Cell: prefix SPU sources with spu_

7 files changed, 1751 insertions, 12 deletions

diff --git a/src/mesa/pipe/cell/spu/Makefile b/src/mesa/pipe/cell/spu/Makefile
index bf97a49e782..b92a756cd50 100644
--- a/src/mesa/pipe/cell/spu/Makefile
+++ b/src/mesa/pipe/cell/spu/Makefile
@@ -16,15 +16,19 @@ PROG_SPU_EMBED_O = $(PROG)_spu-embed.o
 
 
 SOURCES = \
-	main.c \
-	tile.c \
-	tri.c
+	spu_main.c \
+	spu_tile.c \
+	spu_tri.c
 
 SPU_OBJECTS = $(SOURCES:.c=.o) \
 
 INCLUDE_DIRS = -I$(TOP)/src/mesa
 
 
+.c.o:
+	$(SPU_CC) $(SPU_CFLAGS) -c $<
+
+
 # The .a file will be linked into the main/PPU executable
 default: $(PROG_SPU_A)
 
@@ -38,15 +42,6 @@ $(PROG_SPU): $(SPU_OBJECTS)
 	$(SPU_CC) -o $(PROG_SPU) $(SPU_OBJECTS) $(SPU_LFLAGS)
 
 
-main.o: main.c
-	$(SPU_CC) $(SPU_CFLAGS) -c main.c
-
-tile.o: tile.c
-	$(SPU_CC) $(SPU_CFLAGS) -c tile.c
-
-tri.o: tri.c
-	$(SPU_CC) $(SPU_CFLAGS) -c tri.c
-
 
 clean:
 	rm -f *~ *.o *.a *.d $(PROG_SPU)
diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
new file mode 100644
index 00000000000..df708e65d13
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -0,0 +1,553 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+
+/* main() for Cell SPU code */
+
+
+#include <stdio.h>
+#include <libmisc.h>
+#include <spu_mfcio.h>
+
+#include "spu_main.h"
+#include "spu_tri.h"
+#include "spu_tile.h"
+#include "pipe/cell/common.h"
+#include "pipe/p_defines.h"
+
+/*
+helpful headers:
+/usr/lib/gcc/spu/4.1.1/include/spu_mfcio.h
+/opt/ibm/cell-sdk/prototype/sysroot/usr/include/libmisc.h
+*/
+
+static boolean Debug = FALSE;
+
+volatile struct cell_init_info init;
+
+struct framebuffer fb;
+
+
+
+void
+wait_on_mask(unsigned tagMask)
+{
+   mfc_write_tag_mask( tagMask );
+   /* wait for completion of _any_ DMAs specified by tagMask */
+   mfc_read_tag_status_any();
+}
+
+
+static void
+wait_on_mask_all(unsigned tagMask)
+{
+   mfc_write_tag_mask( tagMask );
+   /* wait for completion of _any_ DMAs specified by tagMask */
+   mfc_read_tag_status_all();
+}
+
+
+
+/**
+ * For tiles whose status is TILE_STATUS_CLEAR, write solid-filled
+ * tiles back to the main framebuffer.
+ */
+static void
+really_clear_tiles(uint surfaceIndex)
+{
+   const uint num_tiles = fb.width_tiles * fb.height_tiles;
+   uint i, j;
+
+   if (surfaceIndex == 0) {
+      for (i = 0; i < TILE_SIZE; i++)
+         for (j = 0; j < TILE_SIZE; j++)
+            ctile[i][j] = fb.color_clear_value;
+
+      for (i = init.id; i < num_tiles; i += init.num_spus) {
+         uint tx = i % fb.width_tiles;
+         uint ty = i / fb.width_tiles;
+         if (tile_status[ty][tx] == TILE_STATUS_CLEAR) {
+            put_tile(&fb, tx, ty, (uint *) ctile, TAG_SURFACE_CLEAR, 0);
+         }
+      }
+   }
+   else {
+      for (i = 0; i < TILE_SIZE; i++)
+         for (j = 0; j < TILE_SIZE; j++)
+            ztile[i][j] = fb.depth_clear_value;
+
+      for (i = init.id; i < num_tiles; i += init.num_spus) {
+         uint tx = i % fb.width_tiles;
+         uint ty = i / fb.width_tiles;
+         if (tile_status_z[ty][tx] == TILE_STATUS_CLEAR)
+            put_tile(&fb, tx, ty, (uint *) ctile, TAG_SURFACE_CLEAR, 1);
+      }
+   }
+
+#if 0
+   wait_on_mask(1 << TAG_SURFACE_CLEAR);
+#endif
+}
+
+
+static void
+cmd_clear_surface(const struct cell_command_clear_surface *clear)
+{
+   const uint num_tiles = fb.width_tiles * fb.height_tiles;
+   uint i, j;
+
+   if (Debug)
+      printf("SPU %u: CLEAR SURF %u to 0x%08x\n", init.id,
+             clear->surface, clear->value);
+
+#define CLEAR_OPT 1
+#if CLEAR_OPT
+   /* set all tile's status to CLEAR */
+   if (clear->surface == 0) {
+      memset(tile_status, TILE_STATUS_CLEAR, sizeof(tile_status));
+      fb.color_clear_value = clear->value;
+   }
+   else {
+      memset(tile_status_z, TILE_STATUS_CLEAR, sizeof(tile_status_z));
+      fb.depth_clear_value = clear->value;
+   }
+   return;
+#endif
+
+   if (clear->surface == 0) {
+      for (i = 0; i < TILE_SIZE; i++)
+         for (j = 0; j < TILE_SIZE; j++)
+            ctile[i][j] = clear->value;
+   }
+   else {
+      for (i = 0; i < TILE_SIZE; i++)
+         for (j = 0; j < TILE_SIZE; j++)
+            ztile[i][j] = clear->value;
+   }
+
+   /*
+   printf("SPU: %s num=%d w=%d h=%d\n",
+          __FUNCTION__, num_tiles, fb.width_tiles, fb.height_tiles);
+   */
+
+   for (i = init.id; i < num_tiles; i += init.num_spus) {
+      uint tx = i % fb.width_tiles;
+      uint ty = i / fb.width_tiles;
+      if (clear->surface == 0)
+         put_tile(&fb, tx, ty, (uint *) ctile, TAG_SURFACE_CLEAR, 0);
+      else
+         put_tile(&fb, tx, ty, (uint *) ztile, TAG_SURFACE_CLEAR, 1);
+      /* XXX we don't want this here, but it fixes bad tile results */
+   }
+
+#if 0
+   wait_on_mask(1 << TAG_SURFACE_CLEAR);
+#endif
+}
+
+
+/**
+ * Given a rendering command's bounding box (in pixels) compute the
+ * location of the corresponding screen tile bounding box.
+ */
+static INLINE void
+tile_bounding_box(const struct cell_command_render *render,
+                  uint *txmin, uint *tymin,
+                  uint *box_num_tiles, uint *box_width_tiles)
+{
+#if 1
+   /* Debug: full-window bounding box */
+   uint txmax = fb.width_tiles - 1;
+   uint tymax = fb.height_tiles - 1;
+   *txmin = 0;
+   *tymin = 0;
+   *box_num_tiles = fb.width_tiles * fb.height_tiles;
+   *box_width_tiles = fb.width_tiles;
+   (void) render;
+   (void) txmax;
+   (void) tymax;
+#else
+   uint txmax, tymax, box_height_tiles;
+
+   *txmin = (uint) render->xmin / TILE_SIZE;
+   *tymin = (uint) render->ymin / TILE_SIZE;
+   txmax = (uint) render->xmax / TILE_SIZE;
+   tymax = (uint) render->ymax / TILE_SIZE;
+   *box_width_tiles = txmax - *txmin + 1;
+   box_height_tiles = tymax - *tymin + 1;
+   *box_num_tiles = *box_width_tiles * box_height_tiles;
+#endif
+#if 0
+   printf("Render bounds: %g, %g  ...  %g, %g\n",
+          render->xmin, render->ymin, render->xmax, render->ymax);
+   printf("Render tiles:  %u, %u .. %u, %u\n", *txmin, *tymin, txmax, tymax);
+#endif
+}
+
+
+static void
+cmd_render(const struct cell_command_render *render)
+{
+   /* we'll DMA into these buffers */
+   ubyte vertex_data[CELL_MAX_VBUF_SIZE] ALIGN16_ATTRIB;
+   ushort indexes[CELL_MAX_VBUF_INDEXES] ALIGN16_ATTRIB;
+
+   uint i, j, vertex_size, vertex_bytes, index_bytes;
+
+   if (Debug)
+      printf("SPU %u: RENDER prim %u, indices: %u, nr_vert: %u\n",
+             init.id,
+             render->prim_type,
+             render->num_verts,
+             render->num_indexes);
+
+
+   ASSERT_ALIGN16(render->vertex_data);
+   ASSERT_ALIGN16(render->index_data);
+
+   vertex_size = render->num_attribs * 4 * sizeof(float);
+
+   /* how much vertex data */
+   vertex_bytes = render->num_verts * vertex_size;
+   index_bytes = render->num_indexes * sizeof(ushort);
+   if (index_bytes < 8)
+      index_bytes = 8;
+   else
+      index_bytes = (index_bytes + 15) & ~0xf; /* multiple of 16 */
+
+   /*
+   printf("VBUF: indices at %p,  vertices at %p  vertex_bytes %u  ind_bytes %u\n",
+          render->index_data, render->vertex_data, vertex_bytes, index_bytes);
+   */
+
+   /* get vertex data from main memory */
+   mfc_get(vertex_data,  /* dest */
+           (unsigned int) render->vertex_data,  /* src */
+           vertex_bytes,  /* size */
+           TAG_VERTEX_BUFFER,
+           0, /* tid */
+           0  /* rid */);
+
+   /* get index data from main memory */
+   mfc_get(indexes,  /* dest */
+           (unsigned int) render->index_data,  /* src */
+           index_bytes,
+           TAG_INDEX_BUFFER,
+           0, /* tid */
+           0  /* rid */);
+
+   wait_on_mask_all((1 << TAG_VERTEX_BUFFER) |
+                    (1 << TAG_INDEX_BUFFER));
+
+   /* find tiles which intersect the prim bounding box */
+   uint txmin, tymin, box_width_tiles, box_num_tiles;
+#if 0
+   tile_bounding_box(render, &txmin, &tymin,
+                     &box_num_tiles, &box_width_tiles);
+#else
+   txmin = 0;
+   tymin = 0;
+   box_num_tiles = fb.width_tiles * fb.height_tiles;
+   box_width_tiles = fb.width_tiles;
+#endif
+
+   /* make sure any pending clears have completed */
+   wait_on_mask(1 << TAG_SURFACE_CLEAR);
+
+   /* loop over tiles */
+   for (i = init.id; i < box_num_tiles; i += init.num_spus) {
+      const uint tx = txmin + i % box_width_tiles;
+      const uint ty = tymin + i / box_width_tiles;
+
+      ASSERT(tx < fb.width_tiles);
+      ASSERT(ty < fb.height_tiles);
+
+      /* Start fetching color/z tiles.  We'll wait for completion when
+       * we need read/write to them later in triangle rasterization.
+       */
+      if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+         if (tile_status_z[ty][tx] != TILE_STATUS_CLEAR) {
+            get_tile(&fb, tx, ty, (uint *) ztile, TAG_READ_TILE_Z, 1);
+         }
+      }
+
+      if (tile_status[ty][tx] != TILE_STATUS_CLEAR) {
+         get_tile(&fb, tx, ty, (uint *) ctile, TAG_READ_TILE_COLOR, 0);
+      }
+
+      ASSERT(render->prim_type == PIPE_PRIM_TRIANGLES);
+
+      /* loop over tris */
+      for (j = 0; j < render->num_indexes; j += 3) {
+         const float *v0, *v1, *v2;
+
+         v0 = (const float *) (vertex_data + indexes[j+0] * vertex_size);
+         v1 = (const float *) (vertex_data + indexes[j+1] * vertex_size);
+         v2 = (const float *) (vertex_data + indexes[j+2] * vertex_size);
+
+         tri_draw(v0, v1, v2, tx, ty);
+      }
+
+      /* write color/z tiles back to main framebuffer, if dirtied */
+      if (tile_status[ty][tx] == TILE_STATUS_DIRTY) {
+         put_tile(&fb, tx, ty, (uint *) ctile, TAG_WRITE_TILE_COLOR, 0);
+         tile_status[ty][tx] = TILE_STATUS_DEFINED;
+      }
+      if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+         if (tile_status_z[ty][tx] == TILE_STATUS_DIRTY) {
+            put_tile(&fb, tx, ty, (uint *) ztile, TAG_WRITE_TILE_Z, 1);
+            tile_status_z[ty][tx] = TILE_STATUS_DEFINED;
+         }
+      }
+
+      /* XXX move these... */
+      wait_on_mask(1 << TAG_WRITE_TILE_COLOR);
+      if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+         wait_on_mask(1 << TAG_WRITE_TILE_Z);
+      }
+   }
+}
+
+
+static void
+cmd_framebuffer(const struct cell_command_framebuffer *cmd)
+{
+   if (Debug)
+      printf("SPU %u: FRAMEBUFFER: %d x %d at %p, cformat 0x%x  zformat 0x%x\n",
+             init.id,
+             cmd->width,
+             cmd->height,
+             cmd->color_start,
+             cmd->color_format,
+             cmd->depth_format);
+
+   fb.color_start = cmd->color_start;
+   fb.depth_start = cmd->depth_start;
+   fb.color_format = cmd->color_format;
+   fb.depth_format = cmd->depth_format;
+   fb.width = cmd->width;
+   fb.height = cmd->height;
+   fb.width_tiles = (fb.width + TILE_SIZE - 1) / TILE_SIZE;
+   fb.height_tiles = (fb.height + TILE_SIZE - 1) / TILE_SIZE;
+}
+
+
+static void
+cmd_finish(void)
+{
+   if (Debug)
+      printf("SPU %u: FINISH\n", init.id);
+   really_clear_tiles(0);
+   /* wait for all outstanding DMAs to finish */
+   mfc_write_tag_mask(~0);
+   mfc_read_tag_status_all();
+   /* send mbox message to PPU */
+   spu_write_out_mbox(CELL_CMD_FINISH);
+}
+
+
+/**
+ * Execute a batch of commands
+ * The opcode param encodes the location of the buffer and its size.
+ */
+static void
+cmd_batch(uint opcode)
+{
+   const uint buf = (opcode >> 8) & 0xff;
+   uint size = (opcode >> 16);
+   uint buffer[CELL_BATCH_BUFFER_SIZE / 4] ALIGN16_ATTRIB;
+   const uint usize = size / sizeof(uint);
+   uint pos;
+
+   if (Debug)
+      printf("SPU %u: BATCH buffer %u, len %u, from %p\n",
+             init.id, buf, size, init.batch_buffers[buf]);
+
+   ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH);
+
+   ASSERT_ALIGN16(init.batch_buffers[buf]);
+
+   size = (size + 0xf) & ~0xf;
+
+   mfc_get(buffer,  /* dest */
+           (unsigned int) init.batch_buffers[buf],  /* src */
+           size,
+           TAG_BATCH_BUFFER,
+           0, /* tid */
+           0  /* rid */);
+   wait_on_mask(1 << TAG_BATCH_BUFFER);
+
+   for (pos = 0; pos < usize; /* no incr */) {
+      switch (buffer[pos]) {
+      case CELL_CMD_FRAMEBUFFER:
+         {
+            struct cell_command_framebuffer *fb
+               = (struct cell_command_framebuffer *) &buffer[pos];
+            cmd_framebuffer(fb);
+            pos += sizeof(*fb) / 4;
+         }
+         break;
+      case CELL_CMD_CLEAR_SURFACE:
+         {
+            struct cell_command_clear_surface *clr
+               = (struct cell_command_clear_surface *) &buffer[pos];
+            cmd_clear_surface(clr);
+            pos += sizeof(*clr) / 4;
+         }
+         break;
+      case CELL_CMD_RENDER:
+         {
+            struct cell_command_render *render
+               = (struct cell_command_render *) &buffer[pos];
+            cmd_render(render);
+            pos += sizeof(*render) / 4;
+         }
+         break;
+      case CELL_CMD_FINISH:
+         cmd_finish();
+         pos += 1;
+         break;
+      default:
+         printf("SPU %u: bad opcode: 0x%x\n", init.id, buffer[pos]);
+         ASSERT(0);
+         break;
+      }
+   }
+
+   if (Debug)
+      printf("SPU %u: BATCH complete\n", init.id);
+}
+
+
+/**
+ * Temporary/simple main loop for SPEs: Get a command, execute it, repeat.
+ */
+static void
+main_loop(void)
+{
+   struct cell_command cmd;
+   int exitFlag = 0;
+
+   if (Debug)
+      printf("SPU %u: Enter main loop\n", init.id);
+
+   ASSERT((sizeof(struct cell_command) & 0xf) == 0);
+   ASSERT_ALIGN16(&cmd);
+
+   while (!exitFlag) {
+      unsigned opcode;
+      int tag = 0;
+
+      if (Debug)
+         printf("SPU %u: Wait for cmd...\n", init.id);
+
+      /* read/wait from mailbox */
+      opcode = (unsigned int) spu_read_in_mbox();
+
+      if (Debug)
+         printf("SPU %u: got cmd 0x%x\n", init.id, opcode);
+
+      /* command payload */
+      mfc_get(&cmd,  /* dest */
+              (unsigned int) init.cmd, /* src */
+              sizeof(struct cell_command), /* bytes */
+              tag,
+              0, /* tid */
+              0  /* rid */);
+      wait_on_mask( 1 << tag );
+
+      switch (opcode & CELL_CMD_OPCODE_MASK) {
+      case CELL_CMD_EXIT:
+         if (Debug)
+            printf("SPU %u: EXIT\n", init.id);
+         exitFlag = 1;
+         break;
+      case CELL_CMD_FRAMEBUFFER:
+         cmd_framebuffer(&cmd.fb);
+         break;
+      case CELL_CMD_CLEAR_SURFACE:
+         cmd_clear_surface(&cmd.clear);
+         break;
+      case CELL_CMD_RENDER:
+         cmd_render(&cmd.render);
+         break;
+      case CELL_CMD_BATCH:
+         cmd_batch(opcode);
+         break;
+      case CELL_CMD_FINISH:
+         cmd_finish();
+         break;
+      default:
+         printf("Bad opcode!\n");
+      }
+
+   }
+
+   if (Debug)
+      printf("SPU %u: Exit main loop\n", init.id);
+}
+
+
+
+static void
+one_time_init(void)
+{
+   memset(tile_status, TILE_STATUS_DEFINED, sizeof(tile_status));
+   memset(tile_status_z, TILE_STATUS_DEFINED, sizeof(tile_status_z));
+}
+
+
+/**
+ * SPE entrypoint.
+ * Note: example programs declare params as 'unsigned long long' but
+ * that doesn't work.
+ */
+int
+main(unsigned long speid, unsigned long argp)
+{
+   int tag = 0;
+
+   (void) speid;
+
+   one_time_init();
+
+   if (Debug)
+      printf("SPU: main() speid=%lu\n", speid);
+
+   mfc_get(&init,  /* dest */
+           (unsigned int) argp, /* src */
+           sizeof(struct cell_init_info), /* bytes */
+           tag,
+           0, /* tid */
+           0  /* rid */);
+   wait_on_mask( 1 << tag );
+
+
+   main_loop();
+
+   return 0;
+}
diff --git a/src/mesa/pipe/cell/spu/spu_main.h b/src/mesa/pipe/cell/spu/spu_main.h
new file mode 100644
index 00000000000..9ef8cf07093
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_main.h
@@ -0,0 +1,78 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+#ifndef SPU_MAIN_H
+#define SPU_MAIN_H
+
+
+#include "pipe/cell/common.h"
+
+
+extern volatile struct cell_init_info init;
+
+struct framebuffer {
+   void *color_start;              /**< addr of color surface in main memory */
+   void *depth_start;              /**< addr of depth surface in main memory */
+   enum pipe_format color_format;
+   enum pipe_format depth_format;
+   uint width, height;             /**< size in pixels */
+   uint width_tiles, height_tiles; /**< width and height in tiles */
+
+   uint color_clear_value;
+   uint depth_clear_value;
+};
+
+/* XXX Collect these globals in a struct: */
+
+extern struct framebuffer fb;
+
+
+/* DMA TAGS */
+
+#define TAG_SURFACE_CLEAR     10
+#define TAG_VERTEX_BUFFER     11
+#define TAG_READ_TILE_COLOR   12
+#define TAG_READ_TILE_Z       13
+#define TAG_WRITE_TILE_COLOR  14
+#define TAG_WRITE_TILE_Z      15
+#define TAG_INDEX_BUFFER      16
+#define TAG_BATCH_BUFFER      17
+
+/** The standard assert macro doesn't seem to work on SPUs */
+#define ASSERT(x) \
+   if (!(x)) { \
+      fprintf(stderr, "SPU %d: %s:%d: %s(): assertion %s failed.\n", \
+              init.id, __FILE__, __LINE__, __FUNCTION__, #x); \
+      exit(1); \
+   }
+
+
+void
+wait_on_mask(unsigned tag);
+
+
+#endif /* SPU_MAIN_H */
diff --git a/src/mesa/pipe/cell/spu/spu_tile.c b/src/mesa/pipe/cell/spu/spu_tile.c
new file mode 100644
index 00000000000..8fd9ab2905b
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_tile.c
@@ -0,0 +1,115 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+
+
+#include "spu_tile.h"
+
+
+
+uint ctile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+ushort ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+
+ubyte tile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+ubyte tile_status_z[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+
+
+
+void
+get_tile(const struct framebuffer *fb, uint tx, uint ty, uint *tile,
+         int tag, int zBuf)
+{
+   const uint offset = ty * fb->width_tiles + tx;
+   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? 2 : 4);
+   const ubyte *src = zBuf ? fb->depth_start : fb->color_start;
+
+   src += offset * bytesPerTile;
+
+   ASSERT(tx < fb->width_tiles);
+   ASSERT(ty < fb->height_tiles);
+   ASSERT_ALIGN16(tile);
+   /*
+   printf("get_tile:  dest: %p  src: 0x%x  size: %d\n",
+          tile, (unsigned int) src, bytesPerTile);
+   */
+   mfc_get(tile,  /* dest in local memory */
+           (unsigned int) src, /* src in main memory */
+           bytesPerTile,
+           tag,
+           0, /* tid */
+           0  /* rid */);
+}
+
+
+void
+put_tile(const struct framebuffer *fb, uint tx, uint ty, const uint *tile,
+         int tag, int zBuf)
+{
+   const uint offset = ty * fb->width_tiles + tx;
+   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? 2 : 4);
+   ubyte *dst = zBuf ? fb->depth_start : fb->color_start;
+
+   dst += offset * bytesPerTile;
+
+   ASSERT(tx < fb->width_tiles);
+   ASSERT(ty < fb->height_tiles);
+   ASSERT_ALIGN16(tile);
+   /*
+   printf("put_tile:  src: %p  dst: 0x%x  size: %d\n",
+          tile, (unsigned int) dst, bytesPerTile);
+   */
+   mfc_put((void *) tile,  /* src in local memory */
+           (unsigned int) dst,  /* dst in main memory */
+           bytesPerTile,
+           tag,
+           0, /* tid */
+           0  /* rid */);
+}
+
+
+void
+clear_tile(uint tile[TILE_SIZE][TILE_SIZE], uint value)
+{
+   uint i, j;
+   for (i = 0; i < TILE_SIZE; i++) {
+      for (j = 0; j < TILE_SIZE; j++) {
+         tile[i][j] = value;
+      }
+   }
+}
+
+void
+clear_tile_z(ushort tile[TILE_SIZE][TILE_SIZE], uint value)
+{
+   uint i, j;
+   for (i = 0; i < TILE_SIZE; i++) {
+      for (j = 0; j < TILE_SIZE; j++) {
+         tile[i][j] = value;
+      }
+   }
+}
+
diff --git a/src/mesa/pipe/cell/spu/spu_tile.h b/src/mesa/pipe/cell/spu/spu_tile.h
new file mode 100644
index 00000000000..2ad469b5847
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_tile.h
@@ -0,0 +1,69 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+#ifndef SPU_TILE_H
+#define SPU_TILE_H
+
+
+#include <libmisc.h>
+#include <spu_mfcio.h>
+#include "spu_main.h"
+#include "pipe/cell/common.h"
+
+
+#define MAX_WIDTH 1024
+#define MAX_HEIGHT 1024
+
+
+extern uint ctile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+extern ushort ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+
+
+#define TILE_STATUS_CLEAR   1
+#define TILE_STATUS_DEFINED 2  /**< defined pixel data */
+#define TILE_STATUS_DIRTY   3  /**< modified, but not put back yet */
+
+extern ubyte tile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+extern ubyte tile_status_z[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+
+
+void
+get_tile(const struct framebuffer *fb, uint tx, uint ty, uint *tile,
+          int tag, int zBuf);
+
+void
+put_tile(const struct framebuffer *fb, uint tx, uint ty, const uint *tile,
+         int tag, int zBuf);
+
+void
+clear_tile(uint tile[TILE_SIZE][TILE_SIZE], uint value);
+
+void
+clear_tile_z(ushort tile[TILE_SIZE][TILE_SIZE], uint value);
+
+
+#endif /* SPU_TILE_H */
diff --git a/src/mesa/pipe/cell/spu/spu_tri.c b/src/mesa/pipe/cell/spu/spu_tri.c
new file mode 100644
index 00000000000..af24e42435d
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_tri.c
@@ -0,0 +1,892 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+/**
+ * Triangle rendering within a tile.
+ */
+
+#include "pipe/p_compiler.h"
+#include "pipe/p_format.h"
+#include "pipe/p_util.h"
+#include "spu_main.h"
+#include "spu_tile.h"
+#include "spu_tri.h"
+
+
+
+/**
+ * Simplified types taken from other parts of Gallium
+ */
+
+struct vertex_header {
+   float data[2][4];  /* pos and color */
+};
+
+struct prim_header {
+   struct vertex_header *v[3];
+};
+
+
+
+
+#if 1
+
+/* XXX fix this */
+#undef CEILF
+#define CEILF(X) ((float) (int) ((X) + 0.99999))
+
+
+#define QUAD_TOP_LEFT     0
+#define QUAD_TOP_RIGHT    1
+#define QUAD_BOTTOM_LEFT  2
+#define QUAD_BOTTOM_RIGHT 3
+#define MASK_TOP_LEFT     (1 << QUAD_TOP_LEFT)
+#define MASK_TOP_RIGHT    (1 << QUAD_TOP_RIGHT)
+#define MASK_BOTTOM_LEFT  (1 << QUAD_BOTTOM_LEFT)
+#define MASK_BOTTOM_RIGHT (1 << QUAD_BOTTOM_RIGHT)
+#define MASK_ALL          0xf
+
+#define PIPE_MAX_SHADER_INPUTS 8 /* XXX temp */
+
+#endif
+
+
+#define DEBUG_VERTS 0
+
+/**
+ * Triangle edge info
+ */
+struct edge {
+   float dx;		/**< X(v1) - X(v0), used only during setup */
+   float dy;		/**< Y(v1) - Y(v0), used only during setup */
+   float dxdy;		/**< dx/dy */
+   float sx, sy;	/**< first sample point coord */
+   int lines;		/**< number of lines on this edge */
+};
+
+
+struct interp_coef
+{
+   float a0[4];
+   float dadx[4];
+   float dady[4];
+};
+
+/**
+ * Triangle setup info (derived from draw_stage).
+ * Also used for line drawing (taking some liberties).
+ */
+struct setup_stage {
+
+   /* Vertices are just an array of floats making up each attribute in
+    * turn.  Currently fixed at 4 floats, but should change in time.
+    * Codegen will help cope with this.
+    */
+   const struct vertex_header *vmax;
+   const struct vertex_header *vmid;
+   const struct vertex_header *vmin;
+   const struct vertex_header *vprovoke;
+
+   struct edge ebot;
+   struct edge etop;
+   struct edge emaj;
+
+   float oneoverarea;
+
+   uint tx, ty;
+
+   int cliprect_minx, cliprect_maxx, cliprect_miny, cliprect_maxy;
+
+#if 0
+   struct tgsi_interp_coef coef[PIPE_MAX_SHADER_INPUTS];
+#else
+   struct interp_coef coef[PIPE_MAX_SHADER_INPUTS];
+#endif
+
+#if 0
+   struct quad_header quad; 
+#endif
+
+   struct {
+      int left[2];   /**< [0] = row0, [1] = row1 */
+      int right[2];
+      int y;
+      unsigned y_flags;
+      unsigned mask;     /**< mask of MASK_BOTTOM/TOP_LEFT/RIGHT bits */
+   } span;
+};
+
+
+#if 0
+/**
+ * Basically a cast wrapper.
+ */
+static INLINE struct setup_stage *setup_stage( struct draw_stage *stage )
+{
+   return (struct setup_stage *)stage;
+}
+#endif
+
+#if 0
+/**
+ * Clip setup->quad against the scissor/surface bounds.
+ */
+static INLINE void
+quad_clip(struct setup_stage *setup)
+{
+   const struct pipe_scissor_state *cliprect = &setup->softpipe->cliprect;
+   const int minx = (int) cliprect->minx;
+   const int maxx = (int) cliprect->maxx;
+   const int miny = (int) cliprect->miny;
+   const int maxy = (int) cliprect->maxy;
+
+   if (setup->quad.x0 >= maxx ||
+       setup->quad.y0 >= maxy ||
+       setup->quad.x0 + 1 < minx ||
+       setup->quad.y0 + 1 < miny) {
+      /* totally clipped */
+      setup->quad.mask = 0x0;
+      return;
+   }
+   if (setup->quad.x0 < minx)
+      setup->quad.mask &= (MASK_BOTTOM_RIGHT | MASK_TOP_RIGHT);
+   if (setup->quad.y0 < miny)
+      setup->quad.mask &= (MASK_BOTTOM_LEFT | MASK_BOTTOM_RIGHT);
+   if (setup->quad.x0 == maxx - 1)
+      setup->quad.mask &= (MASK_BOTTOM_LEFT | MASK_TOP_LEFT);
+   if (setup->quad.y0 == maxy - 1)
+      setup->quad.mask &= (MASK_TOP_LEFT | MASK_TOP_RIGHT);
+}
+#endif
+
+#if 0
+/**
+ * Emit a quad (pass to next stage) with clipping.
+ */
+static INLINE void
+clip_emit_quad(struct setup_stage *setup)
+{
+   quad_clip(setup);
+   if (setup->quad.mask) {
+      struct softpipe_context *sp = setup->softpipe;
+      sp->quad.first->run(sp->quad.first, &setup->quad);
+   }
+}
+#endif
+
+/**
+ * Evaluate attribute coefficients (plane equations) to compute
+ * attribute values for the four fragments in a quad.
+ * Eg: four colors will be compute.
+ */
+static INLINE void
+eval_coeff( struct setup_stage *setup, uint slot,
+            float x, float y, float result[4][4])
+{
+   uint i;
+   const float *dadx = setup->coef[slot].dadx;
+   const float *dady = setup->coef[slot].dady;
+
+   /* loop over XYZW comps */
+   for (i = 0; i < 4; i++) {
+      result[QUAD_TOP_LEFT][i] = setup->coef[slot].a0[i] + x * dadx[i] + y * dady[i];
+      result[QUAD_TOP_RIGHT][i] = result[0][i] + dadx[i];
+      result[QUAD_BOTTOM_LEFT][i] = result[0][i] + dady[i];
+      result[QUAD_BOTTOM_RIGHT][i] = result[0][i] + dadx[i] + dady[i];
+   }
+}
+
+
+static INLINE void
+eval_z( struct setup_stage *setup,
+        float x, float y, float result[4])
+{
+   uint slot = 0;
+   uint i = 2;
+   const float *dadx = setup->coef[slot].dadx;
+   const float *dady = setup->coef[slot].dady;
+
+   result[QUAD_TOP_LEFT] = setup->coef[slot].a0[i] + x * dadx[i] + y * dady[i];
+   result[QUAD_TOP_RIGHT] = result[0] + dadx[i];
+   result[QUAD_BOTTOM_LEFT] = result[0] + dady[i];
+   result[QUAD_BOTTOM_RIGHT] = result[0] + dadx[i] + dady[i];
+}
+
+
+static INLINE uint
+pack_color(const float color[4])
+{
+   uint r = (uint) (color[0] * 255.0);
+   uint g = (uint) (color[1] * 255.0);
+   uint b = (uint) (color[2] * 255.0);
+   uint a = (uint) (color[3] * 255.0);
+   switch (fb.color_format) {
+   case PIPE_FORMAT_A8R8G8B8_UNORM:
+      return (a << 24) | (r << 16) | (g << 8) | b;
+   case PIPE_FORMAT_B8G8R8A8_UNORM:
+      return (b << 24) | (g << 16) | (r << 8) | a;
+   default:
+      ASSERT(0);
+      return 0;
+   }
+}
+
+
+/**
+ * Emit a quad (pass to next stage).  No clipping is done.
+ */
+static INLINE void
+emit_quad( struct setup_stage *setup, int x, int y, unsigned mask )
+{
+#if 0
+   struct softpipe_context *sp = setup->softpipe;
+   setup->quad.x0 = x;
+   setup->quad.y0 = y;
+   setup->quad.mask = mask;
+   sp->quad.first->run(sp->quad.first, &setup->quad);
+#else
+   /* Cell: "write" quad fragments to the tile by setting prim color */
+   int ix = x - setup->cliprect_minx;
+   int iy = y - setup->cliprect_miny;
+   float colors[4][4];
+   uint z;
+
+   eval_coeff(setup, 1, (float) x, (float) y, colors);
+
+   if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+      float zvals[4];
+      eval_z(setup, (float) x, (float) y, zvals);
+
+      if (tile_status_z[setup->ty][setup->tx] == TILE_STATUS_CLEAR) {
+         /* now, _really_ clear the tile */
+         clear_tile_z(ztile, fb.depth_clear_value);
+      }
+      else {
+         /* make sure we've got the tile from main mem */
+         wait_on_mask(1 << TAG_READ_TILE_Z);
+      }
+      tile_status_z[setup->ty][setup->tx] = TILE_STATUS_DIRTY;
+
+      if (mask & MASK_TOP_LEFT) {
+         z = (uint) (zvals[0] * 65535.0);
+         if (z < ztile[iy][ix])
+            ztile[iy][ix] = z;
+         else
+            mask &= ~MASK_TOP_LEFT;
+      }
+
+      if (mask & MASK_TOP_RIGHT) {
+         z = (uint) (zvals[1] * 65535.0);
+         if (z < ztile[iy][ix+1])
+            ztile[iy][ix+1] = z;
+         else
+            mask &= ~MASK_TOP_RIGHT;
+      }
+
+      if (mask & MASK_BOTTOM_LEFT) {
+         z = (uint) (zvals[2] * 65535.0);
+         if (z < ztile[iy+1][ix])
+            ztile[iy+1][ix] = z;
+         else
+            mask &= ~MASK_BOTTOM_LEFT;
+      }
+
+      if (mask & MASK_BOTTOM_RIGHT) {
+         z = (uint) (zvals[3] * 65535.0);
+         if (z < ztile[iy+1][ix+1])
+            ztile[iy+1][ix+1] = z;
+         else
+            mask &= ~MASK_BOTTOM_RIGHT;
+      }
+   }
+
+   if (mask) {
+      if (tile_status[setup->ty][setup->tx] == TILE_STATUS_CLEAR) {
+         /* now, _really_ clear the tile */
+         clear_tile(ctile, fb.color_clear_value);
+      }
+      else {
+         /* make sure we've got the tile from main mem */
+         wait_on_mask(1 << TAG_READ_TILE_COLOR);
+      }
+      tile_status[setup->ty][setup->tx] = TILE_STATUS_DIRTY;
+
+      if (mask & MASK_TOP_LEFT)
+         ctile[iy][ix] = pack_color(colors[QUAD_TOP_LEFT]);
+      if (mask & MASK_TOP_RIGHT)
+         ctile[iy][ix+1] = pack_color(colors[QUAD_TOP_RIGHT]);
+      if (mask & MASK_BOTTOM_LEFT)
+         ctile[iy+1][ix] = pack_color(colors[QUAD_BOTTOM_LEFT]);
+      if (mask & MASK_BOTTOM_RIGHT)
+         ctile[iy+1][ix+1] = pack_color(colors[QUAD_BOTTOM_RIGHT]);
+   }
+#endif
+}
+
+
+/**
+ * Given an X or Y coordinate, return the block/quad coordinate that it
+ * belongs to.
+ */
+static INLINE int block( int x )
+{
+   return x & ~1;
+}
+
+
+/**
+ * Compute mask which indicates which pixels in the 2x2 quad are actually inside
+ * the triangle's bounds.
+ *
+ * this is pretty nasty...  may need to rework flush_spans again to
+ * fix it, if possible.
+ */
+static unsigned calculate_mask( struct setup_stage *setup, int x )
+{
+   unsigned mask = 0x0;
+
+   if (x >= setup->span.left[0] && x < setup->span.right[0]) 
+      mask |= MASK_TOP_LEFT;
+
+   if (x >= setup->span.left[1] && x < setup->span.right[1]) 
+      mask |= MASK_BOTTOM_LEFT;
+      
+   if (x+1 >= setup->span.left[0] && x+1 < setup->span.right[0]) 
+      mask |= MASK_TOP_RIGHT;
+
+   if (x+1 >= setup->span.left[1] && x+1 < setup->span.right[1]) 
+      mask |= MASK_BOTTOM_RIGHT;
+
+   return mask;
+}
+
+
+/**
+ * Render a horizontal span of quads
+ */
+static void flush_spans( struct setup_stage *setup )
+{
+   int minleft, maxright;
+   int x;
+
+   switch (setup->span.y_flags) {
+   case 0x3:
+      /* both odd and even lines written (both quad rows) */
+      minleft = MIN2(setup->span.left[0], setup->span.left[1]);
+      maxright = MAX2(setup->span.right[0], setup->span.right[1]);
+      break;
+
+   case 0x1:
+      /* only even line written (quad top row) */
+      minleft = setup->span.left[0];
+      maxright = setup->span.right[0];
+      break;
+
+   case 0x2:
+      /* only odd line written (quad bottom row) */
+      minleft = setup->span.left[1];
+      maxright = setup->span.right[1];
+      break;
+
+   default:
+      return;
+   }
+
+   /* XXX this loop could be moved into the above switch cases and
+    * calculate_mask() could be simplified a bit...
+    */
+   for (x = block(minleft); x <= block(maxright); x += 2) {
+      emit_quad( setup, x, setup->span.y, 
+                 calculate_mask( setup, x ) );
+   }
+
+   setup->span.y = 0;
+   setup->span.y_flags = 0;
+   setup->span.right[0] = 0;
+   setup->span.right[1] = 0;
+}
+
+#if DEBUG_VERTS
+static void print_vertex(const struct setup_stage *setup,
+                         const struct vertex_header *v)
+{
+   int i;
+   fprintf(stderr, "Vertex: (%p)\n", v);
+   for (i = 0; i < setup->quad.nr_attrs; i++) {
+      fprintf(stderr, "  %d: %f %f %f %f\n",  i, 
+              v->data[i][0], v->data[i][1], v->data[i][2], v->data[i][3]);
+   }
+}
+#endif
+
+static boolean setup_sort_vertices( struct setup_stage *setup,
+				      const struct prim_header *prim )
+{
+   const struct vertex_header *v0 = prim->v[0];
+   const struct vertex_header *v1 = prim->v[1];
+   const struct vertex_header *v2 = prim->v[2];
+
+#if DEBUG_VERTS
+   fprintf(stderr, "Triangle:\n");
+   print_vertex(setup, v0);
+   print_vertex(setup, v1);
+   print_vertex(setup, v2);
+#endif
+
+   setup->vprovoke = v2;
+
+   /* determine bottom to top order of vertices */
+   {
+      float y0 = v0->data[0][1];
+      float y1 = v1->data[0][1];
+      float y2 = v2->data[0][1];
+      if (y0 <= y1) {
+	 if (y1 <= y2) {
+	    /* y0<=y1<=y2 */
+	    setup->vmin = v0;   
+	    setup->vmid = v1;   
+	    setup->vmax = v2;
+	 }
+	 else if (y2 <= y0) {
+	    /* y2<=y0<=y1 */
+	    setup->vmin = v2;   
+	    setup->vmid = v0;   
+	    setup->vmax = v1;   
+	 }
+	 else {
+	    /* y0<=y2<=y1 */
+	    setup->vmin = v0;   
+	    setup->vmid = v2;   
+	    setup->vmax = v1;  
+	 }
+      }
+      else {
+	 if (y0 <= y2) {
+	    /* y1<=y0<=y2 */
+	    setup->vmin = v1;   
+	    setup->vmid = v0;   
+	    setup->vmax = v2;  
+	 }
+	 else if (y2 <= y1) {
+	    /* y2<=y1<=y0 */
+	    setup->vmin = v2;   
+	    setup->vmid = v1;   
+	    setup->vmax = v0;  
+	 }
+	 else {
+	    /* y1<=y2<=y0 */
+	    setup->vmin = v1;   
+	    setup->vmid = v2;   
+	    setup->vmax = v0;
+	 }
+      }
+   }
+
+   /* Check if triangle is completely outside the tile bounds */
+   if (setup->vmin->data[0][1] > setup->cliprect_maxy)
+      return FALSE;
+   if (setup->vmax->data[0][1] < setup->cliprect_miny)
+      return FALSE;
+   if (setup->vmin->data[0][0] < setup->cliprect_minx &&
+       setup->vmid->data[0][0] < setup->cliprect_minx &&
+       setup->vmax->data[0][0] < setup->cliprect_minx)
+      return FALSE;
+   if (setup->vmin->data[0][0] > setup->cliprect_maxx &&
+       setup->vmid->data[0][0] > setup->cliprect_maxx &&
+       setup->vmax->data[0][0] > setup->cliprect_maxx)
+      return FALSE;
+
+   setup->ebot.dx = setup->vmid->data[0][0] - setup->vmin->data[0][0];
+   setup->ebot.dy = setup->vmid->data[0][1] - setup->vmin->data[0][1];
+   setup->emaj.dx = setup->vmax->data[0][0] - setup->vmin->data[0][0];
+   setup->emaj.dy = setup->vmax->data[0][1] - setup->vmin->data[0][1];
+   setup->etop.dx = setup->vmax->data[0][0] - setup->vmid->data[0][0];
+   setup->etop.dy = setup->vmax->data[0][1] - setup->vmid->data[0][1];
+
+   /*
+    * Compute triangle's area.  Use 1/area to compute partial
+    * derivatives of attributes later.
+    *
+    * The area will be the same as prim->det, but the sign may be
+    * different depending on how the vertices get sorted above.
+    *
+    * To determine whether the primitive is front or back facing we
+    * use the prim->det value because its sign is correct.
+    */
+   {
+      const float area = (setup->emaj.dx * setup->ebot.dy - 
+			    setup->ebot.dx * setup->emaj.dy);
+
+      setup->oneoverarea = 1.0f / area;
+      /*
+      _mesa_printf("%s one-over-area %f  area %f  det %f\n",
+                   __FUNCTION__, setup->oneoverarea, area, prim->det );
+      */
+   }
+
+#if 0
+   /* We need to know if this is a front or back-facing triangle for:
+    *  - the GLSL gl_FrontFacing fragment attribute (bool)
+    *  - two-sided stencil test
+    */
+   setup->quad.facing = (prim->det > 0.0) ^ (setup->softpipe->rasterizer->front_winding == PIPE_WINDING_CW);
+#endif
+
+   return TRUE;
+}
+
+
+#if 0
+/**
+ * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
+ * The value value comes from vertex->data[slot][i].
+ * The result will be put into setup->coef[slot].a0[i].
+ * \param slot  which attribute slot 
+ * \param i  which component of the slot (0..3)
+ */
+static void const_coeff( struct setup_stage *setup,
+			 unsigned slot,
+			 unsigned i )
+{
+   assert(slot < PIPE_MAX_SHADER_INPUTS);
+   assert(i <= 3);
+
+   setup->coef[slot].dadx[i] = 0;
+   setup->coef[slot].dady[i] = 0;
+
+   /* need provoking vertex info!
+    */
+   setup->coef[slot].a0[i] = setup->vprovoke->data[slot][i];
+}
+#endif
+
+
+/**
+ * Compute a0, dadx and dady for a linearly interpolated coefficient,
+ * for a triangle.
+ */
+static void tri_linear_coeff( struct setup_stage *setup,
+                              uint slot, uint firstComp, uint lastComp )
+{
+   uint i;
+   for (i = firstComp; i < lastComp; i++) {
+      float botda = setup->vmid->data[slot][i] - setup->vmin->data[slot][i];
+      float majda = setup->vmax->data[slot][i] - setup->vmin->data[slot][i];
+      float a = setup->ebot.dy * majda - botda * setup->emaj.dy;
+      float b = setup->emaj.dx * botda - majda * setup->ebot.dx;
+   
+      ASSERT(slot < PIPE_MAX_SHADER_INPUTS);
+
+      setup->coef[slot].dadx[i] = a * setup->oneoverarea;
+      setup->coef[slot].dady[i] = b * setup->oneoverarea;
+
+      /* calculate a0 as the value which would be sampled for the
+       * fragment at (0,0), taking into account that we want to sample at
+       * pixel centers, in other words (0.5, 0.5).
+       *
+       * this is neat but unfortunately not a good way to do things for
+       * triangles with very large values of dadx or dady as it will
+       * result in the subtraction and re-addition from a0 of a very
+       * large number, which means we'll end up loosing a lot of the
+       * fractional bits and precision from a0.  the way to fix this is
+       * to define a0 as the sample at a pixel center somewhere near vmin
+       * instead - i'll switch to this later.
+       */
+      setup->coef[slot].a0[i] = (setup->vmin->data[slot][i] - 
+                                 (setup->coef[slot].dadx[i] * (setup->vmin->data[0][0] - 0.5f) + 
+                                  setup->coef[slot].dady[i] * (setup->vmin->data[0][1] - 0.5f)));
+   }
+
+   /*
+   _mesa_printf("attr[%d].%c: %f dx:%f dy:%f\n",
+		slot, "xyzw"[i], 
+		setup->coef[slot].a0[i],
+		setup->coef[slot].dadx[i],
+		setup->coef[slot].dady[i]);
+   */
+}
+
+
+#if 0
+/**
+ * Compute a0, dadx and dady for a perspective-corrected interpolant,
+ * for a triangle.
+ * We basically multiply the vertex value by 1/w before computing
+ * the plane coefficients (a0, dadx, dady).
+ * Later, when we compute the value at a particular fragment position we'll
+ * divide the interpolated value by the interpolated W at that fragment.
+ */
+static void tri_persp_coeff( struct setup_stage *setup,
+                             unsigned slot,
+                             unsigned i )
+{
+   /* premultiply by 1/w:
+    */
+   float mina = setup->vmin->data[slot][i] * setup->vmin->data[0][3];
+   float mida = setup->vmid->data[slot][i] * setup->vmid->data[0][3];
+   float maxa = setup->vmax->data[slot][i] * setup->vmax->data[0][3];
+
+   float botda = mida - mina;
+   float majda = maxa - mina;
+   float a = setup->ebot.dy * majda - botda * setup->emaj.dy;
+   float b = setup->emaj.dx * botda - majda * setup->ebot.dx;
+      
+   /*
+   printf("tri persp %d,%d: %f %f %f\n", slot, i,
+          setup->vmin->data[slot][i],
+          setup->vmid->data[slot][i],
+          setup->vmax->data[slot][i]
+          );
+   */
+
+   assert(slot < PIPE_MAX_SHADER_INPUTS);
+   assert(i <= 3);
+
+   setup->coef[slot].dadx[i] = a * setup->oneoverarea;
+   setup->coef[slot].dady[i] = b * setup->oneoverarea;
+   setup->coef[slot].a0[i] = (mina - 
+			    (setup->coef[slot].dadx[i] * (setup->vmin->data[0][0] - 0.5f) + 
+			     setup->coef[slot].dady[i] * (setup->vmin->data[0][1] - 0.5f)));
+}
+#endif
+
+
+/**
+ * Compute the setup->coef[] array dadx, dady, a0 values.
+ * Must be called after setup->vmin,vmid,vmax,vprovoke are initialized.
+ */
+static void setup_tri_coefficients( struct setup_stage *setup )
+{
+#if 0
+   const enum interp_mode *interp = setup->softpipe->vertex_info.interp_mode;
+   unsigned slot, j;
+
+   /* z and w are done by linear interpolation:
+    */
+   tri_linear_coeff(setup, 0, 2);
+   tri_linear_coeff(setup, 0, 3);
+
+   /* setup interpolation for all the remaining attributes:
+    */
+   for (slot = 1; slot < setup->quad.nr_attrs; slot++) {
+      switch (interp[slot]) {
+      case INTERP_CONSTANT:
+	 for (j = 0; j < NUM_CHANNELS; j++)
+	    const_coeff(setup, slot, j);
+	 break;
+      
+      case INTERP_LINEAR:
+	 for (j = 0; j < NUM_CHANNELS; j++)
+	    tri_linear_coeff(setup, slot, j);
+	 break;
+
+      case INTERP_PERSPECTIVE:
+	 for (j = 0; j < NUM_CHANNELS; j++)
+	    tri_persp_coeff(setup, slot, j);
+	 break;
+
+      default:
+         /* invalid interp mode */
+         assert(0);
+      }
+   }
+#else
+   tri_linear_coeff(setup, 0, 2, 3);  /* slot 0, z */
+   tri_linear_coeff(setup, 1, 0, 4);  /* slot 1, color */
+#endif
+}
+
+
+static void setup_tri_edges( struct setup_stage *setup )
+{
+   float vmin_x = setup->vmin->data[0][0] + 0.5f;
+   float vmid_x = setup->vmid->data[0][0] + 0.5f;
+
+   float vmin_y = setup->vmin->data[0][1] - 0.5f;
+   float vmid_y = setup->vmid->data[0][1] - 0.5f;
+   float vmax_y = setup->vmax->data[0][1] - 0.5f;
+
+   setup->emaj.sy = CEILF(vmin_y);
+   setup->emaj.lines = (int) CEILF(vmax_y - setup->emaj.sy);
+   setup->emaj.dxdy = setup->emaj.dx / setup->emaj.dy;
+   setup->emaj.sx = vmin_x + (setup->emaj.sy - vmin_y) * setup->emaj.dxdy;
+
+   setup->etop.sy = CEILF(vmid_y);
+   setup->etop.lines = (int) CEILF(vmax_y - setup->etop.sy);
+   setup->etop.dxdy = setup->etop.dx / setup->etop.dy;
+   setup->etop.sx = vmid_x + (setup->etop.sy - vmid_y) * setup->etop.dxdy;
+
+   setup->ebot.sy = CEILF(vmin_y);
+   setup->ebot.lines = (int) CEILF(vmid_y - setup->ebot.sy);
+   setup->ebot.dxdy = setup->ebot.dx / setup->ebot.dy;
+   setup->ebot.sx = vmin_x + (setup->ebot.sy - vmin_y) * setup->ebot.dxdy;
+}
+
+
+/**
+ * Render the upper or lower half of a triangle.
+ * Scissoring/cliprect is applied here too.
+ */
+static void subtriangle( struct setup_stage *setup,
+			 struct edge *eleft,
+			 struct edge *eright,
+			 unsigned lines )
+{
+   const int minx = setup->cliprect_minx;
+   const int maxx = setup->cliprect_maxx;
+   const int miny = setup->cliprect_miny;
+   const int maxy = setup->cliprect_maxy;
+   int y, start_y, finish_y;
+   int sy = (int)eleft->sy;
+
+   ASSERT((int)eleft->sy == (int) eright->sy);
+
+   /* clip top/bottom */
+   start_y = sy;
+   finish_y = sy + lines;
+
+   if (start_y < miny)
+      start_y = miny;
+
+   if (finish_y > maxy)
+      finish_y = maxy;
+
+   start_y -= sy;
+   finish_y -= sy;
+
+   /*
+   _mesa_printf("%s %d %d\n", __FUNCTION__, start_y, finish_y);  
+   */
+
+   for (y = start_y; y < finish_y; y++) {
+
+      /* avoid accumulating adds as floats don't have the precision to
+       * accurately iterate large triangle edges that way.  luckily we
+       * can just multiply these days.
+       *
+       * this is all drowned out by the attribute interpolation anyway.
+       */
+      int left = (int)(eleft->sx + y * eleft->dxdy);
+      int right = (int)(eright->sx + y * eright->dxdy);
+
+      /* clip left/right */
+      if (left < minx)
+         left = minx;
+      if (right > maxx)
+         right = maxx;
+
+      if (left < right) {
+         int _y = sy + y;
+         if (block(_y) != setup->span.y) {
+            flush_spans(setup);
+            setup->span.y = block(_y);
+         }
+
+         setup->span.left[_y&1] = left;
+         setup->span.right[_y&1] = right;
+         setup->span.y_flags |= 1<<(_y&1);
+      }
+   }
+
+
+   /* save the values so that emaj can be restarted:
+    */
+   eleft->sx += lines * eleft->dxdy;
+   eright->sx += lines * eright->dxdy;
+   eleft->sy += lines;
+   eright->sy += lines;
+}
+
+
+/**
+ * Do setup for triangle rasterization, then render the triangle.
+ */
+static void
+setup_tri(struct setup_stage *setup, struct prim_header *prim)
+{
+   if (!setup_sort_vertices( setup, prim )) {
+      return; /* totally clipped */
+   }
+
+   setup_tri_coefficients( setup );
+   setup_tri_edges( setup );
+
+#if 0
+   setup->quad.prim = PRIM_TRI;
+#endif
+
+   setup->span.y = 0;
+   setup->span.y_flags = 0;
+   setup->span.right[0] = 0;
+   setup->span.right[1] = 0;
+   /*   setup->span.z_mode = tri_z_mode( setup->ctx ); */
+
+   /*   init_constant_attribs( setup ); */
+      
+   if (setup->oneoverarea < 0.0) {
+      /* emaj on left:
+       */
+      subtriangle( setup, &setup->emaj, &setup->ebot, setup->ebot.lines );
+      subtriangle( setup, &setup->emaj, &setup->etop, setup->etop.lines );
+   }
+   else {
+      /* emaj on right:
+       */
+      subtriangle( setup, &setup->ebot, &setup->emaj, setup->ebot.lines );
+      subtriangle( setup, &setup->etop, &setup->emaj, setup->etop.lines );
+   }
+
+   flush_spans( setup );
+}
+
+
+
+/**
+ * Draw triangle into tile at (tx, ty) (tile coords)
+ * The tile data should have already been fetched.
+ */
+void
+tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty)
+{
+   struct prim_header tri;
+   struct setup_stage setup;
+
+   tri.v[0] = (struct vertex_header *) v0;
+   tri.v[1] = (struct vertex_header *) v1;
+   tri.v[2] = (struct vertex_header *) v2;
+
+   setup.tx = tx;
+   setup.ty = ty;
+
+   /* set clipping bounds to tile bounds */
+   setup.cliprect_minx = tx * TILE_SIZE;
+   setup.cliprect_miny = ty * TILE_SIZE;
+   setup.cliprect_maxx = (tx + 1) * TILE_SIZE;
+   setup.cliprect_maxy = (ty + 1) * TILE_SIZE;
+
+   setup_tri(&setup, &tri);
+}
diff --git a/src/mesa/pipe/cell/spu/spu_tri.h b/src/mesa/pipe/cell/spu/spu_tri.h
new file mode 100644
index 00000000000..86c42b63394
--- /dev/null
+++ b/src/mesa/pipe/cell/spu/spu_tri.h
@@ -0,0 +1,37 @@
+/**************************************************************************
+ * 
+ * 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.
+ * 
+ **************************************************************************/
+
+
+#ifndef SPU_TRI_H
+#define SPU_TRI_H
+
+
+extern void
+tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty);
+
+
+#endif /* SPU_TRI_H */