From 53951531ae7bfd64afae1ae55aac7f6ebd3fe4f5 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 12:35:51 -0600
Subject: cell: propogate blend color to SPUs for the fallback fragment ops
 code

---
 src/gallium/drivers/cell/common.h | 4 ++++
 1 file changed, 4 insertions(+)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index b0169b8e32..3b5a25e165 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -118,12 +118,16 @@
 
 /**
  * Command to specify per-fragment operations state and generated code.
+ * Note that the dsa, blend, blend_color fields are really only needed
+ * for the fallback/C per-pixel code.  They're not used when we generate
+ * dynamic SPU fragment code (which is the normal case).
  */
 struct cell_command_fragment_ops
 {
    uint64_t opcode;      /**< CELL_CMD_STATE_FRAGMENT_OPS */
    struct pipe_depth_stencil_alpha_state dsa;
    struct pipe_blend_state blend;
+   struct pipe_blend_color blend_color;
    unsigned code[SPU_MAX_FRAGMENT_OPS_INSTS];
 };
 
-- 
cgit v1.2.3


From ddeec1ed10d6c12403fe8d30c072ea68f044db99 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 13:55:18 -0600
Subject: cell: simplify spu debug code

---
 src/gallium/drivers/cell/common.h           |  1 +
 src/gallium/drivers/cell/ppu/cell_context.c |  1 +
 src/gallium/drivers/cell/spu/spu_command.c  | 47 +++++++++++++----------------
 src/gallium/drivers/cell/spu/spu_debug.h    |  9 ------
 src/gallium/drivers/cell/spu/spu_main.c     |  9 +-----
 src/gallium/drivers/cell/spu/spu_main.h     | 15 +++++++--
 src/gallium/drivers/cell/spu/spu_render.c   |  7 +++--
 7 files changed, 41 insertions(+), 48 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 3b5a25e165..8ae78265f2 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -111,6 +111,7 @@
 #define CELL_DEBUG_SYNC                 (1 << 2)
 #define CELL_DEBUG_FRAGMENT_OPS         (1 << 3)
 #define CELL_DEBUG_FRAGMENT_OP_FALLBACK (1 << 4)
+#define CELL_DEBUG_CMD                  (1 << 5)
 
 /** Max instructions for doing per-fragment operations */
 #define SPU_MAX_FRAGMENT_OPS_INSTS 64
diff --git a/src/gallium/drivers/cell/ppu/cell_context.c b/src/gallium/drivers/cell/ppu/cell_context.c
index b66aa9c9d9..f8d5eef3ac 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.c
+++ b/src/gallium/drivers/cell/ppu/cell_context.c
@@ -93,6 +93,7 @@ static const struct debug_named_value cell_debug_flags[] = {
    {"sync", CELL_DEBUG_SYNC},      /**< SPUs do synchronous DMA */
    {"fragops", CELL_DEBUG_FRAGMENT_OPS}, /**< SPUs emit fragment ops debug messages*/
    {"fragopfallback", CELL_DEBUG_FRAGMENT_OP_FALLBACK}, /**< SPUs use reference implementation for fragment ops*/
+   {"cmd", CELL_DEBUG_CMD},       /**< SPUs dump command buffer info */
    {NULL, 0}
 };
 
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index b521c3aecf..ebbed3d1dc 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -44,7 +44,6 @@
 #include "spu_tile.h"
 #include "spu_vertex_shader.h"
 #include "spu_dcache.h"
-#include "spu_debug.h"
 #include "cell/common.h"
 
 
@@ -97,7 +96,7 @@ release_buffer(uint buffer)
 static void
 cmd_clear_surface(const struct cell_command_clear_surface *clear)
 {
-   DEBUG_PRINTF("CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value);
+   D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value);
 
    if (clear->surface == 0) {
       spu.fb.color_clear_value = clear->value;
@@ -165,14 +164,14 @@ cmd_clear_surface(const struct cell_command_clear_surface *clear)
 
 #endif /* CLEAR_OPT */
 
-   DEBUG_PRINTF("CLEAR SURF done\n");
+   D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF done\n");
 }
 
 
 static void
 cmd_release_verts(const struct cell_command_release_verts *release)
 {
-   DEBUG_PRINTF("RELEASE VERTS %u\n", release->vertex_buf);
+   D_PRINTF(CELL_DEBUG_CMD, "RELEASE VERTS %u\n", release->vertex_buf);
    ASSERT(release->vertex_buf != ~0U);
    release_buffer(release->vertex_buf);
 }
@@ -189,7 +188,7 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops)
 {
    static int warned = 0;
 
-   DEBUG_PRINTF("CMD_STATE_FRAGMENT_OPS\n");
+   D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n");
    /* Copy SPU code from batch buffer to spu buffer */
    memcpy(spu.fragment_ops_code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4);
    /* Copy state info (for fallback case only) */
@@ -229,7 +228,7 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops)
 static void
 cmd_state_fragment_program(const struct cell_command_fragment_program *fp)
 {
-   DEBUG_PRINTF("CMD_STATE_FRAGMENT_PROGRAM\n");
+   D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_PROGRAM\n");
    /* Copy SPU code from batch buffer to spu buffer */
    memcpy(spu.fragment_program_code, fp->code,
           SPU_MAX_FRAGMENT_PROGRAM_INSTS * 4);
@@ -247,11 +246,11 @@ cmd_state_fs_constants(const uint64_t *buffer, uint pos)
    const float *constants = (const float *) &buffer[pos + 2];
    uint i;
 
-   DEBUG_PRINTF("CMD_STATE_FS_CONSTANTS (%u)\n", num_const);
+   D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FS_CONSTANTS (%u)\n", num_const);
 
    /* Expand each float to float[4] for SOA execution */
    for (i = 0; i < num_const; i++) {
-      DEBUG_PRINTF("  const[%u] = %f\n", i, constants[i]);
+      D_PRINTF(CELL_DEBUG_CMD, "  const[%u] = %f\n", i, constants[i]);
       spu.constants[i] = spu_splats(constants[i]);
    }
 
@@ -263,7 +262,7 @@ cmd_state_fs_constants(const uint64_t *buffer, uint pos)
 static void
 cmd_state_framebuffer(const struct cell_command_framebuffer *cmd)
 {
-   DEBUG_PRINTF("FRAMEBUFFER: %d x %d at %p, cformat 0x%x  zformat 0x%x\n",
+   D_PRINTF(CELL_DEBUG_CMD, "FRAMEBUFFER: %d x %d at %p, cformat 0x%x  zformat 0x%x\n",
              cmd->width,
              cmd->height,
              cmd->color_start,
@@ -352,7 +351,7 @@ cmd_state_sampler(const struct cell_command_sampler *sampler)
 {
    uint unit = sampler->unit;
 
-   DEBUG_PRINTF("SAMPLER [%u]\n", unit);
+   D_PRINTF(CELL_DEBUG_CMD, "SAMPLER [%u]\n", unit);
 
    spu.sampler[unit] = sampler->state;
 
@@ -404,9 +403,7 @@ cmd_state_texture(const struct cell_command_texture *texture)
    const uint unit = texture->unit;
    uint i;
 
-   //if (spu.init.id==0) Debug=1;
-
-   DEBUG_PRINTF("TEXTURE [%u]\n", texture->unit);
+   D_PRINTF(CELL_DEBUG_CMD, "TEXTURE [%u]\n", texture->unit);
 
    spu.texture[unit].max_level = 0;
    spu.texture[unit].target = texture->target;
@@ -416,7 +413,7 @@ cmd_state_texture(const struct cell_command_texture *texture)
       uint height = texture->height[i];
       uint depth = texture->depth[i];
 
-      DEBUG_PRINTF("  LEVEL %u: at %p  size[0] %u x %u\n", i,
+      D_PRINTF(CELL_DEBUG_CMD, "  LEVEL %u: at %p  size[0] %u x %u\n", i,
              texture->start[i], texture->width[i], texture->height[i]);
 
       spu.texture[unit].level[i].start = texture->start[i];
@@ -438,15 +435,13 @@ cmd_state_texture(const struct cell_command_texture *texture)
    }
 
    update_tex_masks(&spu.texture[unit], &spu.sampler[unit], unit);
-
-   //Debug=0;
 }
 
 
 static void
 cmd_state_vertex_info(const struct vertex_info *vinfo)
 {
-   DEBUG_PRINTF("VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs);
+   D_PRINTF(CELL_DEBUG_CMD, "VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs);
    ASSERT(vinfo->num_attribs >= 1);
    ASSERT(vinfo->num_attribs <= 8);
    memcpy(&spu.vertex_info, vinfo, sizeof(*vinfo));
@@ -485,7 +480,7 @@ cmd_state_attrib_fetch(const struct cell_attribute_fetch_code *code)
 static void
 cmd_finish(void)
 {
-   DEBUG_PRINTF("FINISH\n");
+   D_PRINTF(CELL_DEBUG_CMD, "FINISH\n");
    really_clear_tiles(0);
    /* wait for all outstanding DMAs to finish */
    mfc_write_tag_mask(~0);
@@ -510,7 +505,7 @@ cmd_batch(uint opcode)
    const unsigned usize = size / sizeof(buffer[0]);
    uint pos;
 
-   DEBUG_PRINTF("BATCH buffer %u, len %u, from %p\n",
+   D_PRINTF(CELL_DEBUG_CMD, "BATCH buffer %u, len %u, from %p\n",
              buf, size, spu.init.buffers[buf]);
 
    ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH);
@@ -530,7 +525,7 @@ cmd_batch(uint opcode)
    wait_on_mask(1 << TAG_BATCH_BUFFER);
 
    /* Tell PPU we're done copying the buffer to local store */
-   DEBUG_PRINTF("release batch buf %u\n", buf);
+   D_PRINTF(CELL_DEBUG_CMD, "release batch buf %u\n", buf);
    release_buffer(buf);
 
    /*
@@ -663,7 +658,7 @@ cmd_batch(uint opcode)
       }
    }
 
-   DEBUG_PRINTF("BATCH complete\n");
+   D_PRINTF(CELL_DEBUG_CMD, "BATCH complete\n");
 }
 
 
@@ -677,7 +672,7 @@ command_loop(void)
    struct cell_command cmd;
    int exitFlag = 0;
 
-   DEBUG_PRINTF("Enter command loop\n");
+   D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n");
 
    ASSERT((sizeof(struct cell_command) & 0xf) == 0);
    ASSERT_ALIGN16(&cmd);
@@ -686,12 +681,12 @@ command_loop(void)
       unsigned opcode;
       int tag = 0;
 
-      DEBUG_PRINTF("Wait for cmd...\n");
+      D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n");
 
       /* read/wait from mailbox */
       opcode = (unsigned int) spu_read_in_mbox();
 
-      DEBUG_PRINTF("got cmd 0x%x\n", opcode);
+      D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode);
 
       /* command payload */
       mfc_get(&cmd,  /* dest */
@@ -708,7 +703,7 @@ command_loop(void)
 
       switch (opcode & CELL_CMD_OPCODE_MASK) {
       case CELL_CMD_EXIT:
-         DEBUG_PRINTF("EXIT\n");
+         D_PRINTF(CELL_DEBUG_CMD, "EXIT\n");
          exitFlag = 1;
          break;
       case CELL_CMD_VS_EXECUTE:
@@ -725,7 +720,7 @@ command_loop(void)
 
    }
 
-   DEBUG_PRINTF("Exit command loop\n");
+   D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n");
 
    spu_dcache_report();
 }
diff --git a/src/gallium/drivers/cell/spu/spu_debug.h b/src/gallium/drivers/cell/spu/spu_debug.h
index eeec052655..25653dcdcd 100644
--- a/src/gallium/drivers/cell/spu/spu_debug.h
+++ b/src/gallium/drivers/cell/spu/spu_debug.h
@@ -30,28 +30,19 @@
 #define SPU_DEBUG_H
 
 
-/* Set to 0 to disable all extraneous debugging code */
-#define DEBUG 1
-
 #if DEBUG
-extern boolean Debug;
-extern boolean force_fragment_ops_fallback;
 
 /* These debug macros use the unusual construction ", ##__VA_ARGS__"
  * which expands to the expected comma + args if variadic arguments
  * are supplied, but swallows the comma if there are no variadic
  * arguments (which avoids syntax errors that would otherwise occur).
  */
-#define DEBUG_PRINTF(format,...) \
-   if (Debug) \
-      printf("SPU %u: " format, spu.init.id, ##__VA_ARGS__)
 #define D_PRINTF(flag, format,...) \
    if (spu.init.debug_flags & (flag)) \
       printf("SPU %u: " format, spu.init.id, ##__VA_ARGS__)
 
 #else
 
-#define DEBUG_PRINTF(...)
 #define D_PRINTF(...)
 
 #endif
diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c
index 4becd0f92a..c8bb251905 100644
--- a/src/gallium/drivers/cell/spu/spu_main.c
+++ b/src/gallium/drivers/cell/spu/spu_main.c
@@ -40,7 +40,6 @@
 #include "spu_per_fragment_op.h"
 #include "spu_texture.h"
 //#include "spu_test.h"
-#include "spu_debug.h"
 #include "cell/common.h"
 
 
@@ -53,12 +52,6 @@ helpful headers:
 struct spu_global spu;
 
 
-#if DEBUG
-boolean Debug = FALSE;
-boolean force_fragment_ops_fallback = TRUE;
-#endif
-
-
 static void
 one_time_init(void)
 {
@@ -102,7 +95,7 @@ main(main_param_t speid, main_param_t argp)
 
    one_time_init();
 
-   DEBUG_PRINTF("main() speid=%lu\n", (unsigned long) speid);
+   D_PRINTF(CELL_DEBUG_CMD, "main() speid=%lu\n", (unsigned long) speid);
    D_PRINTF(CELL_DEBUG_FRAGMENT_OP_FALLBACK, "using fragment op fallback\n");
 
    /* get initialization data */
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index ca72baea8b..569b9e45d4 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -36,6 +36,19 @@
 #include "pipe/p_state.h"
 
 
+#if DEBUG
+/* These debug macros use the unusual construction ", ##__VA_ARGS__"
+ * which expands to the expected comma + args if variadic arguments
+ * are supplied, but swallows the comma if there are no variadic
+ * arguments (which avoids syntax errors that would otherwise occur).
+ */
+#define D_PRINTF(flag, format,...) \
+   if (spu.init.debug_flags & (flag)) \
+      printf("SPU %u: " format, spu.init.id, ##__VA_ARGS__)
+#else
+#define D_PRINTF(...)
+#endif
+
 
 #define MAX_WIDTH 1024
 #define MAX_HEIGHT 1024
@@ -187,8 +200,6 @@ struct spu_global
 
 
 extern struct spu_global spu;
-extern boolean Debug;
-
 
 
 
diff --git a/src/gallium/drivers/cell/spu/spu_render.c b/src/gallium/drivers/cell/spu/spu_render.c
index 82dbeb26b7..cfff19b6c0 100644
--- a/src/gallium/drivers/cell/spu/spu_render.c
+++ b/src/gallium/drivers/cell/spu/spu_render.c
@@ -177,7 +177,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
    uint i, j;
 
 
-   if (Debug) {
+#if 0
       printf("SPU %u: RENDER prim %u, num_vert=%u  num_ind=%u  "
              "inline_vert=%u\n",
              spu.init.id,
@@ -190,7 +190,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
       printf("       bound: %g, %g .. %g, %g\n",
              render->xmin, render->ymin, render->xmax, render->ymax);
       */
-   }
+#endif
 
    ASSERT(sizeof(*render) % 4 == 0);
    ASSERT(total_vertex_bytes % 16 == 0);
@@ -293,7 +293,8 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
       spu.ztile_status[ty][tx] = spu.cur_ztile_status;
    }
 
-   if (Debug)
+#if 0
       printf("SPU %u: RENDER done\n",
              spu.init.id);
+#endif
 }
-- 
cgit v1.2.3


From 79e96b3a77f7d5c7136b380abcc675c7242d0ffe Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 13:58:58 -0600
Subject: cell: move some CELL_MAX constants

---
 src/gallium/drivers/cell/common.h       |  6 +++++-
 src/gallium/drivers/cell/spu/spu_main.h | 11 ++---------
 2 files changed, 7 insertions(+), 10 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 8ae78265f2..d716a26175 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -68,6 +68,9 @@
 
 #define CELL_MAX_SAMPLERS 4
 #define CELL_MAX_TEXTURE_LEVELS 12  /* 2k x 2k */
+#define CELL_MAX_CONSTANTS 32  /**< number of float[4] constants */
+#define CELL_MAX_WIDTH 1024    /**< max framebuffer width */
+#define CELL_MAX_HEIGHT 1024   /**< max framebuffer width */
 
 #define TILE_SIZE 32
 
@@ -99,13 +102,14 @@
 #define CELL_CMD_VS_EXECUTE          22
 #define CELL_CMD_FLUSH_BUFFER_RANGE  23
 
-
+/** Command/batch buffers */
 #define CELL_NUM_BUFFERS 4
 #define CELL_BUFFER_SIZE (4*1024)  /**< 16KB would be the max */
 
 #define CELL_BUFFER_STATUS_FREE 10
 #define CELL_BUFFER_STATUS_USED 20
 
+/** Debug flags */
 #define CELL_DEBUG_CHECKER              (1 << 0)
 #define CELL_DEBUG_ASM                  (1 << 1)
 #define CELL_DEBUG_SYNC                 (1 << 2)
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index 569b9e45d4..f87495b72d 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -50,13 +50,6 @@
 #endif
 
 
-#define MAX_WIDTH 1024
-#define MAX_HEIGHT 1024
-
-
-#define CELL_MAX_CONSTANTS 32  /**< number of float[4] constants */
-
-
 /**
  * A tile is basically a TILE_SIZE x TILE_SIZE block of 4-byte pixels.
  * The data may be addressed through several different types.
@@ -175,8 +168,8 @@ struct spu_global
    ubyte cur_ctile_status, cur_ztile_status;
 
    /** Status of all tiles in framebuffer */
-   ubyte ctile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
-   ubyte ztile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+   ubyte ctile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+   ubyte ztile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
 
    /** Current fragment ops machine code, at 8-byte boundary */
    uint fragment_ops_code[SPU_MAX_FRAGMENT_OPS_INSTS] ALIGN8_ATTRIB;
-- 
cgit v1.2.3


From 0eb0b0a816764a323af7a8d2b5cb6792f886ce04 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 14:12:55 -0600
Subject: cell: remove some old, pre-batchbuffer stuff

---
 src/gallium/drivers/cell/common.h          | 14 --------------
 src/gallium/drivers/cell/ppu/cell_spu.c    |  5 +----
 src/gallium/drivers/cell/ppu/cell_spu.h    |  3 +--
 src/gallium/drivers/cell/spu/spu_command.c | 19 -------------------
 4 files changed, 2 insertions(+), 39 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index d716a26175..600f1b37a2 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -269,19 +269,6 @@ struct cell_command_texture
 };
 
 
-/** XXX unions don't seem to work */
-/* XXX this should go away; all commands should be placed in batch buffers */
-struct cell_command
-{
-#if 0
-   struct cell_command_framebuffer fb;
-   struct cell_command_clear_surface clear;
-   struct cell_command_render render;
-#endif
-   struct cell_command_vs vs;
-} ALIGN16_ATTRIB;
-
-
 #define MAX_SPU_FUNCTIONS 12
 /**
  * Used to tell the PPU about the address of particular functions in the
@@ -302,7 +289,6 @@ struct cell_init_info
    unsigned id;
    unsigned num_spus;
    unsigned debug_flags;  /**< mask of CELL_DEBUG_x flags */
-   struct cell_command *cmd;
 
    /** Buffers for command batches, vertex/index data */
    ubyte *buffers[CELL_NUM_BUFFERS];
diff --git a/src/gallium/drivers/cell/ppu/cell_spu.c b/src/gallium/drivers/cell/ppu/cell_spu.c
index df020c4146..90745da3d2 100644
--- a/src/gallium/drivers/cell/ppu/cell_spu.c
+++ b/src/gallium/drivers/cell/ppu/cell_spu.c
@@ -126,9 +126,6 @@ cell_start_spus(struct cell_context *cell)
 
    assert(cell->num_spus <= MAX_SPUS);
 
-   ASSERT_ALIGN16(&cell_global.command[0]);
-   ASSERT_ALIGN16(&cell_global.command[1]);
-
    ASSERT_ALIGN16(&cell_global.inits[0]);
    ASSERT_ALIGN16(&cell_global.inits[1]);
 
@@ -141,7 +138,7 @@ cell_start_spus(struct cell_context *cell)
       cell_global.inits[i].id = i;
       cell_global.inits[i].num_spus = cell->num_spus;
       cell_global.inits[i].debug_flags = cell->debug_flags;
-      cell_global.inits[i].cmd = &cell_global.command[i];
+
       for (j = 0; j < CELL_NUM_BUFFERS; j++) {
          cell_global.inits[i].buffers[j] = cell->buffer[j];
       }
diff --git a/src/gallium/drivers/cell/ppu/cell_spu.h b/src/gallium/drivers/cell/ppu/cell_spu.h
index 137f26612e..3443331b01 100644
--- a/src/gallium/drivers/cell/ppu/cell_spu.h
+++ b/src/gallium/drivers/cell/ppu/cell_spu.h
@@ -50,10 +50,9 @@ struct cell_global_info
    pthread_t spe_threads[MAX_SPUS];
 
    /**
-    * Data sent to SPUs
+    * Data sent to SPUs at start-up
     */
    struct cell_init_info inits[MAX_SPUS];
-   struct cell_command command[MAX_SPUS];
 };
 
 
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index ebbed3d1dc..4febd5385b 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -669,38 +669,19 @@ cmd_batch(uint opcode)
 void
 command_loop(void)
 {
-   struct cell_command cmd;
    int exitFlag = 0;
 
    D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n");
 
-   ASSERT((sizeof(struct cell_command) & 0xf) == 0);
-   ASSERT_ALIGN16(&cmd);
-
    while (!exitFlag) {
       unsigned opcode;
-      int tag = 0;
 
       D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n");
 
       /* read/wait from mailbox */
       opcode = (unsigned int) spu_read_in_mbox();
-
       D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode);
 
-      /* command payload */
-      mfc_get(&cmd,  /* dest */
-              (unsigned int) spu.init.cmd, /* src */
-              sizeof(struct cell_command), /* bytes */
-              tag,
-              0, /* tid */
-              0  /* rid */);
-      wait_on_mask( 1 << tag );
-
-      /*
-       * NOTE: most commands should be contained in a batch buffer
-       */
-
       switch (opcode & CELL_CMD_OPCODE_MASK) {
       case CELL_CMD_EXIT:
          D_PRINTF(CELL_DEBUG_CMD, "EXIT\n");
-- 
cgit v1.2.3


From 67f615681c569264eab1bc901473c86cfc54e480 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 14:18:51 -0600
Subject: cell: use CELL_MAX_SPUS consistently.

---
 src/gallium/drivers/cell/common.h       | 2 +-
 src/gallium/drivers/cell/ppu/cell_spu.c | 2 +-
 src/gallium/drivers/cell/ppu/cell_spu.h | 8 +++-----
 3 files changed, 5 insertions(+), 7 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 600f1b37a2..1f6f2d494b 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -64,7 +64,7 @@
 #define ROUNDUP16(k)  (((k) + 0xf) & ~0xf)
 
 
-#define CELL_MAX_SPUS 6
+#define CELL_MAX_SPUS 8
 
 #define CELL_MAX_SAMPLERS 4
 #define CELL_MAX_TEXTURE_LEVELS 12  /* 2k x 2k */
diff --git a/src/gallium/drivers/cell/ppu/cell_spu.c b/src/gallium/drivers/cell/ppu/cell_spu.c
index 90745da3d2..a6e268b362 100644
--- a/src/gallium/drivers/cell/ppu/cell_spu.c
+++ b/src/gallium/drivers/cell/ppu/cell_spu.c
@@ -124,7 +124,7 @@ cell_start_spus(struct cell_context *cell)
 
    one_time_init = TRUE;
 
-   assert(cell->num_spus <= MAX_SPUS);
+   assert(cell->num_spus <= CELL_MAX_SPUS);
 
    ASSERT_ALIGN16(&cell_global.inits[0]);
    ASSERT_ALIGN16(&cell_global.inits[1]);
diff --git a/src/gallium/drivers/cell/ppu/cell_spu.h b/src/gallium/drivers/cell/ppu/cell_spu.h
index 3443331b01..2e965c6301 100644
--- a/src/gallium/drivers/cell/ppu/cell_spu.h
+++ b/src/gallium/drivers/cell/ppu/cell_spu.h
@@ -36,8 +36,6 @@
 #include "cell_context.h"
 
 
-#define MAX_SPUS 8
-
 /**
  * Global vars, for now anyway.
  */
@@ -46,13 +44,13 @@ struct cell_global_info
    /**
     * SPU/SPE handles, etc
     */
-   spe_context_ptr_t spe_contexts[MAX_SPUS];
-   pthread_t spe_threads[MAX_SPUS];
+   spe_context_ptr_t spe_contexts[CELL_MAX_SPUS];
+   pthread_t spe_threads[CELL_MAX_SPUS];
 
    /**
     * Data sent to SPUs at start-up
     */
-   struct cell_init_info inits[MAX_SPUS];
+   struct cell_init_info inits[CELL_MAX_SPUS];
 };
 
 
-- 
cgit v1.2.3


From ec7d6c656178babdf143faa242f7a3df9d0bc22c Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 14:39:16 -0600
Subject: cell: send rasterizer state to SPUs in proper way, remove
 front_winding hack

---
 src/gallium/drivers/cell/common.h              | 18 ++++++++++++++----
 src/gallium/drivers/cell/ppu/cell_state_emit.c |  7 +++++++
 src/gallium/drivers/cell/ppu/cell_vbuf.c       |  1 -
 src/gallium/drivers/cell/spu/spu_command.c     |  8 ++++++++
 src/gallium/drivers/cell/spu/spu_main.h        |  1 +
 src/gallium/drivers/cell/spu/spu_render.c      |  2 +-
 src/gallium/drivers/cell/spu/spu_tri.c         |  4 ++--
 src/gallium/drivers/cell/spu/spu_tri.h         |  2 +-
 8 files changed, 34 insertions(+), 9 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 1f6f2d494b..0ff2c491fb 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -99,8 +99,9 @@
 #define CELL_CMD_STATE_FRAGMENT_PROGRAM 19
 #define CELL_CMD_STATE_ATTRIB_FETCH  20
 #define CELL_CMD_STATE_FS_CONSTANTS  21
-#define CELL_CMD_VS_EXECUTE          22
-#define CELL_CMD_FLUSH_BUFFER_RANGE  23
+#define CELL_CMD_STATE_RASTERIZER    22
+#define CELL_CMD_VS_EXECUTE          23
+#define CELL_CMD_FLUSH_BUFFER_RANGE  24
 
 /** Command/batch buffers */
 #define CELL_NUM_BUFFERS 4
@@ -156,13 +157,23 @@ struct cell_command_fragment_program
  */
 struct cell_command_framebuffer
 {
-   uint64_t opcode;     /**< CELL_CMD_FRAMEBUFFER */
+   uint64_t opcode;     /**< CELL_CMD_STATE_FRAMEBUFFER */
    int width, height;
    void *color_start, *depth_start;
    enum pipe_format color_format, depth_format;
 };
 
 
+/**
+ * Tell SPUs about rasterizer state.
+ */
+struct cell_command_rasterizer
+{
+   uint64_t opcode;    /**< CELL_CMD_STATE_RASTERIZER */
+   struct pipe_rasterizer_state rasterizer;
+};
+
+
 /**
  * Clear framebuffer to the given value/color.
  */
@@ -238,7 +249,6 @@ struct cell_command_render
    float xmin, ymin, xmax, ymax;  /* XXX another dummy field */
    uint min_index;
    boolean inline_verts;
-   uint front_winding; /* the rasterizer needs to be able to determine facing to apply front/back-facing stencil */
 };
 
 
diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c
index d2427584ba..e6387382f2 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_emit.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c
@@ -147,6 +147,13 @@ cell_emit_state(struct cell_context *cell)
 #endif
    }
 
+   if (cell->dirty & (CELL_NEW_RASTERIZER)) {
+      struct cell_command_rasterizer *rast =
+         cell_batch_alloc(cell, sizeof(*rast));
+      rast->opcode = CELL_CMD_STATE_RASTERIZER;
+      rast->rasterizer = *cell->rasterizer;
+   }
+
    if (cell->dirty & (CELL_NEW_FS)) {
       /* Send new fragment program to SPUs */
       struct cell_command_fragment_program *fp
diff --git a/src/gallium/drivers/cell/ppu/cell_vbuf.c b/src/gallium/drivers/cell/ppu/cell_vbuf.c
index 578ddf62dc..aa63435b93 100644
--- a/src/gallium/drivers/cell/ppu/cell_vbuf.c
+++ b/src/gallium/drivers/cell/ppu/cell_vbuf.c
@@ -214,7 +214,6 @@ cell_vbuf_draw(struct vbuf_render *vbr,
 
       render->opcode = CELL_CMD_RENDER;
       render->prim_type = cvbr->prim;
-      render->front_winding = cell->rasterizer->front_winding;
 
       render->num_indexes = nr_indices;
       render->min_index = min_index;
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index 4febd5385b..d2c282a022 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -583,6 +583,14 @@ cmd_batch(uint opcode)
       case CELL_CMD_STATE_FS_CONSTANTS:
          pos = cmd_state_fs_constants(buffer, pos);
          break;
+      case CELL_CMD_STATE_RASTERIZER:
+         {
+            struct cell_command_rasterizer *rast =
+               (struct cell_command_rasterizer *) &buffer[pos];
+            spu.rasterizer = rast->rasterizer;
+            pos += sizeof(*rast) / 8;
+         }
+         break;
       case CELL_CMD_STATE_SAMPLER:
          {
             struct cell_command_sampler *sampler
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index f87495b72d..4099e52699 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -153,6 +153,7 @@ struct spu_global
    struct pipe_blend_state blend;
    struct pipe_blend_color blend_color;
    struct pipe_sampler_state sampler[PIPE_MAX_SAMPLERS];
+   struct pipe_rasterizer_state rasterizer;
    struct spu_texture texture[PIPE_MAX_SAMPLERS];
    struct vertex_info vertex_info;
 
diff --git a/src/gallium/drivers/cell/spu/spu_render.c b/src/gallium/drivers/cell/spu/spu_render.c
index cfff19b6c0..75a7f75abc 100644
--- a/src/gallium/drivers/cell/spu/spu_render.c
+++ b/src/gallium/drivers/cell/spu/spu_render.c
@@ -279,7 +279,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
          v1 = (const float *) (vertices + indexes[j+1] * vertex_size);
          v2 = (const float *) (vertices + indexes[j+2] * vertex_size);
 
-         drawn += tri_draw(v0, v1, v2, tx, ty, render->front_winding);
+         drawn += tri_draw(v0, v1, v2, tx, ty);
       }
 
       //printf("SPU %u: drew %u of %u\n", spu.init.id, drawn, render->num_indexes/3);
diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c
index 2417db8960..1519b8cd7e 100644
--- a/src/gallium/drivers/cell/spu/spu_tri.c
+++ b/src/gallium/drivers/cell/spu/spu_tri.c
@@ -775,7 +775,7 @@ determinant(const float *v0, const float *v1, const float *v2)
  */
 boolean
 tri_draw(const float *v0, const float *v1, const float *v2,
-         uint tx, uint ty, uint front_winding)
+         uint tx, uint ty)
 {
    setup.tx = tx;
    setup.ty = ty;
@@ -790,7 +790,7 @@ tri_draw(const float *v0, const float *v1, const float *v2,
     * which will be needed for front/back-face stencil application
     */
    float det = determinant(v0, v1, v2);
-   setup.facing = (det > 0.0) ^ (front_winding == PIPE_WINDING_CW);
+   setup.facing = (det > 0.0) ^ (spu.rasterizer.front_winding == PIPE_WINDING_CW);
 
    if (!setup_sort_vertices((struct vertex_header *) v0,
                             (struct vertex_header *) v1,
diff --git a/src/gallium/drivers/cell/spu/spu_tri.h b/src/gallium/drivers/cell/spu/spu_tri.h
index abc3d35160..aa694dd7c9 100644
--- a/src/gallium/drivers/cell/spu/spu_tri.h
+++ b/src/gallium/drivers/cell/spu/spu_tri.h
@@ -31,7 +31,7 @@
 
 
 extern boolean
-tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty, uint front_winding);
+tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty);
 
 
 #endif /* SPU_TRI_H */
-- 
cgit v1.2.3


From 0116ee1d1c341726b6ed23c2dddc4515e8a34385 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 15 Oct 2008 20:46:43 -0600
Subject: cell: start some performance measurements

Use the spu_write_decrementer() and spu_read_decrementer() functions to
measure time.  Convert to milliseconds according to the system timebase value.
---
 src/gallium/drivers/cell/common.h          |  1 +
 src/gallium/drivers/cell/ppu/cell_spu.c    | 31 ++++++++++++++++++++++++++++++
 src/gallium/drivers/cell/spu/spu_command.c | 15 +++++++++++++++
 src/gallium/drivers/cell/spu/spu_render.c  |  9 ++++++++-
 4 files changed, 55 insertions(+), 1 deletion(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 0ff2c491fb..469d56cda8 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -299,6 +299,7 @@ struct cell_init_info
    unsigned id;
    unsigned num_spus;
    unsigned debug_flags;  /**< mask of CELL_DEBUG_x flags */
+   float inv_timebase;    /**< 1.0/timebase, for perf measurement */
 
    /** Buffers for command batches, vertex/index data */
    ubyte *buffers[CELL_NUM_BUFFERS];
diff --git a/src/gallium/drivers/cell/ppu/cell_spu.c b/src/gallium/drivers/cell/ppu/cell_spu.c
index a6e268b362..28e5e6d706 100644
--- a/src/gallium/drivers/cell/ppu/cell_spu.c
+++ b/src/gallium/drivers/cell/ppu/cell_spu.c
@@ -52,6 +52,35 @@ helpful headers:
 struct cell_global_info cell_global;
 
 
+/**
+ * Scan /proc/cpuinfo to determine the timebase for the system.
+ * This is used by the SPUs to convert 'decrementer' ticks to seconds.
+ * There may be a better way to get this value...
+ */
+static unsigned
+get_timebase(void)
+{
+   FILE *f = fopen("/proc/cpuinfo", "r");
+   unsigned timebase;
+
+   assert(f);
+   while (!feof(f)) {
+      char line[80];
+      fgets(line, sizeof(line), f);
+      if (strncmp(line, "timebase", 8) == 0) {
+         char *colon = strchr(line, ':');
+         if (colon) {
+            timebase = atoi(colon + 2);
+            break;
+         }
+      }
+   }
+   fclose(f);
+
+   return timebase;
+}
+
+
 /**
  * Write a 1-word message to the given SPE mailbox.
  */
@@ -115,6 +144,7 @@ cell_start_spus(struct cell_context *cell)
 {
    static boolean one_time_init = FALSE;
    uint i, j;
+   uint timebase = get_timebase();
 
    if (one_time_init) {
       fprintf(stderr, "PPU: Multiple rendering contexts not yet supported "
@@ -138,6 +168,7 @@ cell_start_spus(struct cell_context *cell)
       cell_global.inits[i].id = i;
       cell_global.inits[i].num_spus = cell->num_spus;
       cell_global.inits[i].debug_flags = cell->debug_flags;
+      cell_global.inits[i].inv_timebase = 1000.0f / timebase;
 
       for (j = 0; j < CELL_NUM_BUFFERS; j++) {
          cell_global.inits[i].buffers[j] = cell->buffer[j];
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index d2c282a022..57d265fef7 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -670,6 +670,8 @@ cmd_batch(uint opcode)
 }
 
 
+#define PERF 0
+
 
 /**
  * Main loop for SPEs: Get a command, execute it, repeat.
@@ -678,6 +680,7 @@ void
 command_loop(void)
 {
    int exitFlag = 0;
+   uint t0, t1;
 
    D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n");
 
@@ -686,10 +689,16 @@ command_loop(void)
 
       D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n");
 
+      if (PERF)
+         spu_write_decrementer(~0);
+
       /* read/wait from mailbox */
       opcode = (unsigned int) spu_read_in_mbox();
       D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode);
 
+      if (PERF)
+         t0 = spu_read_decrementer();
+
       switch (opcode & CELL_CMD_OPCODE_MASK) {
       case CELL_CMD_EXIT:
          D_PRINTF(CELL_DEBUG_CMD, "EXIT\n");
@@ -707,6 +716,12 @@ command_loop(void)
          printf("Bad opcode 0x%x!\n", opcode & CELL_CMD_OPCODE_MASK);
       }
 
+      if (PERF) {
+         t1 = spu_read_decrementer();
+         printf("wait mbox time: %gms   batch time: %gms\n",
+                (~0u - t0) * spu.init.inv_timebase,
+                (t0 - t1) * spu.init.inv_timebase);
+      }
    }
 
    D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n");
diff --git a/src/gallium/drivers/cell/spu/spu_render.c b/src/gallium/drivers/cell/spu/spu_render.c
index 802455bf79..5515bb55c9 100644
--- a/src/gallium/drivers/cell/spu/spu_render.c
+++ b/src/gallium/drivers/cell/spu/spu_render.c
@@ -175,6 +175,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
    const ubyte *vertices;
    const ushort *indexes;
    uint i, j;
+   uint num_tiles;
 
    D_PRINTF(CELL_DEBUG_CMD,
             "RENDER prim=%u num_vert=%u num_ind=%u inline_vert=%u\n",
@@ -242,6 +243,8 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
    wait_on_mask(1 << TAG_SURFACE_CLEAR); /* XXX temporary */
 
 
+   num_tiles = 0;
+
    /**
     ** loop over tiles, rendering tris
     **/
@@ -255,6 +258,8 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
       if (!my_tile(tx, ty))
          continue;
 
+      num_tiles++;
+
       spu.cur_ctile_status = spu.ctile_status[ty][tx];
       spu.cur_ztile_status = spu.ztile_status[ty][tx];
 
@@ -284,5 +289,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
       spu.ztile_status[ty][tx] = spu.cur_ztile_status;
    }
 
-   D_PRINTF(CELL_DEBUG_CMD, "RENDER done\n");
+   D_PRINTF(CELL_DEBUG_CMD,
+            "RENDER done (%u tiles hit)\n",
+            num_tiles);
 }
-- 
cgit v1.2.3


From 9fa8671c73fa44a95e2ea7fed6047bddb042796f Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Thu, 16 Oct 2008 20:25:28 -0600
Subject: cell: add new debug flag (cache) to report texture cache stats on
 exit

---
 src/gallium/drivers/cell/common.h           | 1 +
 src/gallium/drivers/cell/ppu/cell_context.c | 1 +
 src/gallium/drivers/cell/spu/spu_command.c  | 3 ++-
 src/gallium/drivers/cell/spu/spu_dcache.c   | 4 +++-
 4 files changed, 7 insertions(+), 2 deletions(-)

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 469d56cda8..9ca4e9d67e 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -117,6 +117,7 @@
 #define CELL_DEBUG_FRAGMENT_OPS         (1 << 3)
 #define CELL_DEBUG_FRAGMENT_OP_FALLBACK (1 << 4)
 #define CELL_DEBUG_CMD                  (1 << 5)
+#define CELL_DEBUG_CACHE                (1 << 6)
 
 /** Max instructions for doing per-fragment operations */
 #define SPU_MAX_FRAGMENT_OPS_INSTS 64
diff --git a/src/gallium/drivers/cell/ppu/cell_context.c b/src/gallium/drivers/cell/ppu/cell_context.c
index 4dad490ce1..7a2d93ecb4 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.c
+++ b/src/gallium/drivers/cell/ppu/cell_context.c
@@ -94,6 +94,7 @@ static const struct debug_named_value cell_debug_flags[] = {
    {"fragops", CELL_DEBUG_FRAGMENT_OPS}, /**< SPUs emit fragment ops debug messages*/
    {"fragopfallback", CELL_DEBUG_FRAGMENT_OP_FALLBACK}, /**< SPUs use reference implementation for fragment ops*/
    {"cmd", CELL_DEBUG_CMD},       /**< SPUs dump command buffer info */
+   {"cache", CELL_DEBUG_CACHE},   /**< report texture cache stats on exit */
    {NULL, 0}
 };
 
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index ff4a52d79a..9c853c0961 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -720,5 +720,6 @@ command_loop(void)
 
    D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n");
 
-   spu_dcache_report();
+   if (spu.init.debug_flags & CELL_DEBUG_CACHE)
+      spu_dcache_report();
 }
diff --git a/src/gallium/drivers/cell/spu/spu_dcache.c b/src/gallium/drivers/cell/spu/spu_dcache.c
index 167404cdc5..a6d67634fd 100644
--- a/src/gallium/drivers/cell/spu/spu_dcache.c
+++ b/src/gallium/drivers/cell/spu/spu_dcache.c
@@ -36,7 +36,9 @@
 #define CACHE_SET_TAGID(set)  (((set) & 0x03) + TAG_DCACHE0)
 #define CACHE_LOG2NNWAY       2
 #define CACHE_LOG2NSETS       6
-/*#define CACHE_STATS           1*/
+#ifdef DEBUG
+#define CACHE_STATS           1
+#endif
 #include <cache-api.h>
 
 /* Yes folks, this is ugly.
-- 
cgit v1.2.3


From 70dd4379d2cd54f229c3940312537912470218d3 Mon Sep 17 00:00:00 2001
From: Brian Paul <brian.paul@tungstengraphics.com>
Date: Wed, 22 Oct 2008 10:34:13 -0600
Subject: cell: implement fencing for texture buffers

If we delete a texture, we need to keep the underlying tiled data buffer
around until any rendering that references it has completed.
Keep a list of buffers referenced by a rendering batch.  Unref/free them when
the associated batch's fence is executed/signalled.
---
 src/gallium/drivers/cell/common.h              |  25 ++++
 src/gallium/drivers/cell/ppu/Makefile          |   1 +
 src/gallium/drivers/cell/ppu/cell_batch.c      |  32 +++++
 src/gallium/drivers/cell/ppu/cell_context.c    |   6 +
 src/gallium/drivers/cell/ppu/cell_context.h    |  21 ++++
 src/gallium/drivers/cell/ppu/cell_fence.c      | 158 +++++++++++++++++++++++++
 src/gallium/drivers/cell/ppu/cell_fence.h      |  57 +++++++++
 src/gallium/drivers/cell/ppu/cell_state_emit.c |   2 +-
 src/gallium/drivers/cell/ppu/cell_texture.c    |  33 ++++--
 src/gallium/drivers/cell/ppu/cell_texture.h    |   5 +-
 src/gallium/drivers/cell/ppu/cell_vbuf.c       |   6 +
 src/gallium/drivers/cell/spu/spu_command.c     |  38 +++++-
 src/gallium/drivers/cell/spu/spu_main.h        |   2 +-
 13 files changed, 367 insertions(+), 19 deletions(-)
 create mode 100644 src/gallium/drivers/cell/ppu/cell_fence.c
 create mode 100644 src/gallium/drivers/cell/ppu/cell_fence.h

(limited to 'src/gallium/drivers/cell/common.h')

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index 9ca4e9d67e..23fb0b0831 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -102,6 +102,8 @@
 #define CELL_CMD_STATE_RASTERIZER    22
 #define CELL_CMD_VS_EXECUTE          23
 #define CELL_CMD_FLUSH_BUFFER_RANGE  24
+#define CELL_CMD_FENCE               25
+
 
 /** Command/batch buffers */
 #define CELL_NUM_BUFFERS 4
@@ -123,6 +125,29 @@
 #define SPU_MAX_FRAGMENT_OPS_INSTS 64
 
 
+
+#define CELL_FENCE_IDLE      0
+#define CELL_FENCE_EMITTED   1
+#define CELL_FENCE_SIGNALLED 2
+
+struct cell_fence
+{
+   /** There's a 16-byte status qword per SPU */
+   volatile uint status[CELL_MAX_SPUS][4];
+};
+
+
+/**
+ * Fence command sent to SPUs.  In response, the SPUs will write
+ * CELL_FENCE_STATUS_SIGNALLED back to the fence status word in main memory.
+ */
+struct cell_command_fence
+{
+   uint64_t opcode;      /**< CELL_CMD_FENCE */
+   struct cell_fence *fence;
+};
+
+
 /**
  * Command to specify per-fragment operations state and generated code.
  * Note that the dsa, blend, blend_color fields are really only needed
diff --git a/src/gallium/drivers/cell/ppu/Makefile b/src/gallium/drivers/cell/ppu/Makefile
index b28f4c5c31..9358a47284 100644
--- a/src/gallium/drivers/cell/ppu/Makefile
+++ b/src/gallium/drivers/cell/ppu/Makefile
@@ -24,6 +24,7 @@ SOURCES = \
 	cell_clear.c \
 	cell_context.c \
 	cell_draw_arrays.c \
+	cell_fence.c \
 	cell_flush.c \
 	cell_gen_fragment.c \
 	cell_gen_fp.c \
diff --git a/src/gallium/drivers/cell/ppu/cell_batch.c b/src/gallium/drivers/cell/ppu/cell_batch.c
index 01254aed60..448b723d85 100644
--- a/src/gallium/drivers/cell/ppu/cell_batch.c
+++ b/src/gallium/drivers/cell/ppu/cell_batch.c
@@ -28,6 +28,7 @@
 
 #include "cell_context.h"
 #include "cell_batch.h"
+#include "cell_fence.h"
 #include "cell_spu.h"
 
 
@@ -63,6 +64,10 @@ cell_get_empty_buffer(struct cell_context *cell)
                printf("PPU: ALLOC BUFFER %u, %u tries\n", buf, tries);
                */
                prev_buffer = buf;
+
+               /* release tex buffer associated w/ prev use of this batch buf */
+               cell_free_fenced_buffers(cell, &cell->fenced_buffers[buf]);
+
                return buf;
             }
          }
@@ -84,6 +89,26 @@ cell_get_empty_buffer(struct cell_context *cell)
 }
 
 
+/**
+ * Append a fence command to the current batch buffer.
+ * Note that we're sure there's always room for this because of the
+ * adjusted size check in cell_batch_free_space().
+ */
+static void
+emit_fence(struct cell_context *cell)
+{
+   const uint batch = cell->cur_batch;
+   const uint size = cell->buffer_size[batch];
+   struct cell_command_fence *fence_cmd;
+
+   ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE);
+
+   fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size);
+   fence_cmd->opcode = CELL_CMD_FENCE;
+   fence_cmd->fence = &cell->fenced_buffers[batch].fence;
+}
+
+
 /**
  * Flush the current batch buffer to the SPUs.
  * An empty buffer will be found and set as the new current batch buffer
@@ -102,6 +127,12 @@ cell_batch_flush(struct cell_context *cell)
    if (size == 0)
       return;
 
+   /* Before we use this batch buffer, make sure any fenced texture buffers
+    * are released.
+    */
+   if (cell->fenced_buffers[batch].head)
+      emit_fence(cell);
+
    flushing = TRUE;
 
    assert(batch < CELL_NUM_BUFFERS);
@@ -142,6 +173,7 @@ uint
 cell_batch_free_space(const struct cell_context *cell)
 {
    uint free = CELL_BUFFER_SIZE - cell->buffer_size[cell->cur_batch];
+   free -= sizeof(struct cell_command_fence);
    return free;
 }
 
diff --git a/src/gallium/drivers/cell/ppu/cell_context.c b/src/gallium/drivers/cell/ppu/cell_context.c
index 7a2d93ecb4..22d552d8e3 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.c
+++ b/src/gallium/drivers/cell/ppu/cell_context.c
@@ -47,6 +47,7 @@
 #include "cell_clear.h"
 #include "cell_context.h"
 #include "cell_draw_arrays.h"
+#include "cell_fence.h"
 #include "cell_flush.h"
 #include "cell_state.h"
 #include "cell_surface.h"
@@ -104,6 +105,7 @@ cell_create_context(struct pipe_screen *screen,
                     struct cell_winsys *cws)
 {
    struct cell_context *cell;
+   uint i;
 
    /* some fields need to be 16-byte aligned, so align the whole object */
    cell = (struct cell_context*) align_malloc(sizeof(struct cell_context), 16);
@@ -151,6 +153,10 @@ cell_create_context(struct pipe_screen *screen,
                                               cell_debug_flags, 
                                               0 );
 
+   for (i = 0; i < CELL_NUM_BUFFERS; i++)
+      cell_fence_init(&cell->fenced_buffers[i].fence);
+
+
    /*
     * SPU stuff
     */
diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h
index ad1f4829a4..4491ae8cdf 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.h
+++ b/src/gallium/drivers/cell/ppu/cell_context.h
@@ -81,6 +81,19 @@ struct cell_fragment_ops_key
 };
 
 
+struct cell_buffer_node;
+
+/**
+ * Fenced buffer list.  List of buffers which can be unreferenced after
+ * the fence has been executed/signalled.
+ */
+struct cell_buffer_list
+{
+   struct cell_fence fence;
+   struct cell_buffer_node *head;
+};
+
+
 /**
  * Per-context state, subclass of pipe_context.
  */
@@ -154,6 +167,14 @@ struct cell_context
    uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BUFFERS][4] ALIGN16_ATTRIB;
 
 
+   /** Associated with each command/batch buffer is a list of pipe_buffers
+    * that are fenced.  When the last command in a buffer is executed, the
+    * fence will be signalled, indicating that any pipe_buffers preceeding
+    * that fence can be unreferenced (and probably freed).
+    */
+   struct cell_buffer_list fenced_buffers[CELL_NUM_BUFFERS];
+
+
    struct spe_function attrib_fetch;
    unsigned attrib_fetch_offsets[PIPE_MAX_ATTRIBS];
 
diff --git a/src/gallium/drivers/cell/ppu/cell_fence.c b/src/gallium/drivers/cell/ppu/cell_fence.c
new file mode 100644
index 0000000000..ffb3bea12b
--- /dev/null
+++ b/src/gallium/drivers/cell/ppu/cell_fence.c
@@ -0,0 +1,158 @@
+/**************************************************************************
+ * 
+ * Copyright 2008 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 <unistd.h>
+#include "util/u_memory.h"
+#include "pipe/p_inlines.h"
+#include "cell_context.h"
+#include "cell_batch.h"
+#include "cell_fence.h"
+#include "cell_texture.h"
+
+
+void
+cell_fence_init(struct cell_fence *fence)
+{
+   uint i;
+   for (i = 0; i < CELL_MAX_SPUS; i++) {
+      fence->status[i][0] = CELL_FENCE_IDLE;
+   }
+}
+
+
+boolean
+cell_fence_signalled(const struct cell_context *cell,
+                     const struct cell_fence *fence)
+{
+   uint i;
+   for (i = 0; i < cell->num_spus; i++) {
+      //ASSERT(fence->status[i][0] != CELL_FENCE_IDLE);
+      if (fence->status[i][0] == CELL_FENCE_EMITTED)
+         return FALSE;
+   }
+   return TRUE;
+}
+
+
+void
+cell_fence_finish(const struct cell_context *cell,
+                  const struct cell_fence *fence)
+{
+   while (!cell_fence_signalled(cell, fence)) {
+      usleep(10);
+   }
+}
+
+
+
+
+struct cell_buffer_node
+{
+   struct pipe_buffer *buffer;
+   struct cell_buffer_node *next;
+};
+
+
+static void
+cell_add_buffer_to_list(struct cell_context *cell,
+                        struct cell_buffer_list *list,
+                        struct pipe_buffer *buffer)
+{
+   struct pipe_screen *ps = cell->pipe.screen;
+   struct cell_buffer_node *node = CALLOC_STRUCT(cell_buffer_node);
+   /* create new list node which references the buffer, insert at head */
+   if (node) {
+      pipe_buffer_reference(ps, &node->buffer, buffer);
+      node->next = list->head;
+      list->head = node;
+   }
+}
+
+
+/**
+ * Wait for completion of the given fence, then unreference any buffers
+ * on the list.
+ * This typically unrefs/frees texture buffers after any rendering which uses
+ * them has completed.
+ */
+void
+cell_free_fenced_buffers(struct cell_context *cell,
+                         struct cell_buffer_list *list)
+{
+   if (list->head) {
+      struct pipe_screen *ps = cell->pipe.screen;
+      struct cell_buffer_node *node;
+
+      cell_fence_finish(cell, &list->fence);
+
+      /* traverse the list, unreferencing buffers, freeing nodes */
+      node = list->head;
+      while (node) {
+         struct cell_buffer_node *next = node->next;
+         assert(node->buffer);
+         pipe_buffer_unmap(ps, node->buffer);
+#if 0
+         printf("Unref buffer %p\n", node->buffer);
+         if (node->buffer->refcount == 1)
+            printf("   Delete!\n");
+#endif
+         pipe_buffer_reference(ps, &node->buffer, NULL);
+         FREE(node);
+         node = next;
+      }
+      list->head = NULL;
+   }
+}
+
+
+/**
+ * This should be called for each render command.
+ * Any texture buffers that are current bound will be added to a fenced
+ * list to be freed later when the fence is executed/signalled.
+ */
+void
+cell_add_fenced_textures(struct cell_context *cell)
+{
+   struct cell_buffer_list *list = &cell->fenced_buffers[cell->cur_batch];
+   uint i;
+
+   for (i = 0; i < cell->num_textures; i++) {
+      struct cell_texture *ct = cell->texture[i];
+      if (ct) {
+         uint level;
+         for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
+            if (ct->tiled_buffer[level]) {
+#if 0
+               printf("Adding texture %p buffer %p to list\n",
+                      ct, ct->tiled_buffer[level]);
+#endif
+               cell_add_buffer_to_list(cell, list, ct->tiled_buffer[level]);
+            }
+         }
+      }
+   }
+}
diff --git a/src/gallium/drivers/cell/ppu/cell_fence.h b/src/gallium/drivers/cell/ppu/cell_fence.h
new file mode 100644
index 0000000000..536b4ba411
--- /dev/null
+++ b/src/gallium/drivers/cell/ppu/cell_fence.h
@@ -0,0 +1,57 @@
+/**************************************************************************
+ * 
+ * Copyright 2008 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 CELL_FENCE_H
+#define CELL_FENCE_H
+
+
+extern void
+cell_fence_init(struct cell_fence *fence);
+
+
+extern boolean
+cell_fence_signalled(const struct cell_context *cell,
+                     const struct cell_fence *fence);
+
+
+extern void
+cell_fence_finish(const struct cell_context *cell,
+                  const struct cell_fence *fence);
+
+
+
+extern void
+cell_free_fenced_buffers(struct cell_context *cell,
+                         struct cell_buffer_list *list);
+
+
+extern void
+cell_add_fenced_textures(struct cell_context *cell);
+
+
+#endif /* CELL_FENCE_H */
diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c
index effcd2a1e1..dd2d7f7d1e 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_emit.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c
@@ -225,7 +225,7 @@ cell_emit_state(struct cell_context *cell)
             if (cell->texture[i]) {
                uint level;
                for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
-                  texture->start[level] = cell->texture[i]->tiled_data[level];
+                  texture->start[level] = cell->texture[i]->tiled_mapped[level];
                   texture->width[level] = cell->texture[i]->base.width[level];
                   texture->height[level] = cell->texture[i]->base.height[level];
                   texture->depth[level] = cell->texture[i]->base.depth[level];
diff --git a/src/gallium/drivers/cell/ppu/cell_texture.c b/src/gallium/drivers/cell/ppu/cell_texture.c
index 9c6741f1bc..9ac2f3bbb9 100644
--- a/src/gallium/drivers/cell/ppu/cell_texture.c
+++ b/src/gallium/drivers/cell/ppu/cell_texture.c
@@ -136,6 +136,9 @@ cell_texture_release(struct pipe_screen *screen,
        __FUNCTION__, (void *) *pt, (*pt)->refcount - 1);
    */
    if (--(*pt)->refcount <= 0) {
+      /* Delete this texture now.
+       * But note that the underlying pipe_buffer may linger...
+       */
       struct cell_texture *ct = cell_texture(*pt);
       uint i;
 
@@ -146,14 +149,12 @@ cell_texture_release(struct pipe_screen *screen,
       pipe_buffer_reference(screen, &ct->buffer, NULL);
 
       for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
-         if (ct->tiled_data[i]) {
-            /* XXX need to use a fenced buffer for tiled data so that
-             * it's properly freed after rendering has completed.
-             * Disabling this free() allows glDrawPixels to work for now.
-             */
-#if 0
-            align_free(ct->tiled_data[i]);
-#endif
+         /* Unreference the tiled image buffer.
+          * It may not actually be deleted until a fence is hit.
+          */
+         if (ct->tiled_buffer[i]) {
+            ct->tiled_mapped[i] = NULL;
+            winsys_buffer_reference(screen->winsys, &ct->tiled_buffer[i], NULL);
          }
       }
 
@@ -234,12 +235,18 @@ cell_twiddle_texture(struct pipe_screen *screen,
          int offset = bufWidth * bufHeight * 4 * surface->face;
          uint *dst;
 
-         if (!ct->tiled_data[level]) {
-            ct->tiled_data[level] =
-               align_malloc(bufWidth * bufHeight * 4 * numFaces, 16);
+         if (!ct->tiled_buffer[level]) {
+            /* allocate buffer for tiled data now */
+            struct pipe_winsys *ws = screen->winsys;
+            uint bytes = bufWidth * bufHeight * 4 * numFaces;
+            ct->tiled_buffer[level] = ws->buffer_create(ws, 16,
+                                                        PIPE_BUFFER_USAGE_PIXEL,
+                                                        bytes);
+            /* and map it */
+            ct->tiled_mapped[level] = ws->buffer_map(ws, ct->tiled_buffer[level],
+                                                     PIPE_BUFFER_USAGE_GPU_READ);
          }
-
-         dst = (uint *) ((ubyte *) ct->tiled_data[level] + offset);
+         dst = (uint *) ((ubyte *) ct->tiled_mapped[level] + offset);
 
          twiddle_image_uint(texWidth, texHeight, TILE_SIZE, dst,
                             surface->stride, src);
diff --git a/src/gallium/drivers/cell/ppu/cell_texture.h b/src/gallium/drivers/cell/ppu/cell_texture.h
index a0757091b0..2f5fe0dd1b 100644
--- a/src/gallium/drivers/cell/ppu/cell_texture.h
+++ b/src/gallium/drivers/cell/ppu/cell_texture.h
@@ -48,7 +48,10 @@ struct cell_texture
    struct pipe_buffer *buffer;
    unsigned long buffer_size;
 
-   void *tiled_data[CELL_MAX_TEXTURE_LEVELS];  /* XXX this may be temporary */ /*ALIGN16*/
+   /** Texture data in tiled layout is held here */
+   struct pipe_buffer *tiled_buffer[CELL_MAX_TEXTURE_LEVELS];
+   /** Mapped, tiled texture data */
+   void *tiled_mapped[CELL_MAX_TEXTURE_LEVELS];
 };
 
 
diff --git a/src/gallium/drivers/cell/ppu/cell_vbuf.c b/src/gallium/drivers/cell/ppu/cell_vbuf.c
index aa63435b93..65ba51b6bb 100644
--- a/src/gallium/drivers/cell/ppu/cell_vbuf.c
+++ b/src/gallium/drivers/cell/ppu/cell_vbuf.c
@@ -38,6 +38,7 @@
 
 #include "cell_batch.h"
 #include "cell_context.h"
+#include "cell_fence.h"
 #include "cell_flush.h"
 #include "cell_spu.h"
 #include "cell_vbuf.h"
@@ -108,6 +109,11 @@ cell_vbuf_release_vertices(struct vbuf_render *vbr, void *vertices,
           __FUNCTION__, cvbr->vertex_buf, vertices_used);
    */
 
+   /* Make sure texture buffers aren't released until we're done rendering
+    * with them.
+    */
+   cell_add_fenced_textures(cell);
+
    /* Tell SPUs they can release the vert buf */
    if (cvbr->vertex_buf != ~0U) {
       struct cell_command_release_verts *release
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index 9c853c0961..a6ed29ea63 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -76,9 +76,10 @@ static void
 release_buffer(uint buffer)
 {
    /* Evidently, using less than a 16-byte status doesn't work reliably */
-   static const uint status[4] ALIGN16_ATTRIB
-      = {CELL_BUFFER_STATUS_FREE, 0, 0, 0};
-
+   static const vector unsigned int status = {CELL_BUFFER_STATUS_FREE,
+                                              CELL_BUFFER_STATUS_FREE,
+                                              CELL_BUFFER_STATUS_FREE,
+                                              CELL_BUFFER_STATUS_FREE};
    const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer);
    uint *dst = spu.init.buffer_status + index;
 
@@ -93,6 +94,29 @@ release_buffer(uint buffer)
 }
 
 
+/**
+ * Write CELL_FENCE_SIGNALLED back to the fence status qword in main memory.
+ * There's a qword of status per SPU.
+ */
+static void
+cmd_fence(struct cell_command_fence *fence_cmd)
+{
+   static const vector unsigned int status = {CELL_FENCE_SIGNALLED,
+                                              CELL_FENCE_SIGNALLED,
+                                              CELL_FENCE_SIGNALLED,
+                                              CELL_FENCE_SIGNALLED};
+   uint *dst = (uint *) fence_cmd->fence;
+   dst += 4 * spu.init.id;  /* main store/memory address, not local store */
+
+   mfc_put((void *) &status,    /* src in local memory */
+           (unsigned int) dst,  /* dst in main memory */
+           sizeof(status),      /* size */
+           TAG_FENCE,           /* tag */
+           0, /* tid */
+           0  /* rid */);
+}
+
+
 static void
 cmd_clear_surface(const struct cell_command_clear_surface *clear)
 {
@@ -637,6 +661,14 @@ cmd_batch(uint opcode)
          cmd_finish();
          pos += 1;
          break;
+      case CELL_CMD_FENCE:
+         {
+            struct cell_command_fence *fence_cmd =
+               (struct cell_command_fence *) &buffer[pos];
+            cmd_fence(fence_cmd);
+            pos += sizeof(*fence_cmd) / 8;
+         }
+         break;
       case CELL_CMD_RELEASE_VERTS:
          {
             struct cell_command_release_verts *release
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index 95ef4c9244..668af10be2 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -210,7 +210,7 @@ extern struct spu_global spu;
 #define TAG_DCACHE1           21
 #define TAG_DCACHE2           22
 #define TAG_DCACHE3           23
-
+#define TAG_FENCE             24
 
 
 static INLINE void
-- 
cgit v1.2.3