Initial pass at vertex shader on SPU using TGSI VM

All of the code is wired in on the SPU side, but it is not called from
the PPU yet.  Instruction / declaration fetch still needs to be
implemented in spu_exec.c.

9 files changed, 3540 insertions, 1 deletions

diff --git a/src/mesa/pipe/cell/common.h b/src/mesa/pipe/cell/common.h
index d5e86863d4..80a1425ec7 100644
--- a/src/mesa/pipe/cell/common.h
+++ b/src/mesa/pipe/cell/common.h
@@ -83,6 +83,9 @@
 #define CELL_CMD_STATE_SAMPLER       12
 #define CELL_CMD_STATE_TEXTURE       13
 #define CELL_CMD_STATE_VERTEX_INFO   14
+#define CELL_CMD_STATE_VIEWPORT      15
+#define CELL_CMD_STATE_VS_ARRAY_INFO 16
+#define CELL_CMD_VS_EXECUTE          17
 
 
 #define CELL_NUM_BUFFERS 4
@@ -116,6 +119,41 @@ struct cell_command_clear_surface
 } ALIGN16_ATTRIB;
 
 
+/**
+ * Array info used by the vertex shader's vertex puller.
+ */
+struct cell_array_info
+{
+    void *base;               /**< Base address of the 0th element. */
+    uint attr;                /**< Attribute that this state if for. */
+    uint pitch;               /**< Byte pitch from one entry to the next. */
+    enum pipe_format format;  /**< Pipe format of each entry. */
+} ALIGN16_ATTRIB;
+
+
+struct cell_shader_info
+{
+   unsigned processor;
+   unsigned num_outputs;
+
+   void *declarations;
+   unsigned num_declarations;
+   void *instructions;
+   unsigned num_instructions;
+   void *uniforms;
+} ALIGN16_ATTRIB;
+
+
+struct cell_command_vs
+{
+   struct cell_shader_info   shader;
+   void *elts;
+   unsigned num_elts;
+   unsigned bytes_per_elt;
+   void *vOut;
+} ALIGN16_ATTRIB;
+
+
 struct cell_command_render
 {
    uint opcode;       /**< CELL_CMD_RENDER */
diff --git a/src/mesa/pipe/cell/spu/Makefile b/src/mesa/pipe/cell/spu/Makefile
index d5b30e1f27..2d031bfbc6 100644
--- a/src/mesa/pipe/cell/spu/Makefile
+++ b/src/mesa/pipe/cell/spu/Makefile
@@ -20,7 +20,11 @@ SOURCES = \
 	spu_render.c \
 	spu_texture.c \
 	spu_tile.c \
-	spu_tri.c
+	spu_tri.c \
+	spu_exec.c \
+	spu_util.c \
+	spu_vertex_fetch.c \
+	spu_vertex_shader.c
 
 SPU_OBJECTS = $(SOURCES:.c=.o) \
 
diff --git a/src/mesa/pipe/cell/spu/spu_exec.c b/src/mesa/pipe/cell/spu/spu_exec.c
new file mode 100644
index 0000000000..6888e97caf
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_exec.c
@@ -0,0 +1,2355 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * All Rights Reserved.
+ * 
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ * 
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * 
+ **************************************************************************/
+
+/**
+ * TGSI interpretor/executor.
+ *
+ * Flow control information:
+ *
+ * Since we operate on 'quads' (4 pixels or 4 vertices in parallel)
+ * flow control statements (IF/ELSE/ENDIF, LOOP/ENDLOOP) require special
+ * care since a condition may be true for some quad components but false
+ * for other components.
+ *
+ * We basically execute all statements (even if they're in the part of
+ * an IF/ELSE clause that's "not taken") and use a special mask to
+ * control writing to destination registers.  This is the ExecMask.
+ * See store_dest().
+ *
+ * The ExecMask is computed from three other masks (CondMask, LoopMask and
+ * ContMask) which are controlled by the flow control instructions (namely:
+ * (IF/ELSE/ENDIF, LOOP/ENDLOOP and CONT).
+ *
+ *
+ * Authors:
+ *   Michal Krol
+ *   Brian Paul
+ */
+
+#include "pipe/p_compiler.h"
+#include "pipe/p_state.h"
+#include "pipe/p_util.h"
+#include "pipe/p_shader_tokens.h"
+#include "pipe/tgsi/util/tgsi_parse.h"
+#include "pipe/tgsi/util/tgsi_util.h"
+#include "spu_exec.h"
+
+#define TILE_TOP_LEFT     0
+#define TILE_TOP_RIGHT    1
+#define TILE_BOTTOM_LEFT  2
+#define TILE_BOTTOM_RIGHT 3
+
+/*
+ * Shorthand locations of various utility registers (_I = Index, _C = Channel)
+ */
+#define TEMP_0_I           TGSI_EXEC_TEMP_00000000_I
+#define TEMP_0_C           TGSI_EXEC_TEMP_00000000_C
+#define TEMP_7F_I          TGSI_EXEC_TEMP_7FFFFFFF_I
+#define TEMP_7F_C          TGSI_EXEC_TEMP_7FFFFFFF_C
+#define TEMP_80_I          TGSI_EXEC_TEMP_80000000_I
+#define TEMP_80_C          TGSI_EXEC_TEMP_80000000_C
+#define TEMP_FF_I          TGSI_EXEC_TEMP_FFFFFFFF_I
+#define TEMP_FF_C          TGSI_EXEC_TEMP_FFFFFFFF_C
+#define TEMP_1_I           TGSI_EXEC_TEMP_ONE_I
+#define TEMP_1_C           TGSI_EXEC_TEMP_ONE_C
+#define TEMP_2_I           TGSI_EXEC_TEMP_TWO_I
+#define TEMP_2_C           TGSI_EXEC_TEMP_TWO_C
+#define TEMP_128_I         TGSI_EXEC_TEMP_128_I
+#define TEMP_128_C         TGSI_EXEC_TEMP_128_C
+#define TEMP_M128_I        TGSI_EXEC_TEMP_MINUS_128_I
+#define TEMP_M128_C        TGSI_EXEC_TEMP_MINUS_128_C
+#define TEMP_KILMASK_I     TGSI_EXEC_TEMP_KILMASK_I
+#define TEMP_KILMASK_C     TGSI_EXEC_TEMP_KILMASK_C
+#define TEMP_OUTPUT_I      TGSI_EXEC_TEMP_OUTPUT_I
+#define TEMP_OUTPUT_C      TGSI_EXEC_TEMP_OUTPUT_C
+#define TEMP_PRIMITIVE_I   TGSI_EXEC_TEMP_PRIMITIVE_I
+#define TEMP_PRIMITIVE_C   TGSI_EXEC_TEMP_PRIMITIVE_C
+#define TEMP_R0            TGSI_EXEC_TEMP_R0
+
+#define FOR_EACH_CHANNEL(CHAN)\
+   for (CHAN = 0; CHAN < 4; CHAN++)
+
+#define IS_CHANNEL_ENABLED(INST, CHAN)\
+   ((INST).FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define IS_CHANNEL_ENABLED2(INST, CHAN)\
+   ((INST).FullDstRegisters[1].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define FOR_EACH_ENABLED_CHANNEL(INST, CHAN)\
+   FOR_EACH_CHANNEL( CHAN )\
+      if (IS_CHANNEL_ENABLED( INST, CHAN ))
+
+#define FOR_EACH_ENABLED_CHANNEL2(INST, CHAN)\
+   FOR_EACH_CHANNEL( CHAN )\
+      if (IS_CHANNEL_ENABLED2( INST, CHAN ))
+
+
+/** The execution mask depends on the conditional mask and the loop mask */
+#define UPDATE_EXEC_MASK(MACH) \
+      MACH->ExecMask = MACH->CondMask & MACH->LoopMask & MACH->ContMask & MACH->FuncMask
+
+
+#define CHAN_X  0
+#define CHAN_Y  1
+#define CHAN_Z  2
+#define CHAN_W  3
+
+
+
+/**
+ * Initialize machine state by expanding tokens to full instructions,
+ * allocating temporary storage, setting up constants, etc.
+ * After this, we can call spu_exec_machine_run() many times.
+ */
+void
+spu_exec_machine_init(struct spu_exec_machine *mach,
+                      uint numSamplers,
+                      struct spu_sampler *samplers,
+                      unsigned processor)
+{
+   uint i;
+
+   mach->Samplers = samplers;
+   mach->Processor = processor;
+   mach->Addrs = &mach->Temps[TGSI_EXEC_NUM_TEMPS];
+
+   /* Setup constants. */
+   for( i = 0; i < 4; i++ ) {
+      mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].u[i] = 0x00000000;
+      mach->Temps[TEMP_7F_I].xyzw[TEMP_7F_C].u[i] = 0x7FFFFFFF;
+      mach->Temps[TEMP_80_I].xyzw[TEMP_80_C].u[i] = 0x80000000;
+      mach->Temps[TEMP_FF_I].xyzw[TEMP_FF_C].u[i] = 0xFFFFFFFF;
+      mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].f[i] = 1.0f;
+      mach->Temps[TEMP_2_I].xyzw[TEMP_2_C].f[i] = 2.0f;
+      mach->Temps[TEMP_128_I].xyzw[TEMP_128_C].f[i] = 128.0f;
+      mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C].f[i] = -128.0f;
+   }
+}
+
+
+static void
+micro_abs(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] = (float) fabs( (double) src->f[0] );
+   dst->f[1] = (float) fabs( (double) src->f[1] );
+   dst->f[2] = (float) fabs( (double) src->f[2] );
+   dst->f[3] = (float) fabs( (double) src->f[3] );
+}
+
+static void
+micro_add(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] + src1->f[0];
+   dst->f[1] = src0->f[1] + src1->f[1];
+   dst->f[2] = src0->f[2] + src1->f[2];
+   dst->f[3] = src0->f[3] + src1->f[3];
+}
+
+static void
+micro_iadd(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] + src1->i[0];
+   dst->i[1] = src0->i[1] + src1->i[1];
+   dst->i[2] = src0->i[2] + src1->i[2];
+   dst->i[3] = src0->i[3] + src1->i[3];
+}
+
+static void
+micro_and(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] & src1->u[0];
+   dst->u[1] = src0->u[1] & src1->u[1];
+   dst->u[2] = src0->u[2] & src1->u[2];
+   dst->u[3] = src0->u[3] & src1->u[3];
+}
+
+static void
+micro_ceil(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) ceil( (double) src->f[0] );
+   dst->f[1] = (float) ceil( (double) src->f[1] );
+   dst->f[2] = (float) ceil( (double) src->f[2] );
+   dst->f[3] = (float) ceil( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_cos(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) cos( (double) src->f[0] );
+   dst->f[1] = (float) cos( (double) src->f[1] );
+   dst->f[2] = (float) cos( (double) src->f[2] );
+   dst->f[3] = (float) cos( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_ddx(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] =
+   dst->f[1] =
+   dst->f[2] =
+   dst->f[3] = src->f[TILE_BOTTOM_RIGHT] - src->f[TILE_BOTTOM_LEFT];
+}
+
+static void
+micro_ddy(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] =
+   dst->f[1] =
+   dst->f[2] =
+   dst->f[3] = src->f[TILE_TOP_LEFT] - src->f[TILE_BOTTOM_LEFT];
+}
+
+static void
+micro_div(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] / src1->f[0];
+   dst->f[1] = src0->f[1] / src1->f[1];
+   dst->f[2] = src0->f[2] / src1->f[2];
+   dst->f[3] = src0->f[3] / src1->f[3];
+}
+
+static void
+micro_udiv(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] / src1->u[0];
+   dst->u[1] = src0->u[1] / src1->u[1];
+   dst->u[2] = src0->u[2] / src1->u[2];
+   dst->u[3] = src0->u[3] / src1->u[3];
+}
+
+static void
+micro_eq(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->f[0] = src0->f[0] == src1->f[0] ? src2->f[0] : src3->f[0];
+   dst->f[1] = src0->f[1] == src1->f[1] ? src2->f[1] : src3->f[1];
+   dst->f[2] = src0->f[2] == src1->f[2] ? src2->f[2] : src3->f[2];
+   dst->f[3] = src0->f[3] == src1->f[3] ? src2->f[3] : src3->f[3];
+}
+
+static void
+micro_ieq(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->i[0] = src0->i[0] == src1->i[0] ? src2->i[0] : src3->i[0];
+   dst->i[1] = src0->i[1] == src1->i[1] ? src2->i[1] : src3->i[1];
+   dst->i[2] = src0->i[2] == src1->i[2] ? src2->i[2] : src3->i[2];
+   dst->i[3] = src0->i[3] == src1->i[3] ? src2->i[3] : src3->i[3];
+}
+
+static void
+micro_exp2(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src)
+{
+#if 0
+   dst->f[0] = (float) pow( 2.0, (double) src->f[0] );
+   dst->f[1] = (float) pow( 2.0, (double) src->f[1] );
+   dst->f[2] = (float) pow( 2.0, (double) src->f[2] );
+   dst->f[3] = (float) pow( 2.0, (double) src->f[3] );
+#endif
+}
+
+static void
+micro_f2it(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->i[0] = (int) src->f[0];
+   dst->i[1] = (int) src->f[1];
+   dst->i[2] = (int) src->f[2];
+   dst->i[3] = (int) src->f[3];
+}
+
+static void
+micro_f2ut(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->u[0] = (uint) src->f[0];
+   dst->u[1] = (uint) src->f[1];
+   dst->u[2] = (uint) src->f[2];
+   dst->u[3] = (uint) src->f[3];
+}
+
+static void
+micro_flr(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) floor( (double) src->f[0] );
+   dst->f[1] = (float) floor( (double) src->f[1] );
+   dst->f[2] = (float) floor( (double) src->f[2] );
+   dst->f[3] = (float) floor( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_frc(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = src->f[0] - (float) floor( (double) src->f[0] );
+   dst->f[1] = src->f[1] - (float) floor( (double) src->f[1] );
+   dst->f[2] = src->f[2] - (float) floor( (double) src->f[2] );
+   dst->f[3] = src->f[3] - (float) floor( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_ge(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->f[0] = src0->f[0] >= src1->f[0] ? src2->f[0] : src3->f[0];
+   dst->f[1] = src0->f[1] >= src1->f[1] ? src2->f[1] : src3->f[1];
+   dst->f[2] = src0->f[2] >= src1->f[2] ? src2->f[2] : src3->f[2];
+   dst->f[3] = src0->f[3] >= src1->f[3] ? src2->f[3] : src3->f[3];
+}
+
+static void
+micro_i2f(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] = (float) src->i[0];
+   dst->f[1] = (float) src->i[1];
+   dst->f[2] = (float) src->i[2];
+   dst->f[3] = (float) src->i[3];
+}
+
+static void
+micro_lg2(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) log( (double) src->f[0] ) * 1.442695f;
+   dst->f[1] = (float) log( (double) src->f[1] ) * 1.442695f;
+   dst->f[2] = (float) log( (double) src->f[2] ) * 1.442695f;
+   dst->f[3] = (float) log( (double) src->f[3] ) * 1.442695f;
+#endif
+}
+
+static void
+micro_lt(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->f[0] = src0->f[0] < src1->f[0] ? src2->f[0] : src3->f[0];
+   dst->f[1] = src0->f[1] < src1->f[1] ? src2->f[1] : src3->f[1];
+   dst->f[2] = src0->f[2] < src1->f[2] ? src2->f[2] : src3->f[2];
+   dst->f[3] = src0->f[3] < src1->f[3] ? src2->f[3] : src3->f[3];
+}
+
+static void
+micro_ilt(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->i[0] = src0->i[0] < src1->i[0] ? src2->i[0] : src3->i[0];
+   dst->i[1] = src0->i[1] < src1->i[1] ? src2->i[1] : src3->i[1];
+   dst->i[2] = src0->i[2] < src1->i[2] ? src2->i[2] : src3->i[2];
+   dst->i[3] = src0->i[3] < src1->i[3] ? src2->i[3] : src3->i[3];
+}
+
+static void
+micro_ult(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2,
+   const union spu_exec_channel *src3 )
+{
+   dst->u[0] = src0->u[0] < src1->u[0] ? src2->u[0] : src3->u[0];
+   dst->u[1] = src0->u[1] < src1->u[1] ? src2->u[1] : src3->u[1];
+   dst->u[2] = src0->u[2] < src1->u[2] ? src2->u[2] : src3->u[2];
+   dst->u[3] = src0->u[3] < src1->u[3] ? src2->u[3] : src3->u[3];
+}
+
+static void
+micro_max(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] > src1->f[0] ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] > src1->f[1] ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] > src1->f[2] ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] > src1->f[3] ? src0->f[3] : src1->f[3];
+}
+
+static void
+micro_imax(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] > src1->i[0] ? src0->i[0] : src1->i[0];
+   dst->i[1] = src0->i[1] > src1->i[1] ? src0->i[1] : src1->i[1];
+   dst->i[2] = src0->i[2] > src1->i[2] ? src0->i[2] : src1->i[2];
+   dst->i[3] = src0->i[3] > src1->i[3] ? src0->i[3] : src1->i[3];
+}
+
+static void
+micro_umax(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] > src1->u[0] ? src0->u[0] : src1->u[0];
+   dst->u[1] = src0->u[1] > src1->u[1] ? src0->u[1] : src1->u[1];
+   dst->u[2] = src0->u[2] > src1->u[2] ? src0->u[2] : src1->u[2];
+   dst->u[3] = src0->u[3] > src1->u[3] ? src0->u[3] : src1->u[3];
+}
+
+static void
+micro_min(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] < src1->f[0] ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] < src1->f[1] ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] < src1->f[2] ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] < src1->f[3] ? src0->f[3] : src1->f[3];
+}
+
+static void
+micro_imin(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] < src1->i[0] ? src0->i[0] : src1->i[0];
+   dst->i[1] = src0->i[1] < src1->i[1] ? src0->i[1] : src1->i[1];
+   dst->i[2] = src0->i[2] < src1->i[2] ? src0->i[2] : src1->i[2];
+   dst->i[3] = src0->i[3] < src1->i[3] ? src0->i[3] : src1->i[3];
+}
+
+static void
+micro_umin(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] < src1->u[0] ? src0->u[0] : src1->u[0];
+   dst->u[1] = src0->u[1] < src1->u[1] ? src0->u[1] : src1->u[1];
+   dst->u[2] = src0->u[2] < src1->u[2] ? src0->u[2] : src1->u[2];
+   dst->u[3] = src0->u[3] < src1->u[3] ? src0->u[3] : src1->u[3];
+}
+
+static void
+micro_umod(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] % src1->u[0];
+   dst->u[1] = src0->u[1] % src1->u[1];
+   dst->u[2] = src0->u[2] % src1->u[2];
+   dst->u[3] = src0->u[3] % src1->u[3];
+}
+
+static void
+micro_mul(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] * src1->f[0];
+   dst->f[1] = src0->f[1] * src1->f[1];
+   dst->f[2] = src0->f[2] * src1->f[2];
+   dst->f[3] = src0->f[3] * src1->f[3];
+}
+
+static void
+micro_imul(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] * src1->i[0];
+   dst->i[1] = src0->i[1] * src1->i[1];
+   dst->i[2] = src0->i[2] * src1->i[2];
+   dst->i[3] = src0->i[3] * src1->i[3];
+}
+
+static void
+micro_imul64(
+   union spu_exec_channel *dst0,
+   union spu_exec_channel *dst1,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst1->i[0] = src0->i[0] * src1->i[0];
+   dst1->i[1] = src0->i[1] * src1->i[1];
+   dst1->i[2] = src0->i[2] * src1->i[2];
+   dst1->i[3] = src0->i[3] * src1->i[3];
+   dst0->i[0] = 0;
+   dst0->i[1] = 0;
+   dst0->i[2] = 0;
+   dst0->i[3] = 0;
+}
+
+static void
+micro_umul64(
+   union spu_exec_channel *dst0,
+   union spu_exec_channel *dst1,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst1->u[0] = src0->u[0] * src1->u[0];
+   dst1->u[1] = src0->u[1] * src1->u[1];
+   dst1->u[2] = src0->u[2] * src1->u[2];
+   dst1->u[3] = src0->u[3] * src1->u[3];
+   dst0->u[0] = 0;
+   dst0->u[1] = 0;
+   dst0->u[2] = 0;
+   dst0->u[3] = 0;
+}
+
+static void
+micro_movc(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1,
+   const union spu_exec_channel *src2 )
+{
+   dst->u[0] = src0->u[0] ? src1->u[0] : src2->u[0];
+   dst->u[1] = src0->u[1] ? src1->u[1] : src2->u[1];
+   dst->u[2] = src0->u[2] ? src1->u[2] : src2->u[2];
+   dst->u[3] = src0->u[3] ? src1->u[3] : src2->u[3];
+}
+
+static void
+micro_neg(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] = -src->f[0];
+   dst->f[1] = -src->f[1];
+   dst->f[2] = -src->f[2];
+   dst->f[3] = -src->f[3];
+}
+
+static void
+micro_ineg(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->i[0] = -src->i[0];
+   dst->i[1] = -src->i[1];
+   dst->i[2] = -src->i[2];
+   dst->i[3] = -src->i[3];
+}
+
+static void
+micro_not(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->u[0] = ~src->u[0];
+   dst->u[1] = ~src->u[1];
+   dst->u[2] = ~src->u[2];
+   dst->u[3] = ~src->u[3];
+}
+
+static void
+micro_or(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] | src1->u[0];
+   dst->u[1] = src0->u[1] | src1->u[1];
+   dst->u[2] = src0->u[2] | src1->u[2];
+   dst->u[3] = src0->u[3] | src1->u[3];
+}
+
+static void
+micro_pow(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+#if 0
+   dst->f[0] = (float) pow( (double) src0->f[0], (double) src1->f[0] );
+   dst->f[1] = (float) pow( (double) src0->f[1], (double) src1->f[1] );
+   dst->f[2] = (float) pow( (double) src0->f[2], (double) src1->f[2] );
+   dst->f[3] = (float) pow( (double) src0->f[3], (double) src1->f[3] );
+#endif
+}
+
+static void
+micro_rnd(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) floor( (double) (src->f[0] + 0.5f) );
+   dst->f[1] = (float) floor( (double) (src->f[1] + 0.5f) );
+   dst->f[2] = (float) floor( (double) (src->f[2] + 0.5f) );
+   dst->f[3] = (float) floor( (double) (src->f[3] + 0.5f) );
+#endif
+}
+
+static void
+micro_shl(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] << src1->i[0];
+   dst->i[1] = src0->i[1] << src1->i[1];
+   dst->i[2] = src0->i[2] << src1->i[2];
+   dst->i[3] = src0->i[3] << src1->i[3];
+}
+
+static void
+micro_ishr(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] >> src1->i[0];
+   dst->i[1] = src0->i[1] >> src1->i[1];
+   dst->i[2] = src0->i[2] >> src1->i[2];
+   dst->i[3] = src0->i[3] >> src1->i[3];
+}
+
+static void
+micro_trunc(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0 )
+{
+   dst->f[0] = (float) (int) src0->f[0];
+   dst->f[1] = (float) (int) src0->f[1];
+   dst->f[2] = (float) (int) src0->f[2];
+   dst->f[3] = (float) (int) src0->f[3];
+}
+
+static void
+micro_ushr(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] >> src1->u[0];
+   dst->u[1] = src0->u[1] >> src1->u[1];
+   dst->u[2] = src0->u[2] >> src1->u[2];
+   dst->u[3] = src0->u[3] >> src1->u[3];
+}
+
+static void
+micro_sin(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) sin( (double) src->f[0] );
+   dst->f[1] = (float) sin( (double) src->f[1] );
+   dst->f[2] = (float) sin( (double) src->f[2] );
+   dst->f[3] = (float) sin( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_sqrt( union spu_exec_channel *dst,
+            const union spu_exec_channel *src )
+{
+#if 0
+   dst->f[0] = (float) sqrt( (double) src->f[0] );
+   dst->f[1] = (float) sqrt( (double) src->f[1] );
+   dst->f[2] = (float) sqrt( (double) src->f[2] );
+   dst->f[3] = (float) sqrt( (double) src->f[3] );
+#endif
+}
+
+static void
+micro_sub(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] - src1->f[0];
+   dst->f[1] = src0->f[1] - src1->f[1];
+   dst->f[2] = src0->f[2] - src1->f[2];
+   dst->f[3] = src0->f[3] - src1->f[3];
+}
+
+static void
+micro_u2f(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src )
+{
+   dst->f[0] = (float) src->u[0];
+   dst->f[1] = (float) src->u[1];
+   dst->f[2] = (float) src->u[2];
+   dst->f[3] = (float) src->u[3];
+}
+
+static void
+micro_xor(
+   union spu_exec_channel *dst,
+   const union spu_exec_channel *src0,
+   const union spu_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] ^ src1->u[0];
+   dst->u[1] = src0->u[1] ^ src1->u[1];
+   dst->u[2] = src0->u[2] ^ src1->u[2];
+   dst->u[3] = src0->u[3] ^ src1->u[3];
+}
+
+static void
+fetch_src_file_channel(
+   const struct spu_exec_machine *mach,
+   const uint file,
+   const uint swizzle,
+   const union spu_exec_channel *index,
+   union spu_exec_channel *chan )
+{
+   switch( swizzle ) {
+   case TGSI_EXTSWIZZLE_X:
+   case TGSI_EXTSWIZZLE_Y:
+   case TGSI_EXTSWIZZLE_Z:
+   case TGSI_EXTSWIZZLE_W:
+      switch( file ) {
+      case TGSI_FILE_CONSTANT:
+         chan->f[0] = mach->Consts[index->i[0]][swizzle];
+         chan->f[1] = mach->Consts[index->i[1]][swizzle];
+         chan->f[2] = mach->Consts[index->i[2]][swizzle];
+         chan->f[3] = mach->Consts[index->i[3]][swizzle];
+         break;
+
+      case TGSI_FILE_INPUT:
+         chan->u[0] = mach->Inputs[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Inputs[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Inputs[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Inputs[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      case TGSI_FILE_TEMPORARY:
+         chan->u[0] = mach->Temps[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Temps[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Temps[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Temps[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      case TGSI_FILE_IMMEDIATE:
+         assert( index->i[0] < (int) mach->ImmLimit );
+         assert( index->i[1] < (int) mach->ImmLimit );
+         assert( index->i[2] < (int) mach->ImmLimit );
+         assert( index->i[3] < (int) mach->ImmLimit );
+
+         chan->f[0] = mach->Imms[index->i[0]][swizzle];
+         chan->f[1] = mach->Imms[index->i[1]][swizzle];
+         chan->f[2] = mach->Imms[index->i[2]][swizzle];
+         chan->f[3] = mach->Imms[index->i[3]][swizzle];
+         break;
+
+      case TGSI_FILE_ADDRESS:
+         chan->u[0] = mach->Addrs[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Addrs[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Addrs[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Addrs[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      case TGSI_FILE_OUTPUT:
+         /* vertex/fragment output vars can be read too */
+         chan->u[0] = mach->Outputs[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Outputs[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Outputs[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Outputs[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      default:
+         assert( 0 );
+      }
+      break;
+
+   case TGSI_EXTSWIZZLE_ZERO:
+      *chan = mach->Temps[TEMP_0_I].xyzw[TEMP_0_C];
+      break;
+
+   case TGSI_EXTSWIZZLE_ONE:
+      *chan = mach->Temps[TEMP_1_I].xyzw[TEMP_1_C];
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
+static void
+fetch_source(
+   const struct spu_exec_machine *mach,
+   union spu_exec_channel *chan,
+   const struct tgsi_full_src_register *reg,
+   const uint chan_index )
+{
+   union spu_exec_channel index;
+   uint swizzle;
+
+   index.i[0] =
+   index.i[1] =
+   index.i[2] =
+   index.i[3] = reg->SrcRegister.Index;
+
+   if (reg->SrcRegister.Indirect) {
+      union spu_exec_channel index2;
+      union spu_exec_channel indir_index;
+
+      index2.i[0] =
+      index2.i[1] =
+      index2.i[2] =
+      index2.i[3] = reg->SrcRegisterInd.Index;
+
+      swizzle = tgsi_util_get_src_register_swizzle(&reg->SrcRegisterInd,
+                                                   CHAN_X);
+      fetch_src_file_channel(
+         mach,
+         reg->SrcRegisterInd.File,
+         swizzle,
+         &index2,
+         &indir_index );
+
+      index.i[0] += indir_index.i[0];
+      index.i[1] += indir_index.i[1];
+      index.i[2] += indir_index.i[2];
+      index.i[3] += indir_index.i[3];
+   }
+
+   if( reg->SrcRegister.Dimension ) {
+      switch( reg->SrcRegister.File ) {
+      case TGSI_FILE_INPUT:
+         index.i[0] *= 17;
+         index.i[1] *= 17;
+         index.i[2] *= 17;
+         index.i[3] *= 17;
+         break;
+      case TGSI_FILE_CONSTANT:
+         index.i[0] *= 4096;
+         index.i[1] *= 4096;
+         index.i[2] *= 4096;
+         index.i[3] *= 4096;
+         break;
+      default:
+         assert( 0 );
+      }
+
+      index.i[0] += reg->SrcRegisterDim.Index;
+      index.i[1] += reg->SrcRegisterDim.Index;
+      index.i[2] += reg->SrcRegisterDim.Index;
+      index.i[3] += reg->SrcRegisterDim.Index;
+
+      if (reg->SrcRegisterDim.Indirect) {
+         union spu_exec_channel index2;
+         union spu_exec_channel indir_index;
+
+         index2.i[0] =
+         index2.i[1] =
+         index2.i[2] =
+         index2.i[3] = reg->SrcRegisterDimInd.Index;
+
+         swizzle = tgsi_util_get_src_register_swizzle( &reg->SrcRegisterDimInd, CHAN_X );
+         fetch_src_file_channel(
+            mach,
+            reg->SrcRegisterDimInd.File,
+            swizzle,
+            &index2,
+            &indir_index );
+
+         index.i[0] += indir_index.i[0];
+         index.i[1] += indir_index.i[1];
+         index.i[2] += indir_index.i[2];
+         index.i[3] += indir_index.i[3];
+      }
+   }
+
+   swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+   fetch_src_file_channel(
+      mach,
+      reg->SrcRegister.File,
+      swizzle,
+      &index,
+      chan );
+
+   switch (tgsi_util_get_full_src_register_sign_mode( reg, chan_index )) {
+   case TGSI_UTIL_SIGN_CLEAR:
+      micro_abs( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_SET:
+      micro_abs( chan, chan );
+      micro_neg( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_TOGGLE:
+      micro_neg( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_KEEP:
+      break;
+   }
+
+   if (reg->SrcRegisterExtMod.Complement) {
+      micro_sub( chan, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], chan );
+   }
+}
+
+static void
+store_dest(
+   struct spu_exec_machine *mach,
+   const union spu_exec_channel *chan,
+   const struct tgsi_full_dst_register *reg,
+   const struct tgsi_full_instruction *inst,
+   uint chan_index )
+{
+   union spu_exec_channel *dst;
+
+   switch( reg->DstRegister.File ) {
+   case TGSI_FILE_NULL:
+      return;
+
+   case TGSI_FILE_OUTPUT:
+      dst = &mach->Outputs[mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0]
+                           + reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   case TGSI_FILE_TEMPORARY:
+      dst = &mach->Temps[reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   case TGSI_FILE_ADDRESS:
+      dst = &mach->Addrs[reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   default:
+      assert( 0 );
+      return;
+   }
+
+   switch (inst->Instruction.Saturate)
+   {
+   case TGSI_SAT_NONE:
+      if (mach->ExecMask & 0x1)
+         dst->i[0] = chan->i[0];
+      if (mach->ExecMask & 0x2)
+         dst->i[1] = chan->i[1];
+      if (mach->ExecMask & 0x4)
+         dst->i[2] = chan->i[2];
+      if (mach->ExecMask & 0x8)
+         dst->i[3] = chan->i[3];
+      break;
+
+   case TGSI_SAT_ZERO_ONE:
+      /* XXX need to obey ExecMask here */
+      micro_max(dst, chan, &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C]);
+      micro_min(dst, dst, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C]);
+      break;
+
+   case TGSI_SAT_MINUS_PLUS_ONE:
+      assert( 0 );
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
+#define FETCH(VAL,INDEX,CHAN)\
+    fetch_source (mach, VAL, &inst->FullSrcRegisters[INDEX], CHAN)
+
+#define STORE(VAL,INDEX,CHAN)\
+    store_dest (mach, VAL, &inst->FullDstRegisters[INDEX], inst, CHAN )
+
+
+/**
+ * Execute ARB-style KIL which is predicated by a src register.
+ * Kill fragment if any of the four values is less than zero.
+ */
+static void
+exec_kilp(struct spu_exec_machine *mach,
+          const struct tgsi_full_instruction *inst)
+{
+   uint uniquemask;
+   uint chan_index;
+   uint kilmask = 0; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
+   union spu_exec_channel r[1];
+
+   /* This mask stores component bits that were already tested. Note that
+    * we test if the value is less than zero, so 1.0 and 0.0 need not to be
+    * tested. */
+   uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+
+   for (chan_index = 0; chan_index < 4; chan_index++)
+   {
+      uint swizzle;
+      uint i;
+
+      /* unswizzle channel */
+      swizzle = tgsi_util_get_full_src_register_extswizzle (
+                        &inst->FullSrcRegisters[0],
+                        chan_index);
+
+      /* check if the component has not been already tested */
+      if (uniquemask & (1 << swizzle))
+         continue;
+      uniquemask |= 1 << swizzle;
+
+      FETCH(&r[0], 0, chan_index);
+      for (i = 0; i < 4; i++)
+         if (r[0].f[i] < 0.0f)
+            kilmask |= 1 << i;
+   }
+
+   mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
+}
+
+
+/*
+ * Fetch a texel using STR texture coordinates.
+ */
+static void
+fetch_texel( struct spu_sampler *sampler,
+             const union spu_exec_channel *s,
+             const union spu_exec_channel *t,
+             const union spu_exec_channel *p,
+             float lodbias,  /* XXX should be float[4] */
+             union spu_exec_channel *r,
+             union spu_exec_channel *g,
+             union spu_exec_channel *b,
+             union spu_exec_channel *a )
+{
+   uint j;
+   float rgba[NUM_CHANNELS][QUAD_SIZE];
+
+   sampler->get_samples(sampler, s->f, t->f, p->f, lodbias, rgba);
+
+   for (j = 0; j < 4; j++) {
+      r->f[j] = rgba[0][j];
+      g->f[j] = rgba[1][j];
+      b->f[j] = rgba[2][j];
+      a->f[j] = rgba[3][j];
+   }
+}
+
+
+static void
+exec_tex(struct spu_exec_machine *mach,
+         const struct tgsi_full_instruction *inst,
+         boolean biasLod)
+{
+   const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index;
+   union spu_exec_channel r[8];
+   uint chan_index;
+   float lodBias;
+
+   /*   printf("Sampler %u unit %u\n", sampler, unit); */
+
+   switch (inst->InstructionExtTexture.Texture) {
+   case TGSI_TEXTURE_1D:
+
+      FETCH(&r[0], 0, CHAN_X);
+
+      switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide) {
+      case TGSI_EXTSWIZZLE_W:
+         FETCH(&r[1], 0, CHAN_W);
+         micro_div( &r[0], &r[0], &r[1] );
+         break;
+
+      case TGSI_EXTSWIZZLE_ONE:
+         break;
+
+      default:
+         assert (0);
+      }
+
+      if (biasLod) {
+         FETCH(&r[1], 0, CHAN_W);
+         lodBias = r[2].f[0];
+      }
+      else
+         lodBias = 0.0;
+
+      fetch_texel(&mach->Samplers[unit],
+                  &r[0], NULL, NULL, lodBias,  /* S, T, P, BIAS */
+                  &r[0], &r[1], &r[2], &r[3]); /* R, G, B, A */
+      break;
+
+   case TGSI_TEXTURE_2D:
+   case TGSI_TEXTURE_RECT:
+
+      FETCH(&r[0], 0, CHAN_X);
+      FETCH(&r[1], 0, CHAN_Y);
+      FETCH(&r[2], 0, CHAN_Z);
+
+      switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide) {
+      case TGSI_EXTSWIZZLE_W:
+         FETCH(&r[3], 0, CHAN_W);
+         micro_div( &r[0], &r[0], &r[3] );
+         micro_div( &r[1], &r[1], &r[3] );
+         micro_div( &r[2], &r[2], &r[3] );
+         break;
+
+      case TGSI_EXTSWIZZLE_ONE:
+         break;
+
+      default:
+         assert (0);
+      }
+
+      if (biasLod) {
+         FETCH(&r[3], 0, CHAN_W);
+         lodBias = r[3].f[0];
+      }
+      else
+         lodBias = 0.0;
+
+      fetch_texel(&mach->Samplers[unit],
+                  &r[0], &r[1], &r[2], lodBias,  /* inputs */
+                  &r[0], &r[1], &r[2], &r[3]);  /* outputs */
+      break;
+
+   case TGSI_TEXTURE_3D:
+   case TGSI_TEXTURE_CUBE:
+
+      FETCH(&r[0], 0, CHAN_X);
+      FETCH(&r[1], 0, CHAN_Y);
+      FETCH(&r[2], 0, CHAN_Z);
+
+      switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide) {
+      case TGSI_EXTSWIZZLE_W:
+         FETCH(&r[3], 0, CHAN_W);
+         micro_div( &r[0], &r[0], &r[3] );
+         micro_div( &r[1], &r[1], &r[3] );
+         micro_div( &r[2], &r[2], &r[3] );
+         break;
+
+      case TGSI_EXTSWIZZLE_ONE:
+         break;
+
+      default:
+         assert (0);
+      }
+
+      if (biasLod) {
+         FETCH(&r[3], 0, CHAN_W);
+         lodBias = r[3].f[0];
+      }
+      else
+         lodBias = 0.0;
+
+      fetch_texel(&mach->Samplers[unit],
+                  &r[0], &r[1], &r[2], lodBias,
+                  &r[0], &r[1], &r[2], &r[3]);
+      break;
+
+   default:
+      assert (0);
+   }
+
+   FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+      STORE( &r[chan_index], 0, chan_index );
+   }
+}
+
+
+
+static void
+constant_interpolation(
+   struct spu_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   unsigned i;
+
+   for( i = 0; i < QUAD_SIZE; i++ ) {
+      mach->Inputs[attrib].xyzw[chan].f[i] = mach->InterpCoefs[attrib].a0[chan];
+   }
+}
+
+static void
+linear_interpolation(
+   struct spu_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   const float x = mach->QuadPos.xyzw[0].f[0];
+   const float y = mach->QuadPos.xyzw[1].f[0];
+   const float dadx = mach->InterpCoefs[attrib].dadx[chan];
+   const float dady = mach->InterpCoefs[attrib].dady[chan];
+   const float a0 = mach->InterpCoefs[attrib].a0[chan] + dadx * x + dady * y;
+   mach->Inputs[attrib].xyzw[chan].f[0] = a0;
+   mach->Inputs[attrib].xyzw[chan].f[1] = a0 + dadx;
+   mach->Inputs[attrib].xyzw[chan].f[2] = a0 + dady;
+   mach->Inputs[attrib].xyzw[chan].f[3] = a0 + dadx + dady;
+}
+
+static void
+perspective_interpolation(
+   struct spu_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   const float x = mach->QuadPos.xyzw[0].f[0];
+   const float y = mach->QuadPos.xyzw[1].f[0];
+   const float dadx = mach->InterpCoefs[attrib].dadx[chan];
+   const float dady = mach->InterpCoefs[attrib].dady[chan];
+   const float a0 = mach->InterpCoefs[attrib].a0[chan] + dadx * x + dady * y;
+   const float *w = mach->QuadPos.xyzw[3].f;
+   /* divide by W here */
+   mach->Inputs[attrib].xyzw[chan].f[0] = a0 / w[0];
+   mach->Inputs[attrib].xyzw[chan].f[1] = (a0 + dadx) / w[1];
+   mach->Inputs[attrib].xyzw[chan].f[2] = (a0 + dady) / w[2];
+   mach->Inputs[attrib].xyzw[chan].f[3] = (a0 + dadx + dady) / w[3];
+}
+
+
+typedef void (* interpolation_func)(
+   struct spu_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan );
+
+static void
+exec_declaration(struct spu_exec_machine *mach,
+                 const struct tgsi_full_declaration *decl)
+{
+   if( mach->Processor == TGSI_PROCESSOR_FRAGMENT ) {
+      if( decl->Declaration.File == TGSI_FILE_INPUT ) {
+         unsigned first, last, mask;
+         interpolation_func interp;
+
+         assert( decl->Declaration.Declare == TGSI_DECLARE_RANGE );
+
+         first = decl->u.DeclarationRange.First;
+         last = decl->u.DeclarationRange.Last;
+         mask = decl->Declaration.UsageMask;
+
+         switch( decl->Interpolation.Interpolate ) {
+         case TGSI_INTERPOLATE_CONSTANT:
+            interp = constant_interpolation;
+            break;
+
+         case TGSI_INTERPOLATE_LINEAR:
+            interp = linear_interpolation;
+            break;
+
+         case TGSI_INTERPOLATE_PERSPECTIVE:
+            interp = perspective_interpolation;
+            break;
+
+         default:
+            assert( 0 );
+         }
+
+         if( mask == TGSI_WRITEMASK_XYZW ) {
+            unsigned i, j;
+
+            for( i = first; i <= last; i++ ) {
+               for( j = 0; j < NUM_CHANNELS; j++ ) {
+                  interp( mach, i, j );
+               }
+            }
+         }
+         else {
+            unsigned i, j;
+
+            for( j = 0; j < NUM_CHANNELS; j++ ) {
+               if( mask & (1 << j) ) {
+                  for( i = first; i <= last; i++ ) {
+                     interp( mach, i, j );
+                  }
+               }
+            }
+         }
+      }
+   }
+}
+
+static void
+exec_instruction(
+   struct spu_exec_machine *mach,
+   const struct tgsi_full_instruction *inst,
+   int *pc )
+{
+   uint chan_index;
+   union spu_exec_channel r[8];
+
+   (*pc)++;
+
+   switch (inst->Instruction.Opcode) {
+   case TGSI_OPCODE_ARL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 FETCH( &r[0], 0, chan_index );
+	 micro_f2it( &r[0], &r[0] );
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MOV:
+   /* TGSI_OPCODE_SWZ */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LIT:
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y ) || IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	 FETCH( &r[0], 0, CHAN_X );
+	 if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+	    micro_max( &r[0], &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+	    STORE( &r[0], 0, CHAN_Y );
+	 }
+
+	 if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	    FETCH( &r[1], 0, CHAN_Y );
+	    micro_max( &r[1], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+
+	    FETCH( &r[2], 0, CHAN_W );
+	    micro_min( &r[2], &r[2], &mach->Temps[TEMP_128_I].xyzw[TEMP_128_C] );
+	    micro_max( &r[2], &r[2], &mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C] );
+	    micro_pow( &r[1], &r[1], &r[2] );
+	    micro_lt( &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+	    STORE( &r[0], 0, CHAN_Z );
+	 }
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_RCP:
+   /* TGSI_OPCODE_RECIP */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_div( &r[0], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_RSQ:
+   /* TGSI_OPCODE_RECIPSQRT */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_sqrt( &r[0], &r[0] );
+      micro_div( &r[0], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_EXP:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_LOG:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_MUL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+      {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+
+         micro_mul( &r[0], &r[0], &r[1] );
+
+         STORE(&r[0], 0, chan_index);
+      }
+      break;
+
+   case TGSI_OPCODE_ADD:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_add( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DP3:
+   /* TGSI_OPCODE_DOT3 */
+      FETCH( &r[0], 0, CHAN_X );
+      FETCH( &r[1], 1, CHAN_X );
+      micro_mul( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Y );
+      FETCH( &r[2], 1, CHAN_Y );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Z );
+      FETCH( &r[2], 1, CHAN_Z );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_DP4:
+    /* TGSI_OPCODE_DOT4 */
+       FETCH(&r[0], 0, CHAN_X);
+       FETCH(&r[1], 1, CHAN_X);
+
+       micro_mul( &r[0], &r[0], &r[1] );
+
+       FETCH(&r[1], 0, CHAN_Y);
+       FETCH(&r[2], 1, CHAN_Y);
+
+       micro_mul( &r[1], &r[1], &r[2] );
+       micro_add( &r[0], &r[0], &r[1] );
+
+       FETCH(&r[1], 0, CHAN_Z);
+       FETCH(&r[2], 1, CHAN_Z);
+
+       micro_mul( &r[1], &r[1], &r[2] );
+       micro_add( &r[0], &r[0], &r[1] );
+
+       FETCH(&r[1], 0, CHAN_W);
+       FETCH(&r[2], 1, CHAN_W);
+
+       micro_mul( &r[1], &r[1], &r[2] );
+       micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DST:
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+	 FETCH( &r[0], 0, CHAN_Y );
+	 FETCH( &r[1], 1, CHAN_Y);
+	 micro_mul( &r[0], &r[0], &r[1] );
+	 STORE( &r[0], 0, CHAN_Y );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	 FETCH( &r[0], 0, CHAN_Z );
+	 STORE( &r[0], 0, CHAN_Z );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+	 FETCH( &r[0], 1, CHAN_W );
+	 STORE( &r[0], 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_MIN:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+
+         /* XXX use micro_min()?? */
+         micro_lt( &r[0], &r[0], &r[1], &r[0], &r[1] );
+
+         STORE(&r[0], 0, chan_index);
+      }
+      break;
+
+   case TGSI_OPCODE_MAX:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+
+         /* XXX use micro_max()?? */
+         micro_lt( &r[0], &r[0], &r[1], &r[1], &r[0] );
+
+         STORE(&r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SLT:
+   /* TGSI_OPCODE_SETLT */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_lt( &r[0], &r[0], &r[1], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SGE:
+   /* TGSI_OPCODE_SETGE */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_ge( &r[0], &r[0], &r[1], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MAD:
+   /* TGSI_OPCODE_MADD */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_mul( &r[0], &r[0], &r[1] );
+         FETCH( &r[1], 2, chan_index );
+         micro_add( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SUB:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+
+         micro_sub( &r[0], &r[0], &r[1] );
+
+         STORE(&r[0], 0, chan_index);
+      }
+      break;
+
+   case TGSI_OPCODE_LERP:
+   /* TGSI_OPCODE_LRP */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+         FETCH(&r[2], 2, chan_index);
+
+         micro_sub( &r[1], &r[1], &r[2] );
+         micro_mul( &r[0], &r[0], &r[1] );
+         micro_add( &r[0], &r[0], &r[2] );
+
+         STORE(&r[0], 0, chan_index);
+      }
+      break;
+
+   case TGSI_OPCODE_CND:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_CND0:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_DOT2ADD:
+      /* TGSI_OPCODE_DP2A */
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_INDEX:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_NEGATE:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_FRAC:
+   /* TGSI_OPCODE_FRC */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_frc( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_CLAMP:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_FLOOR:
+   /* TGSI_OPCODE_FLR */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_flr( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_ROUND:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_rnd( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_EXPBASE2:
+    /* TGSI_OPCODE_EX2 */
+      FETCH(&r[0], 0, CHAN_X);
+
+      micro_pow( &r[0], &mach->Temps[TEMP_2_I].xyzw[TEMP_2_C], &r[0] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LOGBASE2:
+   /* TGSI_OPCODE_LG2 */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_lg2( &r[0], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_POWER:
+      /* TGSI_OPCODE_POW */
+      FETCH(&r[0], 0, CHAN_X);
+      FETCH(&r[1], 1, CHAN_X);
+
+      micro_pow( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_CROSSPRODUCT:
+      /* TGSI_OPCODE_XPD */
+      FETCH(&r[0], 0, CHAN_Y);
+      FETCH(&r[1], 1, CHAN_Z);
+
+      micro_mul( &r[2], &r[0], &r[1] );
+
+      FETCH(&r[3], 0, CHAN_Z);
+      FETCH(&r[4], 1, CHAN_Y);
+
+      micro_mul( &r[5], &r[3], &r[4] );
+      micro_sub( &r[2], &r[2], &r[5] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+         STORE( &r[2], 0, CHAN_X );
+      }
+
+      FETCH(&r[2], 1, CHAN_X);
+
+      micro_mul( &r[3], &r[3], &r[2] );
+
+      FETCH(&r[5], 0, CHAN_X);
+
+      micro_mul( &r[1], &r[1], &r[5] );
+      micro_sub( &r[3], &r[3], &r[1] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+         STORE( &r[3], 0, CHAN_Y );
+      }
+
+      micro_mul( &r[5], &r[5], &r[4] );
+      micro_mul( &r[0], &r[0], &r[2] );
+      micro_sub( &r[5], &r[5], &r[0] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+         STORE( &r[5], 0, CHAN_Z );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+         STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+      break;
+
+    case TGSI_OPCODE_MULTIPLYMATRIX:
+       assert (0);
+       break;
+
+    case TGSI_OPCODE_ABS:
+       FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+          FETCH(&r[0], 0, chan_index);
+
+          micro_abs( &r[0], &r[0] );
+
+          STORE(&r[0], 0, chan_index);
+       }
+       break;
+
+   case TGSI_OPCODE_RCC:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_DPH:
+      FETCH(&r[0], 0, CHAN_X);
+      FETCH(&r[1], 1, CHAN_X);
+
+      micro_mul( &r[0], &r[0], &r[1] );
+
+      FETCH(&r[1], 0, CHAN_Y);
+      FETCH(&r[2], 1, CHAN_Y);
+
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FETCH(&r[1], 0, CHAN_Z);
+      FETCH(&r[2], 1, CHAN_Z);
+
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FETCH(&r[1], 1, CHAN_W);
+
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_COS:
+      FETCH(&r[0], 0, CHAN_X);
+
+      micro_cos( &r[0], &r[0] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DDX:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_ddx( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DDY:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_ddy( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_KILP:
+      exec_kilp (mach, inst);
+      break;
+
+   case TGSI_OPCODE_KIL:
+      /* for enabled ExecMask bits, set the killed bit */
+      mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= mach->ExecMask;
+      break;
+
+   case TGSI_OPCODE_PK2H:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_PK2US:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_PK4B:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_PK4UB:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_RFL:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_SEQ:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_eq( &r[0], &r[0], &r[1],
+                   &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C],
+                   &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SFL:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_SGT:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_lt( &r[0], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SIN:
+      FETCH( &r[0], 0, CHAN_X );
+      micro_sin( &r[0], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SLE:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_ge( &r[0], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SNE:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_eq( &r[0], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_STR:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_TEX:
+      /* simple texture lookup */
+      /* src[0] = texcoord */
+      /* src[1] = sampler unit */
+      exec_tex(mach, inst, FALSE);
+      break;
+
+   case TGSI_OPCODE_TXB:
+      /* Texture lookup with lod bias */
+      /* src[0] = texcoord (src[0].w = load bias) */
+      /* src[1] = sampler unit */
+      exec_tex(mach, inst, TRUE);
+      break;
+
+   case TGSI_OPCODE_TXD:
+      /* Texture lookup with explict partial derivatives */
+      /* src[0] = texcoord */
+      /* src[1] = d[strq]/dx */
+      /* src[2] = d[strq]/dy */
+      /* src[3] = sampler unit */
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_TXL:
+      /* Texture lookup with explit LOD */
+      /* src[0] = texcoord (src[0].w = load bias) */
+      /* src[1] = sampler unit */
+      exec_tex(mach, inst, TRUE);
+      break;
+
+   case TGSI_OPCODE_UP2H:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_UP2US:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_UP4B:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_UP4UB:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_X2D:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_ARA:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_ARR:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_BRA:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_CAL:
+      /* skip the call if no execution channels are enabled */
+      if (mach->ExecMask) {
+         /* do the call */
+
+         /* push the Cond, Loop, Cont stacks */
+         assert(mach->CondStackTop < TGSI_EXEC_MAX_COND_NESTING);
+         mach->CondStack[mach->CondStackTop++] = mach->CondMask;
+         assert(mach->LoopStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+         mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
+         assert(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+         mach->ContStack[mach->ContStackTop++] = mach->ContMask;
+
+         assert(mach->FuncStackTop < TGSI_EXEC_MAX_CALL_NESTING);
+         mach->FuncStack[mach->FuncStackTop++] = mach->FuncMask;
+
+         /* note that PC was already incremented above */
+         mach->CallStack[mach->CallStackTop++] = *pc;
+         *pc = inst->InstructionExtLabel.Label;
+      }
+      break;
+
+   case TGSI_OPCODE_RET:
+      mach->FuncMask &= ~mach->ExecMask;
+      UPDATE_EXEC_MASK(mach);
+
+      if (mach->ExecMask == 0x0) {
+         /* really return now (otherwise, keep executing */
+
+         if (mach->CallStackTop == 0) {
+            /* returning from main() */
+            *pc = -1;
+            return;
+         }
+         *pc = mach->CallStack[--mach->CallStackTop];
+
+         /* pop the Cond, Loop, Cont stacks */
+         assert(mach->CondStackTop > 0);
+         mach->CondMask = mach->CondStack[--mach->CondStackTop];
+         assert(mach->LoopStackTop > 0);
+         mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
+         assert(mach->ContStackTop > 0);
+         mach->ContMask = mach->ContStack[--mach->ContStackTop];
+         assert(mach->FuncStackTop > 0);
+         mach->FuncMask = mach->FuncStack[--mach->FuncStackTop];
+
+         UPDATE_EXEC_MASK(mach);
+      }
+      break;
+
+   case TGSI_OPCODE_SSG:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_CMP:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH(&r[0], 0, chan_index);
+         FETCH(&r[1], 1, chan_index);
+         FETCH(&r[2], 2, chan_index);
+
+         micro_lt( &r[0], &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &r[1], &r[2] );
+
+         STORE(&r[0], 0, chan_index);
+      }
+      break;
+
+   case TGSI_OPCODE_SCS:
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) || IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+         FETCH( &r[0], 0, CHAN_X );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
+         micro_cos( &r[1], &r[0] );
+         STORE( &r[1], 0, CHAN_X );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+         micro_sin( &r[1], &r[0] );
+         STORE( &r[1], 0, CHAN_Y );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+         STORE( &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], 0, CHAN_Z );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
+         STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_NRM:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_DIV:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DP2:
+      FETCH( &r[0], 0, CHAN_X );
+      FETCH( &r[1], 1, CHAN_X );
+      micro_mul( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Y );
+      FETCH( &r[2], 1, CHAN_Y );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_IF:
+      /* push CondMask */
+      assert(mach->CondStackTop < TGSI_EXEC_MAX_COND_NESTING);
+      mach->CondStack[mach->CondStackTop++] = mach->CondMask;
+      FETCH( &r[0], 0, CHAN_X );
+      /* update CondMask */
+      if( ! r[0].u[0] ) {
+         mach->CondMask &= ~0x1;
+      }
+      if( ! r[0].u[1] ) {
+         mach->CondMask &= ~0x2;
+      }
+      if( ! r[0].u[2] ) {
+         mach->CondMask &= ~0x4;
+      }
+      if( ! r[0].u[3] ) {
+         mach->CondMask &= ~0x8;
+      }
+      UPDATE_EXEC_MASK(mach);
+      /* Todo: If CondMask==0, jump to ELSE */
+      break;
+
+   case TGSI_OPCODE_ELSE:
+      /* invert CondMask wrt previous mask */
+      {
+         uint prevMask;
+         assert(mach->CondStackTop > 0);
+         prevMask = mach->CondStack[mach->CondStackTop - 1];
+         mach->CondMask = ~mach->CondMask & prevMask;
+         UPDATE_EXEC_MASK(mach);
+         /* Todo: If CondMask==0, jump to ENDIF */
+      }
+      break;
+
+   case TGSI_OPCODE_ENDIF:
+      /* pop CondMask */
+      assert(mach->CondStackTop > 0);
+      mach->CondMask = mach->CondStack[--mach->CondStackTop];
+      UPDATE_EXEC_MASK(mach);
+      break;
+
+   case TGSI_OPCODE_END:
+      /* halt execution */
+      *pc = -1;
+      break;
+
+   case TGSI_OPCODE_REP:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_ENDREP:
+       assert (0);
+       break;
+
+   case TGSI_OPCODE_PUSHA:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_POPA:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_CEIL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_ceil( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_I2F:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_i2f( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_NOT:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_not( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_TRUNC:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_trunc( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SHL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_shl( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SHR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_ishr( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_AND:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_and( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_OR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_or( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MOD:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_XOR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_xor( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SAD:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_TXF:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_TXQ:
+      assert (0);
+      break;
+
+   case TGSI_OPCODE_EMIT:
+      mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] += 16;
+      mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]]++;
+      break;
+
+   case TGSI_OPCODE_ENDPRIM:
+      mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]++;
+      mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]] = 0;
+      break;
+
+   case TGSI_OPCODE_LOOP:
+      /* fall-through (for now) */
+   case TGSI_OPCODE_BGNLOOP2:
+      /* push LoopMask and ContMasks */
+      assert(mach->LoopStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+      mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
+      assert(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+      mach->ContStack[mach->ContStackTop++] = mach->ContMask;
+      break;
+
+   case TGSI_OPCODE_ENDLOOP:
+      /* fall-through (for now at least) */
+   case TGSI_OPCODE_ENDLOOP2:
+      /* Restore ContMask, but don't pop */
+      assert(mach->ContStackTop > 0);
+      mach->ContMask = mach->ContStack[mach->ContStackTop - 1];
+      if (mach->LoopMask) {
+         /* repeat loop: jump to instruction just past BGNLOOP */
+         *pc = inst->InstructionExtLabel.Label + 1;
+      }
+      else {
+         /* exit loop: pop LoopMask */
+         assert(mach->LoopStackTop > 0);
+         mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
+         /* pop ContMask */
+         assert(mach->ContStackTop > 0);
+         mach->ContMask = mach->ContStack[--mach->ContStackTop];
+      }
+      UPDATE_EXEC_MASK(mach);
+      break;
+
+   case TGSI_OPCODE_BRK:
+      /* turn off loop channels for each enabled exec channel */
+      mach->LoopMask &= ~mach->ExecMask;
+      /* Todo: if mach->LoopMask == 0, jump to end of loop */
+      UPDATE_EXEC_MASK(mach);
+      break;
+
+   case TGSI_OPCODE_CONT:
+      /* turn off cont channels for each enabled exec channel */
+      mach->ContMask &= ~mach->ExecMask;
+      /* Todo: if mach->LoopMask == 0, jump to end of loop */
+      UPDATE_EXEC_MASK(mach);
+      break;
+
+   case TGSI_OPCODE_BGNSUB:
+      /* no-op */
+      break;
+
+   case TGSI_OPCODE_ENDSUB:
+      /* no-op */
+      break;
+
+   case TGSI_OPCODE_NOISE1:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NOISE2:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NOISE3:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NOISE4:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NOP:
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
+
+/**
+ * Run TGSI interpreter.
+ * \return bitmask of "alive" quad components
+ */
+uint
+spu_exec_machine_run( struct spu_exec_machine *mach )
+{
+   uint i;
+   int pc = 0;
+
+   mach->CondMask = 0xf;
+   mach->LoopMask = 0xf;
+   mach->ContMask = 0xf;
+   mach->FuncMask = 0xf;
+   mach->ExecMask = 0xf;
+
+   mach->CondStackTop = 0; /* temporarily subvert this assertion */
+   assert(mach->CondStackTop == 0);
+   assert(mach->LoopStackTop == 0);
+   assert(mach->ContStackTop == 0);
+   assert(mach->CallStackTop == 0);
+
+   mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] = 0;
+   mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] = 0;
+
+   if( mach->Processor == TGSI_PROCESSOR_GEOMETRY ) {
+      mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0] = 0;
+      mach->Primitives[0] = 0;
+   }
+
+
+   /* execute declarations (interpolants) */
+   for (i = 0; i < mach->NumDeclarations; i++) {
+      exec_declaration( mach, mach->Declarations+i );
+   }
+
+   /* execute instructions, until pc is set to -1 */
+   while (pc != -1) {
+      assert(pc < mach->NumInstructions);
+      exec_instruction( mach, mach->Instructions + pc, &pc );
+   }
+
+#if 0
+   /* we scale from floats in [0,1] to Zbuffer ints in sp_quad_depth_test.c */
+   if (mach->Processor == TGSI_PROCESSOR_FRAGMENT) {
+      /*
+       * Scale back depth component.
+       */
+      for (i = 0; i < 4; i++)
+         mach->Outputs[0].xyzw[2].f[i] *= ctx->DrawBuffer->_DepthMaxF;
+   }
+#endif
+
+   return ~mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0];
+}
+
+
diff --git a/src/mesa/pipe/cell/spu/spu_exec.h b/src/mesa/pipe/cell/spu/spu_exec.h
new file mode 100644
index 0000000000..89e422ba48
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_exec.h
@@ -0,0 +1,171 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * All Rights Reserved.
+ * 
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ * 
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * 
+ **************************************************************************/
+
+#if !defined SPU_EXEC_H
+#define SPU_EXEC_H
+
+#include "pipe/p_compiler.h"
+#include "pipe/tgsi/exec/tgsi_exec.h"
+
+#if defined __cplusplus
+extern "C" {
+#endif
+
+/**
+  * Registers may be treated as float, signed int or unsigned int.
+  */
+union spu_exec_channel
+{
+   float    f[QUAD_SIZE];
+   int      i[QUAD_SIZE];
+   unsigned u[QUAD_SIZE];
+};
+
+/**
+  * A vector[RGBA] of channels[4 pixels]
+  */
+struct spu_exec_vector
+{
+   union spu_exec_channel xyzw[NUM_CHANNELS];
+};
+
+/**
+ * For fragment programs, information for computing fragment input
+ * values from plane equation of the triangle/line.
+ */
+struct spu_interp_coef
+{
+   float a0[NUM_CHANNELS];	/* in an xyzw layout */
+   float dadx[NUM_CHANNELS];
+   float dady[NUM_CHANNELS];
+};
+
+
+struct softpipe_tile_cache;  /**< Opaque to TGSI */
+
+/**
+ * Information for sampling textures, which must be implemented
+ * by code outside the TGSI executor.
+ */
+struct spu_sampler
+{
+   const struct pipe_sampler_state *state;
+   struct pipe_texture *texture;
+   /** Get samples for four fragments in a quad */
+   void (*get_samples)(struct spu_sampler *sampler,
+                       const float s[QUAD_SIZE],
+                       const float t[QUAD_SIZE],
+                       const float p[QUAD_SIZE],
+                       float lodbias,
+                       float rgba[NUM_CHANNELS][QUAD_SIZE]);
+   void *pipe; /*XXX temporary*/
+   struct softpipe_tile_cache *cache;
+};
+
+
+/**
+ * Run-time virtual machine state for executing TGSI shader.
+ */
+struct spu_exec_machine
+{
+   /*
+    * 32 program temporaries
+    * 4  internal temporaries
+    * 1  address
+    */
+   struct spu_exec_vector       Temps[TGSI_EXEC_NUM_TEMPS 
+				      + TGSI_EXEC_NUM_ADDRS + 1]
+       ALIGN16_ATTRIB;
+
+   struct spu_exec_vector       *Addrs;
+
+   struct spu_sampler           *Samplers;
+
+   float                         Imms[TGSI_EXEC_NUM_IMMEDIATES][4];
+   unsigned                      ImmLimit;
+   float                         (*Consts)[4];
+   struct spu_exec_vector       *Inputs;
+   struct spu_exec_vector       *Outputs;
+   unsigned                      Processor;
+
+   /* GEOMETRY processor only. */
+   unsigned                      *Primitives;
+
+   /* FRAGMENT processor only. */
+   const struct spu_interp_coef *InterpCoefs;
+   struct spu_exec_vector       QuadPos;
+
+   /* Conditional execution masks */
+   uint CondMask;  /**< For IF/ELSE/ENDIF */
+   uint LoopMask;  /**< For BGNLOOP/ENDLOOP */
+   uint ContMask;  /**< For loop CONT statements */
+   uint FuncMask;  /**< For function calls */
+   uint ExecMask;  /**< = CondMask & LoopMask */
+
+   /** Condition mask stack (for nested conditionals) */
+   uint CondStack[TGSI_EXEC_MAX_COND_NESTING];
+   int CondStackTop;
+
+   /** Loop mask stack (for nested loops) */
+   uint LoopStack[TGSI_EXEC_MAX_LOOP_NESTING];
+   int LoopStackTop;
+
+   /** Loop continue mask stack (see comments in tgsi_exec.c) */
+   uint ContStack[TGSI_EXEC_MAX_LOOP_NESTING];
+   int ContStackTop;
+
+   /** Function execution mask stack (for executing subroutine code) */
+   uint FuncStack[TGSI_EXEC_MAX_CALL_NESTING];
+   int FuncStackTop;
+
+   /** Function call stack for saving/restoring the program counter */
+   uint CallStack[TGSI_EXEC_MAX_CALL_NESTING];
+   int CallStackTop;
+
+   struct tgsi_full_instruction *Instructions;
+   uint NumInstructions;
+
+   struct tgsi_full_declaration *Declarations;
+   uint NumDeclarations;
+};
+
+
+extern void
+spu_exec_machine_init(struct spu_exec_machine *mach,
+                      uint numSamplers,
+                      struct spu_sampler *samplers,
+                      unsigned processor);
+
+extern uint
+spu_exec_machine_run( struct spu_exec_machine *mach );
+
+
+#if defined __cplusplus
+} /* extern "C" */
+#endif
+
+#endif /* SPU_EXEC_H */
diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
index 6886f283be..9daa3ec735 100644
--- a/src/mesa/pipe/cell/spu/spu_main.c
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -36,6 +36,7 @@
 #include "spu_render.h"
 #include "spu_texture.h"
 #include "spu_tile.h"
+#include "spu_vertex_shader.h"
 #include "pipe/cell/common.h"
 #include "pipe/p_defines.h"
 
@@ -50,6 +51,7 @@ boolean Debug = FALSE;
 
 struct spu_global spu;
 
+struct spu_vs_context draw;
 
 /**
  * Tell the PPU that this SPU has finished copying a buffer to
@@ -264,6 +266,18 @@ cmd_state_vertex_info(const struct vertex_info *vinfo)
 }
 
 
+static void
+cmd_state_vs_array_info(const struct cell_array_info *vs_info)
+{
+   const unsigned attr = vs_info->attr;
+
+   ASSERT(attr < PIPE_ATTRIB_MAX);
+   draw.vertex_fetch.src_ptr[attr] = vs_info->base;
+   draw.vertex_fetch.pitch[attr] = vs_info->pitch;
+   draw.vertex_fetch.format[attr] = vs_info->format;
+   draw.vertex_fetch.dirty = 1;
+}
+
 
 static void
 cmd_finish(void)
@@ -374,6 +388,20 @@ cmd_batch(uint opcode)
          cmd_state_vertex_info((struct vertex_info *) &buffer[pos+1]);
          pos += (1 + sizeof(struct vertex_info) / 4);
          break;
+      case CELL_CMD_STATE_VIEWPORT:
+         (void) memcpy(& draw.viewport, &buffer[pos+1],
+                       sizeof(struct pipe_viewport_state));
+         pos += (1 + sizeof(struct pipe_viewport_state) / 4);
+         break;
+      case CELL_CMD_STATE_VS_ARRAY_INFO:
+         cmd_state_vs_array_info((struct cell_array_info *) &buffer[pos+1]);
+         pos += (1 + sizeof(struct cell_array_info) / 4);
+         break;
+      case CELL_CMD_VS_EXECUTE:
+         spu_execute_vertex_shader(&draw,
+                                   (struct cell_command_vs *) &buffer[pos+1]);
+         pos += (1 + sizeof(struct cell_command_vs) / 4);
+         break;
       default:
          printf("SPU %u: bad opcode: 0x%x\n", spu.init.id, buffer[pos]);
          ASSERT(0);
diff --git a/src/mesa/pipe/cell/spu/spu_util.c b/src/mesa/pipe/cell/spu/spu_util.c
new file mode 100644
index 0000000000..ac373240c1
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_util.c
@@ -0,0 +1,165 @@
+#include "pipe/p_util.h"
+#include "pipe/p_shader_tokens.h"
+#include "pipe/tgsi/util/tgsi_parse.h"
+//#include "tgsi_build.h"
+#include "pipe/tgsi/util/tgsi_util.h"
+
+unsigned
+tgsi_util_get_src_register_swizzle(
+   const struct tgsi_src_register *reg,
+   unsigned component )
+{
+   switch( component ) {
+   case 0:
+      return reg->SwizzleX;
+   case 1:
+      return reg->SwizzleY;
+   case 2:
+      return reg->SwizzleZ;
+   case 3:
+      return reg->SwizzleW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+unsigned
+tgsi_util_get_src_register_extswizzle(
+   const struct tgsi_src_register_ext_swz *reg,
+   unsigned component )
+{
+   switch( component ) {
+   case 0:
+      return reg->ExtSwizzleX;
+   case 1:
+      return reg->ExtSwizzleY;
+   case 2:
+      return reg->ExtSwizzleZ;
+   case 3:
+      return reg->ExtSwizzleW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+unsigned
+tgsi_util_get_full_src_register_extswizzle(
+   const struct tgsi_full_src_register  *reg,
+   unsigned component )
+{
+   unsigned swizzle;
+
+   /*
+    * First, calculate  the   extended swizzle for a given channel. This will give
+    * us either a channel index into the simple swizzle or  a constant 1 or   0.
+    */
+   swizzle = tgsi_util_get_src_register_extswizzle(
+      &reg->SrcRegisterExtSwz,
+      component );
+
+   assert (TGSI_SWIZZLE_X == TGSI_EXTSWIZZLE_X);
+   assert (TGSI_SWIZZLE_Y == TGSI_EXTSWIZZLE_Y);
+   assert (TGSI_SWIZZLE_Z == TGSI_EXTSWIZZLE_Z);
+   assert (TGSI_SWIZZLE_W == TGSI_EXTSWIZZLE_W);
+   assert (TGSI_EXTSWIZZLE_ZERO > TGSI_SWIZZLE_W);
+   assert (TGSI_EXTSWIZZLE_ONE > TGSI_SWIZZLE_W);
+
+   /*
+    * Second, calculate the simple  swizzle  for   the   unswizzled channel index.
+    * Leave the constants intact, they are   not   affected by the   simple swizzle.
+    */
+   if( swizzle <= TGSI_SWIZZLE_W ) {
+      swizzle = tgsi_util_get_src_register_swizzle(
+         &reg->SrcRegister,
+         component );
+   }
+
+   return swizzle;
+}
+
+unsigned
+tgsi_util_get_src_register_extnegate(
+   const  struct tgsi_src_register_ext_swz *reg,
+   unsigned component )
+{
+   switch( component ) {
+   case 0:
+      return reg->NegateX;
+   case 1:
+      return reg->NegateY;
+   case 2:
+      return reg->NegateZ;
+   case 3:
+      return reg->NegateW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+void
+tgsi_util_set_src_register_extnegate(
+   struct tgsi_src_register_ext_swz *reg,
+   unsigned negate,
+   unsigned component )
+{
+   switch( component ) {
+   case 0:
+      reg->NegateX = negate;
+      break;
+   case 1:
+      reg->NegateY = negate;
+      break;
+   case 2:
+      reg->NegateZ = negate;
+      break;
+   case 3:
+      reg->NegateW = negate;
+      break;
+   default:
+      assert( 0 );
+   }
+}
+
+unsigned
+tgsi_util_get_full_src_register_sign_mode(
+   const struct  tgsi_full_src_register *reg,
+   unsigned component )
+{
+   unsigned sign_mode;
+
+   if( reg->SrcRegisterExtMod.Absolute ) {
+      /* Consider only the post-abs negation. */
+
+      if( reg->SrcRegisterExtMod.Negate ) {
+         sign_mode = TGSI_UTIL_SIGN_SET;
+      }
+      else {
+         sign_mode = TGSI_UTIL_SIGN_CLEAR;
+      }
+   }
+   else {
+      /* Accumulate the three negations. */
+
+      unsigned negate;
+
+      negate = reg->SrcRegister.Negate;
+      if( tgsi_util_get_src_register_extnegate( &reg->SrcRegisterExtSwz, component ) ) {
+         negate = !negate;
+      }
+      if( reg->SrcRegisterExtMod.Negate ) {
+         negate = !negate;
+      }
+
+      if( negate ) {
+         sign_mode = TGSI_UTIL_SIGN_TOGGLE;
+      }
+      else {
+         sign_mode = TGSI_UTIL_SIGN_KEEP;
+      }
+   }
+
+   return sign_mode;
+}
diff --git a/src/mesa/pipe/cell/spu/spu_vertex_fetch.c b/src/mesa/pipe/cell/spu/spu_vertex_fetch.c
new file mode 100644
index 0000000000..b8f8c52eed
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_vertex_fetch.c
@@ -0,0 +1,493 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * All Rights Reserved.
+ * 
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ * 
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * 
+ **************************************************************************/
+
+ /*
+  * Authors:
+  *   Keith Whitwell <keith@tungstengraphics.com>
+  */
+
+#include "pipe/p_util.h"
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "spu_exec.h"
+#include "spu_vertex_shader.h"
+
+
+#define DRAW_DBG 0
+
+
+/**
+ * Fetch a float[4] vertex attribute from memory, doing format/type
+ * conversion as needed.
+ *
+ * This is probably needed/dupliocated elsewhere, eg format
+ * conversion, texture sampling etc.
+ */
+#define FETCH_ATTRIB( NAME, SZ, CVT )			\
+static void						\
+fetch_##NAME(const void *ptr, float *attrib)		\
+{							\
+   static const float defaults[4] = { 0,0,0,1 };	\
+   int i;						\
+							\
+   for (i = 0; i < SZ; i++) {				\
+      attrib[i] = CVT;					\
+   }							\
+							\
+   for (; i < 4; i++) {					\
+      attrib[i] = defaults[i];				\
+   }							\
+}
+
+#define CVT_64_FLOAT   (float) ((double *) ptr)[i]
+#define CVT_32_FLOAT   ((float *) ptr)[i]
+
+#define CVT_8_USCALED  (float) ((unsigned char *) ptr)[i]
+#define CVT_16_USCALED (float) ((unsigned short *) ptr)[i]
+#define CVT_32_USCALED (float) ((unsigned int *) ptr)[i]
+
+#define CVT_8_SSCALED  (float) ((char *) ptr)[i]
+#define CVT_16_SSCALED (float) ((short *) ptr)[i]
+#define CVT_32_SSCALED (float) ((int *) ptr)[i]
+
+#define CVT_8_UNORM    (float) ((unsigned char *) ptr)[i] / 255.0f
+#define CVT_16_UNORM   (float) ((unsigned short *) ptr)[i] / 65535.0f
+#define CVT_32_UNORM   (float) ((unsigned int *) ptr)[i] / 4294967295.0f
+
+#define CVT_8_SNORM    (float) ((char *) ptr)[i] / 127.0f
+#define CVT_16_SNORM   (float) ((short *) ptr)[i] / 32767.0f
+#define CVT_32_SNORM   (float) ((int *) ptr)[i] / 2147483647.0f
+
+FETCH_ATTRIB( R64G64B64A64_FLOAT,   4, CVT_64_FLOAT )
+FETCH_ATTRIB( R64G64B64_FLOAT,      3, CVT_64_FLOAT )
+FETCH_ATTRIB( R64G64_FLOAT,         2, CVT_64_FLOAT )
+FETCH_ATTRIB( R64_FLOAT,            1, CVT_64_FLOAT )
+
+FETCH_ATTRIB( R32G32B32A32_FLOAT,   4, CVT_32_FLOAT )
+FETCH_ATTRIB( R32G32B32_FLOAT,      3, CVT_32_FLOAT )
+FETCH_ATTRIB( R32G32_FLOAT,         2, CVT_32_FLOAT )
+FETCH_ATTRIB( R32_FLOAT,            1, CVT_32_FLOAT )
+
+FETCH_ATTRIB( R32G32B32A32_USCALED, 4, CVT_32_USCALED )
+FETCH_ATTRIB( R32G32B32_USCALED,    3, CVT_32_USCALED )
+FETCH_ATTRIB( R32G32_USCALED,       2, CVT_32_USCALED )
+FETCH_ATTRIB( R32_USCALED,          1, CVT_32_USCALED )
+
+FETCH_ATTRIB( R32G32B32A32_SSCALED, 4, CVT_32_SSCALED )
+FETCH_ATTRIB( R32G32B32_SSCALED,    3, CVT_32_SSCALED )
+FETCH_ATTRIB( R32G32_SSCALED,       2, CVT_32_SSCALED )
+FETCH_ATTRIB( R32_SSCALED,          1, CVT_32_SSCALED )
+
+FETCH_ATTRIB( R32G32B32A32_UNORM, 4, CVT_32_UNORM )
+FETCH_ATTRIB( R32G32B32_UNORM,    3, CVT_32_UNORM )
+FETCH_ATTRIB( R32G32_UNORM,       2, CVT_32_UNORM )
+FETCH_ATTRIB( R32_UNORM,          1, CVT_32_UNORM )
+
+FETCH_ATTRIB( R32G32B32A32_SNORM, 4, CVT_32_SNORM )
+FETCH_ATTRIB( R32G32B32_SNORM,    3, CVT_32_SNORM )
+FETCH_ATTRIB( R32G32_SNORM,       2, CVT_32_SNORM )
+FETCH_ATTRIB( R32_SNORM,          1, CVT_32_SNORM )
+
+FETCH_ATTRIB( R16G16B16A16_USCALED, 4, CVT_16_USCALED )
+FETCH_ATTRIB( R16G16B16_USCALED,    3, CVT_16_USCALED )
+FETCH_ATTRIB( R16G16_USCALED,       2, CVT_16_USCALED )
+FETCH_ATTRIB( R16_USCALED,          1, CVT_16_USCALED )
+
+FETCH_ATTRIB( R16G16B16A16_SSCALED, 4, CVT_16_SSCALED )
+FETCH_ATTRIB( R16G16B16_SSCALED,    3, CVT_16_SSCALED )
+FETCH_ATTRIB( R16G16_SSCALED,       2, CVT_16_SSCALED )
+FETCH_ATTRIB( R16_SSCALED,          1, CVT_16_SSCALED )
+
+FETCH_ATTRIB( R16G16B16A16_UNORM, 4, CVT_16_UNORM )
+FETCH_ATTRIB( R16G16B16_UNORM,    3, CVT_16_UNORM )
+FETCH_ATTRIB( R16G16_UNORM,       2, CVT_16_UNORM )
+FETCH_ATTRIB( R16_UNORM,          1, CVT_16_UNORM )
+
+FETCH_ATTRIB( R16G16B16A16_SNORM, 4, CVT_16_SNORM )
+FETCH_ATTRIB( R16G16B16_SNORM,    3, CVT_16_SNORM )
+FETCH_ATTRIB( R16G16_SNORM,       2, CVT_16_SNORM )
+FETCH_ATTRIB( R16_SNORM,          1, CVT_16_SNORM )
+
+FETCH_ATTRIB( R8G8B8A8_USCALED,   4, CVT_8_USCALED )
+FETCH_ATTRIB( R8G8B8_USCALED,     3, CVT_8_USCALED )
+FETCH_ATTRIB( R8G8_USCALED,       2, CVT_8_USCALED )
+FETCH_ATTRIB( R8_USCALED,         1, CVT_8_USCALED )
+
+FETCH_ATTRIB( R8G8B8A8_SSCALED,  4, CVT_8_SSCALED )
+FETCH_ATTRIB( R8G8B8_SSCALED,    3, CVT_8_SSCALED )
+FETCH_ATTRIB( R8G8_SSCALED,      2, CVT_8_SSCALED )
+FETCH_ATTRIB( R8_SSCALED,        1, CVT_8_SSCALED )
+
+FETCH_ATTRIB( R8G8B8A8_UNORM,  4, CVT_8_UNORM )
+FETCH_ATTRIB( R8G8B8_UNORM,    3, CVT_8_UNORM )
+FETCH_ATTRIB( R8G8_UNORM,      2, CVT_8_UNORM )
+FETCH_ATTRIB( R8_UNORM,        1, CVT_8_UNORM )
+
+FETCH_ATTRIB( R8G8B8A8_SNORM,  4, CVT_8_SNORM )
+FETCH_ATTRIB( R8G8B8_SNORM,    3, CVT_8_SNORM )
+FETCH_ATTRIB( R8G8_SNORM,      2, CVT_8_SNORM )
+FETCH_ATTRIB( R8_SNORM,        1, CVT_8_SNORM )
+
+FETCH_ATTRIB( A8R8G8B8_UNORM,       4, CVT_8_UNORM )
+//FETCH_ATTRIB( R8G8B8A8_UNORM,       4, CVT_8_UNORM )
+
+
+
+static spu_fetch_func get_fetch_func( enum pipe_format format )
+{
+#if 0
+   {
+      char tmp[80];
+      pf_sprint_name(tmp, format);
+      _mesa_printf("%s: %s\n", __FUNCTION__, tmp);
+   }
+#endif
+
+   switch (format) {
+   case PIPE_FORMAT_R64_FLOAT:
+      return fetch_R64_FLOAT;
+   case PIPE_FORMAT_R64G64_FLOAT:
+      return fetch_R64G64_FLOAT;
+   case PIPE_FORMAT_R64G64B64_FLOAT:
+      return fetch_R64G64B64_FLOAT;
+   case PIPE_FORMAT_R64G64B64A64_FLOAT:
+      return fetch_R64G64B64A64_FLOAT;
+
+   case PIPE_FORMAT_R32_FLOAT:
+      return fetch_R32_FLOAT;
+   case PIPE_FORMAT_R32G32_FLOAT:
+      return fetch_R32G32_FLOAT;
+   case PIPE_FORMAT_R32G32B32_FLOAT:
+      return fetch_R32G32B32_FLOAT;
+   case PIPE_FORMAT_R32G32B32A32_FLOAT:
+      return fetch_R32G32B32A32_FLOAT;
+
+   case PIPE_FORMAT_R32_UNORM:
+      return fetch_R32_UNORM;
+   case PIPE_FORMAT_R32G32_UNORM:
+      return fetch_R32G32_UNORM;
+   case PIPE_FORMAT_R32G32B32_UNORM:
+      return fetch_R32G32B32_UNORM;
+   case PIPE_FORMAT_R32G32B32A32_UNORM:
+      return fetch_R32G32B32A32_UNORM;
+
+   case PIPE_FORMAT_R32_USCALED:
+      return fetch_R32_USCALED;
+   case PIPE_FORMAT_R32G32_USCALED:
+      return fetch_R32G32_USCALED;
+   case PIPE_FORMAT_R32G32B32_USCALED:
+      return fetch_R32G32B32_USCALED;
+   case PIPE_FORMAT_R32G32B32A32_USCALED:
+      return fetch_R32G32B32A32_USCALED;
+
+   case PIPE_FORMAT_R32_SNORM:
+      return fetch_R32_SNORM;
+   case PIPE_FORMAT_R32G32_SNORM:
+      return fetch_R32G32_SNORM;
+   case PIPE_FORMAT_R32G32B32_SNORM:
+      return fetch_R32G32B32_SNORM;
+   case PIPE_FORMAT_R32G32B32A32_SNORM:
+      return fetch_R32G32B32A32_SNORM;
+
+   case PIPE_FORMAT_R32_SSCALED:
+      return fetch_R32_SSCALED;
+   case PIPE_FORMAT_R32G32_SSCALED:
+      return fetch_R32G32_SSCALED;
+   case PIPE_FORMAT_R32G32B32_SSCALED:
+      return fetch_R32G32B32_SSCALED;
+   case PIPE_FORMAT_R32G32B32A32_SSCALED:
+      return fetch_R32G32B32A32_SSCALED;
+
+   case PIPE_FORMAT_R16_UNORM:
+      return fetch_R16_UNORM;
+   case PIPE_FORMAT_R16G16_UNORM:
+      return fetch_R16G16_UNORM;
+   case PIPE_FORMAT_R16G16B16_UNORM:
+      return fetch_R16G16B16_UNORM;
+   case PIPE_FORMAT_R16G16B16A16_UNORM:
+      return fetch_R16G16B16A16_UNORM;
+
+   case PIPE_FORMAT_R16_USCALED:
+      return fetch_R16_USCALED;
+   case PIPE_FORMAT_R16G16_USCALED:
+      return fetch_R16G16_USCALED;
+   case PIPE_FORMAT_R16G16B16_USCALED:
+      return fetch_R16G16B16_USCALED;
+   case PIPE_FORMAT_R16G16B16A16_USCALED:
+      return fetch_R16G16B16A16_USCALED;
+
+   case PIPE_FORMAT_R16_SNORM:
+      return fetch_R16_SNORM;
+   case PIPE_FORMAT_R16G16_SNORM:
+      return fetch_R16G16_SNORM;
+   case PIPE_FORMAT_R16G16B16_SNORM:
+      return fetch_R16G16B16_SNORM;
+   case PIPE_FORMAT_R16G16B16A16_SNORM:
+      return fetch_R16G16B16A16_SNORM;
+
+   case PIPE_FORMAT_R16_SSCALED:
+      return fetch_R16_SSCALED;
+   case PIPE_FORMAT_R16G16_SSCALED:
+      return fetch_R16G16_SSCALED;
+   case PIPE_FORMAT_R16G16B16_SSCALED:
+      return fetch_R16G16B16_SSCALED;
+   case PIPE_FORMAT_R16G16B16A16_SSCALED:
+      return fetch_R16G16B16A16_SSCALED;
+
+   case PIPE_FORMAT_R8_UNORM:
+      return fetch_R8_UNORM;
+   case PIPE_FORMAT_R8G8_UNORM:
+      return fetch_R8G8_UNORM;
+   case PIPE_FORMAT_R8G8B8_UNORM:
+      return fetch_R8G8B8_UNORM;
+   case PIPE_FORMAT_R8G8B8A8_UNORM:
+      return fetch_R8G8B8A8_UNORM;
+
+   case PIPE_FORMAT_R8_USCALED:
+      return fetch_R8_USCALED;
+   case PIPE_FORMAT_R8G8_USCALED:
+      return fetch_R8G8_USCALED;
+   case PIPE_FORMAT_R8G8B8_USCALED:
+      return fetch_R8G8B8_USCALED;
+   case PIPE_FORMAT_R8G8B8A8_USCALED:
+      return fetch_R8G8B8A8_USCALED;
+
+   case PIPE_FORMAT_R8_SNORM:
+      return fetch_R8_SNORM;
+   case PIPE_FORMAT_R8G8_SNORM:
+      return fetch_R8G8_SNORM;
+   case PIPE_FORMAT_R8G8B8_SNORM:
+      return fetch_R8G8B8_SNORM;
+   case PIPE_FORMAT_R8G8B8A8_SNORM:
+      return fetch_R8G8B8A8_SNORM;
+
+   case PIPE_FORMAT_R8_SSCALED:
+      return fetch_R8_SSCALED;
+   case PIPE_FORMAT_R8G8_SSCALED:
+      return fetch_R8G8_SSCALED;
+   case PIPE_FORMAT_R8G8B8_SSCALED:
+      return fetch_R8G8B8_SSCALED;
+   case PIPE_FORMAT_R8G8B8A8_SSCALED:
+      return fetch_R8G8B8A8_SSCALED;
+
+   case PIPE_FORMAT_A8R8G8B8_UNORM:
+      return fetch_A8R8G8B8_UNORM;
+
+   case 0:
+      return NULL;		/* not sure why this is needed */
+
+   default:
+      assert(0);
+      return NULL;
+   }
+}
+
+
+static void 
+transpose_4x4( float *out, const float *in )
+{
+   /* This can be achieved in 12 sse instructions, plus the final
+    * stores I guess.  This is probably a bit more than that - maybe
+    * 32 or so?
+    */
+   out[0] = in[0];  out[1] = in[4];  out[2] = in[8];   out[3] = in[12];
+   out[4] = in[1];  out[5] = in[5];  out[6] = in[9];   out[7] = in[13];
+   out[8] = in[2];  out[9] = in[6];  out[10] = in[10]; out[11] = in[14];
+   out[12] = in[3]; out[13] = in[7]; out[14] = in[11]; out[15] = in[15];
+}
+
+
+
+static void fetch_xyz_rgb( struct spu_vs_context *draw,
+			   struct spu_exec_machine *machine,
+			   const unsigned *elts,
+			   unsigned count )
+{
+   assert(count <= 4);
+
+//   _mesa_printf("%s\n", __FUNCTION__);
+
+   /* loop over vertex attributes (vertex shader inputs)
+    */
+
+   const unsigned *pitch   = draw->vertex_fetch.pitch;
+   const ubyte **src       = draw->vertex_fetch.src_ptr;
+   int i;
+
+   for (i = 0; i < 4; i++) {
+      {
+	 const float *in = (const float *)(src[0] + elts[i] * pitch[0]);
+	 float *out = &machine->Inputs[0].xyzw[0].f[i];
+	 out[0] = in[0];
+	 out[4] = in[1];
+	 out[8] = in[2];
+ 	 out[12] = 1.0f;
+      }
+
+      {
+	 const float *in = (const float *)(src[1] + elts[i] * pitch[1]);
+	 float *out = &machine->Inputs[1].xyzw[0].f[i];
+	 out[0] = in[0];
+	 out[4] = in[1];
+	 out[8] = in[2];
+ 	 out[12] = 1.0f;
+      }
+   }
+}
+
+
+
+
+static void fetch_xyz_rgb_st( struct spu_vs_context *draw,
+			      struct spu_exec_machine *machine,
+			      const unsigned *elts,
+			      unsigned count )
+{
+   assert(count <= 4);
+
+   /* loop over vertex attributes (vertex shader inputs)
+    */
+
+   const unsigned *pitch   = draw->vertex_fetch.pitch;
+   const ubyte **src       = draw->vertex_fetch.src_ptr;
+   int i;
+
+   for (i = 0; i < 4; i++) {
+      {
+	 const float *in = (const float *)(src[0] + elts[i] * pitch[0]);
+	 float *out = &machine->Inputs[0].xyzw[0].f[i];
+	 out[0] = in[0];
+	 out[4] = in[1];
+	 out[8] = in[2];
+ 	 out[12] = 1.0f;
+      }
+
+      {
+	 const float *in = (const float *)(src[1] + elts[i] * pitch[1]);
+	 float *out = &machine->Inputs[1].xyzw[0].f[i];
+	 out[0] = in[0];
+	 out[4] = in[1];
+	 out[8] = in[2];
+ 	 out[12] = 1.0f;
+      }
+
+      {
+	 const float *in = (const float *)(src[2] + elts[i] * pitch[2]);
+	 float *out = &machine->Inputs[1].xyzw[0].f[i];
+	 out[0] = in[0];
+	 out[4] = in[1];
+	 out[8] = 0.0f;
+ 	 out[12] = 1.0f;
+      }
+   }
+}
+
+
+
+
+/**
+ * Fetch vertex attributes for 'count' vertices.
+ */
+static void generic_vertex_fetch( struct spu_vs_context *draw,
+				  struct spu_exec_machine *machine,
+				  const unsigned *elts,
+				  unsigned count )
+{
+   unsigned nr_attrs = draw->vertex_fetch.nr_attrs;
+   unsigned attr;
+
+   assert(count <= 4);
+
+//   _mesa_printf("%s %d\n", __FUNCTION__, count);
+
+   /* loop over vertex attributes (vertex shader inputs)
+    */
+   for (attr = 0; attr < nr_attrs; attr++) {
+
+      const unsigned pitch   = draw->vertex_fetch.pitch[attr];
+      const ubyte *src = draw->vertex_fetch.src_ptr[attr];
+      const spu_fetch_func fetch = draw->vertex_fetch.fetch[attr];
+      unsigned i;
+      float p[4][4];
+
+
+      /* Fetch four attributes for four vertices.  
+       * 
+       * Could fetch directly into AOS format, but this is meant to be
+       * a prototype for an sse implementation, which would have
+       * difficulties doing that.
+       */
+      for (i = 0; i < count; i++) 
+	 fetch( src + elts[i] * pitch, p[i] );
+
+      /* Be nice and zero out any missing vertices: 
+       */
+      for (/* empty */; i < 4; i++) 
+	 p[i][0] = p[i][1] = p[i][2] = p[i][3] = 0;
+      
+      /* Transpose/swizzle into sse-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_4x4( (float *)&machine->Inputs[attr].xyzw[0].f[0], (float *)p );
+   }
+}
+
+
+void spu_update_vertex_fetch( struct spu_vs_context *draw )
+{
+   unsigned i;
+
+   
+   for (i = 0; i < draw->vertex_fetch.nr_attrs; i++) {
+      draw->vertex_fetch.fetch[i] =
+          get_fetch_func(draw->vertex_fetch.format[i]);
+   }
+
+   draw->vertex_fetch.fetch_func = generic_vertex_fetch;
+
+   switch (draw->vertex_fetch.nr_attrs) {
+   case 2:
+      if (draw->vertex_fetch.format[0] == PIPE_FORMAT_R32G32B32_FLOAT &&
+          draw->vertex_fetch.format[1] == PIPE_FORMAT_R32G32B32_FLOAT)
+          draw->vertex_fetch.fetch_func = fetch_xyz_rgb;
+      break;
+   case 3:
+      if (draw->vertex_fetch.format[0] == PIPE_FORMAT_R32G32B32_FLOAT &&
+          draw->vertex_fetch.format[1] == PIPE_FORMAT_R32G32B32_FLOAT &&
+          draw->vertex_fetch.format[2] == PIPE_FORMAT_R32G32_FLOAT)
+          draw->vertex_fetch.fetch_func = fetch_xyz_rgb_st;
+      break;
+   default:
+      break;
+   }
+}
diff --git a/src/mesa/pipe/cell/spu/spu_vertex_shader.c b/src/mesa/pipe/cell/spu/spu_vertex_shader.c
new file mode 100644
index 0000000000..e694ff729f
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_vertex_shader.c
@@ -0,0 +1,224 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * All Rights Reserved.
+ * 
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ * 
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * 
+ **************************************************************************/
+
+ /*
+  * Authors:
+  *   Keith Whitwell <keith@tungstengraphics.com>
+  *   Brian Paul
+  *   Ian Romanick <idr@us.ibm.com>
+  */
+
+#include "pipe/p_util.h"
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "spu_vertex_shader.h"
+#include "spu_exec.h"
+#include "pipe/draw/draw_private.h"
+#include "pipe/draw/draw_context.h"
+#include "pipe/cell/common.h"
+
+#define DBG_VS 0
+
+
+static INLINE unsigned
+compute_clipmask(const float *clip, /*const*/ float plane[][4], unsigned nr)
+{
+   unsigned mask = 0;
+   unsigned i;
+
+   /* Do the hardwired planes first:
+    */
+   if (-clip[0] + clip[3] < 0) mask |= CLIP_RIGHT_BIT;
+   if ( clip[0] + clip[3] < 0) mask |= CLIP_LEFT_BIT;
+   if (-clip[1] + clip[3] < 0) mask |= CLIP_TOP_BIT;
+   if ( clip[1] + clip[3] < 0) mask |= CLIP_BOTTOM_BIT;
+   if (-clip[2] + clip[3] < 0) mask |= CLIP_FAR_BIT;
+   if ( clip[2] + clip[3] < 0) mask |= CLIP_NEAR_BIT;
+
+   /* Followed by any remaining ones:
+    */
+   for (i = 6; i < nr; i++) {
+      if (dot4(clip, plane[i]) < 0) 
+         mask |= (1<<i);
+   }
+
+   return mask;
+}
+
+
+/**
+ * Transform vertices with the current vertex program/shader
+ * Up to four vertices can be shaded at a time.
+ * \param vbuffer  the input vertex data
+ * \param elts  indexes of four input vertices
+ * \param count  number of vertices to shade [1..4]
+ * \param vOut  array of pointers to four output vertices
+ */
+static void
+run_vertex_program(struct spu_vs_context *draw,
+                   unsigned elts[4], unsigned count,
+                   struct vertex_header *vOut[])
+{
+   struct spu_exec_machine *machine = &draw->machine;
+   unsigned int j;
+
+   ALIGN16_DECL(struct spu_exec_vector, inputs, PIPE_ATTRIB_MAX);
+   ALIGN16_DECL(struct spu_exec_vector, outputs, PIPE_ATTRIB_MAX);
+   const float *scale = draw->viewport.scale;
+   const float *trans = draw->viewport.translate;
+
+   assert(count <= 4);
+
+   /* Consts does not require 16 byte alignment. */
+   ASSERT_ALIGN16(draw->constants);
+   machine->Consts = (float (*)[4]) draw->constants;
+
+   machine->Inputs = ALIGN16_ASSIGN(inputs);
+   machine->Outputs = ALIGN16_ASSIGN(outputs);
+
+   spu_vertex_fetch( draw, machine, elts, count );
+
+   /* run shader */
+   spu_exec_machine_run( machine );
+
+
+   /* store machine results */
+   for (j = 0; j < count; j++) {
+      unsigned slot;
+      float x, y, z, w;
+
+      /* Handle attr[0] (position) specially:
+       *
+       * XXX: Computing the clipmask should be done in the vertex
+       * program as a set of DP4 instructions appended to the
+       * user-provided code.
+       */
+      x = vOut[j]->clip[0] = machine->Outputs[0].xyzw[0].f[j];
+      y = vOut[j]->clip[1] = machine->Outputs[0].xyzw[1].f[j];
+      z = vOut[j]->clip[2] = machine->Outputs[0].xyzw[2].f[j];
+      w = vOut[j]->clip[3] = machine->Outputs[0].xyzw[3].f[j];
+
+      vOut[j]->clipmask = compute_clipmask(vOut[j]->clip, draw->plane,
+					   draw->nr_planes);
+      vOut[j]->edgeflag = 1;
+
+      /* divide by w */
+      w = 1.0f / w;
+      x *= w;
+      y *= w;
+      z *= w;
+
+      /* Viewport mapping */
+      vOut[j]->data[0][0] = x * scale[0] + trans[0];
+      vOut[j]->data[0][1] = y * scale[1] + trans[1];
+      vOut[j]->data[0][2] = z * scale[2] + trans[2];
+      vOut[j]->data[0][3] = w;
+
+#if DBG_VS
+      printf("output[%d]win: %f %f %f %f\n", j,
+             vOut[j]->data[0][0],
+             vOut[j]->data[0][1],
+             vOut[j]->data[0][2],
+             vOut[j]->data[0][3]);
+#endif
+      /* Remaining attributes are packed into sequential post-transform
+       * vertex attrib slots.
+       */
+      for (slot = 1; slot < draw->num_vs_outputs; slot++) {
+         vOut[j]->data[slot][0] = machine->Outputs[slot].xyzw[0].f[j];
+         vOut[j]->data[slot][1] = machine->Outputs[slot].xyzw[1].f[j];
+         vOut[j]->data[slot][2] = machine->Outputs[slot].xyzw[2].f[j];
+         vOut[j]->data[slot][3] = machine->Outputs[slot].xyzw[3].f[j];
+#if DBG_VS
+         printf("output[%d][%d]: %f %f %f %f\n", j, slot,
+                vOut[j]->data[slot][0],
+                vOut[j]->data[slot][1],
+                vOut[j]->data[slot][2],
+                vOut[j]->data[slot][3]);
+#endif
+      }
+   } /* loop over vertices */
+}
+
+
+static void
+spu_bind_vertex_shader(struct spu_vs_context *draw,
+		       void *uniforms,
+		       void *planes,
+		       unsigned nr_planes,
+		       unsigned num_outputs
+		       )
+{
+   draw->constants = (float (*)[4]) uniforms;
+
+   (void) memcpy(draw->plane, planes, sizeof(float) * 4 * nr_planes);
+   draw->nr_planes = nr_planes;
+   draw->num_vs_outputs = num_outputs;
+
+   /* specify the shader to interpret/execute */
+   spu_exec_machine_init(&draw->machine,
+			 PIPE_MAX_SAMPLERS,
+			 NULL /*samplers*/,
+			 PIPE_SHADER_VERTEX);
+}
+
+
+void
+spu_execute_vertex_shader(struct spu_vs_context *draw,
+			  const struct cell_command_vs *vs)
+{
+   unsigned i;
+   unsigned j;
+
+   draw->machine.Instructions = (struct tgsi_full_instruction *)
+       vs->shader.instructions;
+   draw->machine.NumInstructions = vs->shader.num_instructions;
+
+   draw->machine.Declarations = (struct tgsi_full_declaration *)
+       vs->shader.declarations;
+   draw->machine.NumDeclarations = vs->shader.num_declarations;
+
+   spu_bind_vertex_shader(draw, vs->shader.uniforms,
+			  NULL, -1,
+			  vs->shader.num_outputs);
+   
+   for (i = 0; i < vs->num_elts; i += 4) {
+      const unsigned batch_size = MIN2(vs->num_elts - i, 4);
+      unsigned elts[4];
+
+      for (j = 0; j < batch_size; j++) {
+	 switch (vs->bytes_per_elt) {
+	 case 1: elts[j] = ((unsigned char *) vs->elts)[i + j]; break;
+	 case 2: elts[j] = ((unsigned short *)vs->elts)[i + j]; break;
+	 case 4: elts[j] = ((unsigned int *)  vs->elts)[i + j]; break;
+	 }
+      }
+
+      run_vertex_program(draw, elts, batch_size,
+			 (struct vertex_header (*)[]) vs->vOut);
+   }
+}
diff --git a/src/mesa/pipe/cell/spu/spu_vertex_shader.h b/src/mesa/pipe/cell/spu/spu_vertex_shader.h
new file mode 100644
index 0000000000..c52f38fd02
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_vertex_shader.h
@@ -0,0 +1,61 @@
+#ifndef SPU_VERTEX_SHADER_H
+#define SPU_VERTEX_SHADER_H
+
+#include "pipe/p_format.h"
+#include "spu_exec.h"
+
+struct spu_vs_context;
+
+typedef void (*spu_fetch_func)(const void *ptr, float *attrib);
+typedef void (*spu_full_fetch_func)( struct spu_vs_context *draw,
+				     struct spu_exec_machine *machine,
+				     const unsigned *elts,
+				     unsigned count );
+
+struct spu_vs_context {
+   struct pipe_viewport_state viewport;
+
+   struct {
+      const ubyte *src_ptr[PIPE_ATTRIB_MAX];
+      unsigned pitch[PIPE_ATTRIB_MAX];
+      enum pipe_format format[PIPE_ATTRIB_MAX];
+      unsigned nr_attrs;
+      boolean dirty;
+
+      spu_fetch_func fetch[PIPE_ATTRIB_MAX];
+      spu_full_fetch_func fetch_func;
+   } vertex_fetch;
+   
+   /* Clip derived state:
+    */
+   float plane[12][4];
+   unsigned nr_planes;
+
+   struct spu_exec_machine machine;
+   const float (*constants)[4];
+
+   unsigned num_vs_outputs;
+};
+
+extern void spu_update_vertex_fetch(struct spu_vs_context *draw);
+
+static INLINE void spu_vertex_fetch(struct spu_vs_context *draw,
+				    struct spu_exec_machine *machine,
+				    const unsigned *elts,
+				    unsigned count)
+{
+   if (draw->vertex_fetch.dirty) {
+      spu_update_vertex_fetch(draw);
+      draw->vertex_fetch.dirty = 0;
+   }
+   
+   (*draw->vertex_fetch.fetch_func)(draw, machine, elts, count);
+}
+
+struct cell_command_vs;
+
+extern void
+spu_execute_vertex_shader(struct spu_vs_context *draw,
+			  const struct cell_command_vs *vs);
+
+#endif /* SPU_VERTEX_SHADER_H */