R64G64B64A64_FLOAT,   SZ_4, CVT_64_FLOAT, 2 )
+FETCH_ATTRIB( R64G64B64_FLOAT,      SZ_3, CVT_64_FLOAT, 2 )
+FETCH_ATTRIB( R64G64_FLOAT,         SZ_2, CVT_64_FLOAT, 2 )
+FETCH_ATTRIB( R64_FLOAT,            SZ_1, CVT_64_FLOAT, 2 )
 
-FETCH_ATTRIB( R32G32B32A32_FLOAT,   SZ_4, CVT_32_FLOAT )
-FETCH_ATTRIB( R32G32B32_FLOAT,      SZ_3, CVT_32_FLOAT )
-FETCH_ATTRIB( R32G32_FLOAT,         SZ_2, CVT_32_FLOAT )
-FETCH_ATTRIB( R32_FLOAT,            SZ_1, CVT_32_FLOAT )
+FETCH_ATTRIB( R32G32B32A32_FLOAT,   SZ_4, CVT_32_FLOAT, 1 )
+FETCH_ATTRIB( R32G32B32_FLOAT,      SZ_3, CVT_32_FLOAT, 1 )
+FETCH_ATTRIB( R32G32_FLOAT,         SZ_2, CVT_32_FLOAT, 1 )
+FETCH_ATTRIB( R32_FLOAT,            SZ_1, CVT_32_FLOAT, 1 )
 
-FETCH_ATTRIB( R32G32B32A32_USCALED, SZ_4, CVT_32_USCALED )
-FETCH_ATTRIB( R32G32B32_USCALED,    SZ_3, CVT_32_USCALED )
-FETCH_ATTRIB( R32G32_USCALED,       SZ_2, CVT_32_USCALED )
-FETCH_ATTRIB( R32_USCALED,          SZ_1, CVT_32_USCALED )
+FETCH_ATTRIB( R32G32B32A32_USCALED, SZ_4, CVT_32_USCALED, 1 )
+FETCH_ATTRIB( R32G32B32_USCALED,    SZ_3, CVT_32_USCALED, 1 )
+FETCH_ATTRIB( R32G32_USCALED,       SZ_2, CVT_32_USCALED, 1 )
+FETCH_ATTRIB( R32_USCALED,          SZ_1, CVT_32_USCALED, 1 )
 
-FETCH_ATTRIB( R32G32B32A32_SSCALED, SZ_4, CVT_32_SSCALED )
-FETCH_ATTRIB( R32G32B32_SSCALED,    SZ_3, CVT_32_SSCALED )
-FETCH_ATTRIB( R32G32_SSCALED,       SZ_2, CVT_32_SSCALED )
-FETCH_ATTRIB( R32_SSCALED,          SZ_1, CVT_32_SSCALED )
+FETCH_ATTRIB( R32G32B32A32_SSCALED, SZ_4, CVT_32_SSCALED, 1 )
+FETCH_ATTRIB( R32G32B32_SSCALED,    SZ_3, CVT_32_SSCALED, 1 )
+FETCH_ATTRIB( R32G32_SSCALED,       SZ_2, CVT_32_SSCALED, 1 )
+FETCH_ATTRIB( R32_SSCALED,          SZ_1, CVT_32_SSCALED, 1 )
 
-FETCH_ATTRIB( R32G32B32A32_UNORM, SZ_4, CVT_32_UNORM )
-FETCH_ATTRIB( R32G32B32_UNORM,    SZ_3, CVT_32_UNORM )
-FETCH_ATTRIB( R32G32_UNORM,       SZ_2, CVT_32_UNORM )
-FETCH_ATTRIB( R32_UNORM,          SZ_1, CVT_32_UNORM )
+FETCH_ATTRIB( R32G32B32A32_UNORM, SZ_4, CVT_32_UNORM, 1 )
+FETCH_ATTRIB( R32G32B32_UNORM,    SZ_3, CVT_32_UNORM, 1 )
+FETCH_ATTRIB( R32G32_UNORM,       SZ_2, CVT_32_UNORM, 1 )
+FETCH_ATTRIB( R32_UNORM,          SZ_1, CVT_32_UNORM, 1 )
 
-FETCH_ATTRIB( R32G32B32A32_SNORM, SZ_4, CVT_32_SNORM )
-FETCH_ATTRIB( R32G32B32_SNORM,    SZ_3, CVT_32_SNORM )
-FETCH_ATTRIB( R32G32_SNORM,       SZ_2, CVT_32_SNORM )
-FETCH_ATTRIB( R32_SNORM,          SZ_1, CVT_32_SNORM )
+FETCH_ATTRIB( R32G32B32A32_SNORM, SZ_4, CVT_32_SNORM, 1 )
+FETCH_ATTRIB( R32G32B32_SNORM,    SZ_3, CVT_32_SNORM, 1 )
+FETCH_ATTRIB( R32G32_SNORM,       SZ_2, CVT_32_SNORM, 1 )
+FETCH_ATTRIB( R32_SNORM,          SZ_1, CVT_32_SNORM, 1 )
 
-FETCH_ATTRIB( R16G16B16A16_USCALED, SZ_4, CVT_16_USCALED )
-FETCH_ATTRIB( R16G16B16_USCALED,    SZ_3, CVT_16_USCALED )
-FETCH_ATTRIB( R16G16_USCALED,       SZ_2, CVT_16_USCALED )
-FETCH_ATTRIB( R16_USCALED,          SZ_1, CVT_16_USCALED )
+FETCH_ATTRIB( R16G16B16A16_USCALED, SZ_4, CVT_16_USCALED, 1 )
+FETCH_ATTRIB( R16G16B16_USCALED,    SZ_3, CVT_16_USCALED, 1 )
+FETCH_ATTRIB( R16G16_USCALED,       SZ_2, CVT_16_USCALED, 1 )
+FETCH_ATTRIB( R16_USCALED,          SZ_1, CVT_16_USCALED, 1 )
 
-FETCH_ATTRIB( R16G16B16A16_SSCALED, SZ_4, CVT_16_SSCALED )
-FETCH_ATTRIB( R16G16B16_SSCALED,    SZ_3, CVT_16_SSCALED )
-FETCH_ATTRIB( R16G16_SSCALED,       SZ_2, CVT_16_SSCALED )
-FETCH_ATTRIB( R16_SSCALED,          SZ_1, CVT_16_SSCALED )
+FETCH_ATTRIB( R16G16B16A16_SSCALED, SZ_4, CVT_16_SSCALED, 1 )
+FETCH_ATTRIB( R16G16B16_SSCALED,    SZ_3, CVT_16_SSCALED, 1 )
+FETCH_ATTRIB( R16G16_SSCALED,       SZ_2, CVT_16_SSCALED, 1 )
+FETCH_ATTRIB( R16_SSCALED,          SZ_1, CVT_16_SSCALED, 1 )
 
-FETCH_ATTRIB( R16G16B16A16_UNORM, SZ_4, CVT_16_UNORM )
-FETCH_ATTRIB( R16G16B16_UNORM,    SZ_3, CVT_16_UNORM )
-FETCH_ATTRIB( R16G16_UNORM,       SZ_2, CVT_16_UNORM )
-FETCH_ATTRIB( R16_UNORM,          SZ_1, CVT_16_UNORM )
+FETCH_ATTRIB( R16G16B16A16_UNORM, SZ_4, CVT_16_UNORM, 1 )
+FETCH_ATTRIB( R16G16B16_UNORM,    SZ_3, CVT_16_UNORM, 1 )
+FETCH_ATTRIB( R16G16_UNORM,       SZ_2, CVT_16_UNORM, 1 )
+FETCH_ATTRIB( R16_UNORM,          SZ_1, CVT_16_UNORM, 1 )
 
-FETCH_ATTRIB( R16G16B16A16_SNORM, SZ_4, CVT_16_SNORM )
-FETCH_ATTRIB( R16G16B16_SNORM,    SZ_3, CVT_16_SNORM )
-FETCH_ATTRIB( R16G16_SNORM,       SZ_2, CVT_16_SNORM )
-FETCH_ATTRIB( R16_SNORM,          SZ_1, CVT_16_SNORM )
+FETCH_ATTRIB( R16G16B16A16_SNORM, SZ_4, CVT_16_SNORM, 1 )
+FETCH_ATTRIB( R16G16B16_SNORM,    SZ_3, CVT_16_SNORM, 1 )
+FETCH_ATTRIB( R16G16_SNORM,       SZ_2, CVT_16_SNORM, 1 )
+FETCH_ATTRIB( R16_SNORM,          SZ_1, CVT_16_SNORM, 1 )
 
-FETCH_ATTRIB( R8G8B8A8_USCALED,   SZ_4, CVT_8_USCALED )
-FETCH_ATTRIB( R8G8B8_USCALED,     SZ_3, CVT_8_USCALED )
-FETCH_ATTRIB( R8G8_USCALED,       SZ_2, CVT_8_USCALED )
-FETCH_ATTRIB( R8_USCALED,         SZ_1, CVT_8_USCALED )
+FETCH_ATTRIB( R8G8B8A8_USCALED,   SZ_4, CVT_8_USCALED, 1 )
+FETCH_ATTRIB( R8G8B8_USCALED,     SZ_3, CVT_8_USCALED, 1 )
+FETCH_ATTRIB( R8G8_USCALED,       SZ_2, CVT_8_USCALED, 1 )
+FETCH_ATTRIB( R8_USCALED,         SZ_1, CVT_8_USCALED, 1 )
 
-FETCH_ATTRIB( R8G8B8A8_SSCALED,  SZ_4, CVT_8_SSCALED )
-FETCH_ATTRIB( R8G8B8_SSCALED,    SZ_3, CVT_8_SSCALED )
-FETCH_ATTRIB( R8G8_SSCALED,      SZ_2, CVT_8_SSCALED )
-FETCH_ATTRIB( R8_SSCALED,        SZ_1, CVT_8_SSCALED )
+FETCH_ATTRIB( R8G8B8A8_SSCALED,  SZ_4, CVT_8_SSCALED, 1 )
+FETCH_ATTRIB( R8G8B8_SSCALED,    SZ_3, CVT_8_SSCALED, 1 )
+FETCH_ATTRIB( R8G8_SSCALED,      SZ_2, CVT_8_SSCALED, 1 )
+FETCH_ATTRIB( R8_SSCALED,        SZ_1, CVT_8_SSCALED, 1 )
 
-FETCH_ATTRIB( R8G8B8A8_UNORM,  SZ_4, CVT_8_UNORM )
-FETCH_ATTRIB( R8G8B8_UNORM,    SZ_3, CVT_8_UNORM )
-FETCH_ATTRIB( R8G8_UNORM,      SZ_2, CVT_8_UNORM )
-FETCH_ATTRIB( R8_UNORM,        SZ_1, CVT_8_UNORM )
+FETCH_ATTRIB( R8G8B8A8_UNORM,  SZ_4, CVT_8_UNORM, 1 )
+FETCH_ATTRIB( R8G8B8_UNORM,    SZ_3, CVT_8_UNORM, 1 )
+FETCH_ATTRIB( R8G8_UNORM,      SZ_2, CVT_8_UNORM, 1 )
+FETCH_ATTRIB( R8_UNORM,        SZ_1, CVT_8_UNORM, 1 )
 
-FETCH_ATTRIB( R8G8B8A8_SNORM,  SZ_4, CVT_8_SNORM )
-FETCH_ATTRIB( R8G8B8_SNORM,    SZ_3, CVT_8_SNORM )
-FETCH_ATTRIB( R8G8_SNORM,      SZ_2, CVT_8_SNORM )
-FETCH_ATTRIB( R8_SNORM,        SZ_1, CVT_8_SNORM )
+FETCH_ATTRIB( R8G8B8A8_SNORM,  SZ_4, CVT_8_SNORM, 1 )
+FETCH_ATTRIB( R8G8B8_SNORM,    SZ_3, CVT_8_SNORM, 1 )
+FETCH_ATTRIB( R8G8_SNORM,      SZ_2, CVT_8_SNORM, 1 )
+FETCH_ATTRIB( R8_SNORM,        SZ_1, CVT_8_SNORM, 1 )
 
-FETCH_ATTRIB( A8R8G8B8_UNORM,       SZ_4, CVT_8_UNORM )
+FETCH_ATTRIB( A8R8G8B8_UNORM,       SZ_4, CVT_8_UNORM, 1 )
 
 
 
@@ -584,7 +589,6 @@ static void generic_vertex_fetch(struct spu_vs_context *draw,
       unsigned idx;
       const unsigned bytes_per_entry = draw->vertex_fetch.size[attr];
       const unsigned quads_per_entry = (bytes_per_entry + 15) / 16;
-      qword p[4];
       qword in[2 * 4];
 
 
@@ -609,23 +613,7 @@ static void generic_vertex_fetch(struct spu_vs_context *draw,
 
       /* Convert all 4 vertices to vectors of float.
        */
-      idx = 0;
-      for (i = 0; i < 4; i++) {
-         p[i] = (*fetch)(in + idx);
-	 idx += quads_per_entry;
-      }
-
-
-      /* Transpose/swizzle into vector-friendly format.  Currently
-       * assuming that all vertex shader inputs are float[4], but this
-       * isn't true -- if the vertex shader only wants tex0.xy, we
-       * could optimize for that.
-       *
-       * To do so fully without codegen would probably require an
-       * excessive number of fetch functions, but we could at least
-       * minimize the transpose step:
-       */
-      _transpose_matrix4x4(&machine->Inputs[attr].xyzw[0].q, p);
+      (*fetch)(&machine->Inputs[attr].xyzw[0].q, in);
    }
 }
 
diff --git a/src/mesa/pipe/cell/spu/spu_vertex_shader.h b/src/mesa/pipe/cell/spu/spu_vertex_shader.h
index ea044e841da..8b37a239a47 100644
--- a/src/mesa/pipe/cell/spu/spu_vertex_shader.h
+++ b/src/mesa/pipe/cell/spu/spu_vertex_shader.h
@@ -6,7 +6,7 @@
 
 struct spu_vs_context;
 
-typedef qword (*spu_fetch_func)(const qword *qw);
+typedef void (*spu_fetch_func)(qword *out, const qword *in);
 typedef void (*spu_full_fetch_func)( struct spu_vs_context *draw,
 				     struct spu_exec_machine *machine,
 				     const unsigned *elts,