Cell: generalize the batch buffer code for vertex buffers...

7 files changed, 79 insertions, 57 deletions

diff --git a/src/mesa/pipe/cell/common.h b/src/mesa/pipe/cell/common.h
index 0b63ed39be..ce9c381907 100644
--- a/src/mesa/pipe/cell/common.h
+++ b/src/mesa/pipe/cell/common.h
@@ -81,8 +81,8 @@
 #define CELL_CMD_STATE_VERTEX_INFO   13
 
 
-#define CELL_NUM_BATCH_BUFFERS 3
-#define CELL_BATCH_BUFFER_SIZE 1024  /**< 16KB would be the max */
+#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
@@ -147,7 +147,9 @@ struct cell_init_info
    unsigned id;
    unsigned num_spus;
    struct cell_command *cmd;
-   ubyte *batch_buffers[CELL_NUM_BATCH_BUFFERS];
+
+   /** Buffers for command batches, vertex/index data */
+   ubyte *buffers[CELL_NUM_BUFFERS];
    uint *buffer_status;  /**< points at cell_context->buffer_status */
 } ALIGN16_ATTRIB;
 
diff --git a/src/mesa/pipe/cell/ppu/cell_batch.c b/src/mesa/pipe/cell/ppu/cell_batch.c
index c894ef8608..178caa74e1 100644
--- a/src/mesa/pipe/cell/ppu/cell_batch.c
+++ b/src/mesa/pipe/cell/ppu/cell_batch.c
@@ -31,12 +31,46 @@
 #include "cell_spu.h"
 
 
+
+uint
+cell_get_empty_buffer(struct cell_context *cell)
+{
+   uint buf = 0;
+
+   /* Find a buffer that's marked as free by all SPUs */
+   while (1) {
+      uint spu, num_free = 0;
+
+      for (spu = 0; spu < cell->num_spus; spu++) {
+         if (cell->buffer_status[spu][buf][0] == CELL_BUFFER_STATUS_FREE) {
+            num_free++;
+
+            if (num_free == cell->num_spus) {
+               /* found a free buffer, now mark status as used */
+               for (spu = 0; spu < cell->num_spus; spu++) {
+                  cell->buffer_status[spu][buf][0] = CELL_BUFFER_STATUS_USED;
+               }
+               return buf;
+            }
+         }
+         else {
+            break;
+         }
+      }
+
+      /* try next buf */
+      buf = (buf + 1) % CELL_NUM_BUFFERS;
+   }
+}
+
+
+
 void
 cell_batch_flush(struct cell_context *cell)
 {
    static boolean flushing = FALSE;
    uint batch = cell->cur_batch;
-   const uint size = cell->batch_buffer_size[batch];
+   const uint size = cell->buffer_size[batch];
    uint spu, cmd_word;
 
    assert(!flushing);
@@ -46,7 +80,7 @@ cell_batch_flush(struct cell_context *cell)
 
    flushing = TRUE;
 
-   assert(batch < CELL_NUM_BATCH_BUFFERS);
+   assert(batch < CELL_NUM_BUFFERS);
 
    /*
    printf("cell_batch_dispatch: buf %u at %p, size %u\n",
@@ -68,28 +102,9 @@ cell_batch_flush(struct cell_context *cell)
     * array indicating that the PPU can re-use the buffer.
     */
 
+   batch = cell_get_empty_buffer(cell);
 
-   /* Find a buffer that's marked as free by all SPUs */
-   while (1) {
-      uint num_free = 0;
-
-      batch = (batch + 1) % CELL_NUM_BATCH_BUFFERS;
-
-      for (spu = 0; spu < cell->num_spus; spu++) {
-         if (cell->buffer_status[spu][batch][0] == CELL_BUFFER_STATUS_FREE)
-            num_free++;
-      }
-
-      if (num_free == cell->num_spus) {
-         /* found a free buffer, now mark status as used */
-         for (spu = 0; spu < cell->num_spus; spu++) {
-            cell->buffer_status[spu][batch][0] = CELL_BUFFER_STATUS_USED;
-         }
-         break;
-      }
-   }
-
-   cell->batch_buffer_size[batch] = 0;  /* empty */
+   cell->buffer_size[batch] = 0;  /* empty */
    cell->cur_batch = batch;
 
    flushing = FALSE;
@@ -99,8 +114,7 @@ cell_batch_flush(struct cell_context *cell)
 uint
 cell_batch_free_space(const struct cell_context *cell)
 {
-   uint free = CELL_BATCH_BUFFER_SIZE
-      - cell->batch_buffer_size[cell->cur_batch];
+   uint free = CELL_BUFFER_SIZE - cell->buffer_size[cell->cur_batch];
    return free;
 }
 
@@ -117,18 +131,18 @@ cell_batch_append(struct cell_context *cell, const void *cmd, uint length)
    assert(length % 4 == 0);
    assert(cell->cur_batch >= 0);
 
-   size = cell->batch_buffer_size[cell->cur_batch];
+   size = cell->buffer_size[cell->cur_batch];
 
-   if (size + length > CELL_BATCH_BUFFER_SIZE) {
+   if (size + length > CELL_BUFFER_SIZE) {
       cell_batch_flush(cell);
       size = 0;
    }
 
-   assert(size + length <= CELL_BATCH_BUFFER_SIZE);
+   assert(size + length <= CELL_BUFFER_SIZE);
 
-   memcpy(cell->batch_buffer[cell->cur_batch] + size, cmd, length);
+   memcpy(cell->buffer[cell->cur_batch] + size, cmd, length);
 
-   cell->batch_buffer_size[cell->cur_batch] = size + length;
+   cell->buffer_size[cell->cur_batch] = size + length;
 }
 
 
@@ -142,18 +156,18 @@ cell_batch_alloc(struct cell_context *cell, uint bytes)
 
    assert(cell->cur_batch >= 0);
 
-   size = cell->batch_buffer_size[cell->cur_batch];
+   size = cell->buffer_size[cell->cur_batch];
 
-   if (size + bytes > CELL_BATCH_BUFFER_SIZE) {
+   if (size + bytes > CELL_BUFFER_SIZE) {
       cell_batch_flush(cell);
       size = 0;
    }
 
-   assert(size + bytes <= CELL_BATCH_BUFFER_SIZE);
+   assert(size + bytes <= CELL_BUFFER_SIZE);
 
-   pos = (void *) (cell->batch_buffer[cell->cur_batch] + size);
+   pos = (void *) (cell->buffer[cell->cur_batch] + size);
 
-   cell->batch_buffer_size[cell->cur_batch] = size + bytes;
+   cell->buffer_size[cell->cur_batch] = size + bytes;
 
    return pos;
 }
diff --git a/src/mesa/pipe/cell/ppu/cell_batch.h b/src/mesa/pipe/cell/ppu/cell_batch.h
index c4ba7feb3d..b4c96f465a 100644
--- a/src/mesa/pipe/cell/ppu/cell_batch.h
+++ b/src/mesa/pipe/cell/ppu/cell_batch.h
@@ -35,6 +35,9 @@
 struct cell_context;
 
 
+extern uint
+cell_get_empty_buffer(struct cell_context *cell);
+
 extern void
 cell_batch_flush(struct cell_context *cell);
 
diff --git a/src/mesa/pipe/cell/ppu/cell_context.c b/src/mesa/pipe/cell/ppu/cell_context.c
index 8cb0c48f40..e8020a49bc 100644
--- a/src/mesa/pipe/cell/ppu/cell_context.c
+++ b/src/mesa/pipe/cell/ppu/cell_context.c
@@ -254,8 +254,9 @@ cell_create_context(struct pipe_winsys *winsys, struct cell_winsys *cws)
 
    cell_start_spus(cell);
 
-   for (buf = 0; buf < CELL_NUM_BATCH_BUFFERS; buf++) {
-      cell->batch_buffer_size[buf] = 0;
+   /* init command, vertex/index buffer info */
+   for (buf = 0; buf < CELL_NUM_BUFFERS; buf++) {
+      cell->buffer_size[buf] = 0;
 
       /* init batch buffer status values,
        * mark 0th buffer as used, rest as free.
diff --git a/src/mesa/pipe/cell/ppu/cell_context.h b/src/mesa/pipe/cell/ppu/cell_context.h
index 3bd88bfd5b..de65fb5e9a 100644
--- a/src/mesa/pipe/cell/ppu/cell_context.h
+++ b/src/mesa/pipe/cell/ppu/cell_context.h
@@ -102,12 +102,14 @@ struct cell_context
 
    uint num_spus;
 
-   uint batch_buffer_size[CELL_NUM_BATCH_BUFFERS];
-   ubyte batch_buffer[CELL_NUM_BATCH_BUFFERS][CELL_BATCH_BUFFER_SIZE] ALIGN16_ATTRIB;
-   int cur_batch;  /**< which batch buffer is being filled */
+   /** Buffers for command batches, vertex/index data */
+   uint buffer_size[CELL_NUM_BUFFERS];
+   ubyte buffer[CELL_NUM_BUFFERS][CELL_BUFFER_SIZE] ALIGN16_ATTRIB;
+
+   int cur_batch;  /**< which buffer is being filled w/ commands */
 
    /** [4] to ensure 16-byte alignment for each status word */
-   uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BATCH_BUFFERS][4] ALIGN16_ATTRIB;
+   uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BUFFERS][4] ALIGN16_ATTRIB;
 
 };
 
diff --git a/src/mesa/pipe/cell/ppu/cell_spu.c b/src/mesa/pipe/cell/ppu/cell_spu.c
index 4627bc8d1f..7c83a47e57 100644
--- a/src/mesa/pipe/cell/ppu/cell_spu.c
+++ b/src/mesa/pipe/cell/ppu/cell_spu.c
@@ -111,8 +111,8 @@ 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].cmd = &cell_global.command[i];
-      for (j = 0; j < CELL_NUM_BATCH_BUFFERS; j++) {
-         cell_global.inits[i].batch_buffers[j] = cell->batch_buffer[j];
+      for (j = 0; j < CELL_NUM_BUFFERS; j++) {
+         cell_global.inits[i].buffers[j] = cell->buffer[j];
       }
       cell_global.inits[i].buffer_status = &cell->buffer_status[0][0][0];
 
diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
index 0c83900a18..2097683b82 100644
--- a/src/mesa/pipe/cell/spu/spu_main.c
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -473,22 +473,22 @@ cmd_finish(void)
 
 
 /**
- * Tell the PPU that this SPU has finished copying a batch buffer to
+ * Tell the PPU that this SPU has finished copying a buffer to
  * local store and that it may be reused by the PPU.
  * This is done by writting a 16-byte batch-buffer-status block back into
- * main memory (in cell_contex->buffer_status[]).
+ * main memory (in cell_context->buffer_status[]).
  */
 static void
-release_batch_buffer(uint buffer)
+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};
 
-   const uint index = 4 * (spu.init.id * CELL_NUM_BATCH_BUFFERS + buffer);
+   const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer);
    uint *dst = spu.init.buffer_status + index;
 
-   ASSERT(buffer < CELL_NUM_BATCH_BUFFERS);
+   ASSERT(buffer < CELL_NUM_BUFFERS);
 
    /*
    printf("SPU %u: Set batch status buf=%u, index %u, at %p to FREE\n",
@@ -513,24 +513,24 @@ cmd_batch(uint opcode)
 {
    const uint buf = (opcode >> 8) & 0xff;
    uint size = (opcode >> 16);
-   uint buffer[CELL_BATCH_BUFFER_SIZE / 4] ALIGN16_ATTRIB;
+   uint buffer[CELL_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",
-             spu.init.id, buf, size, spu.init.batch_buffers[buf]);
+             spu.init.id, buf, size, spu.init.buffers[buf]);
 
    ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH);
 
-   ASSERT_ALIGN16(spu.init.batch_buffers[buf]);
+   ASSERT_ALIGN16(spu.init.buffers[buf]);
 
    size = ROUNDUP16(size);
 
-   ASSERT_ALIGN16(spu.init.batch_buffers[buf]);
+   ASSERT_ALIGN16(spu.init.buffers[buf]);
 
    mfc_get(buffer,  /* dest */
-           (unsigned int) spu.init.batch_buffers[buf],  /* src */
+           (unsigned int) spu.init.buffers[buf],  /* src */
            size,
            TAG_BATCH_BUFFER,
            0, /* tid */
@@ -538,7 +538,7 @@ cmd_batch(uint opcode)
    wait_on_mask(1 << TAG_BATCH_BUFFER);
 
    /* Tell PPU we're done copying the buffer to local store */
-   release_batch_buffer(buf);
+   release_buffer(buf);
 
    for (pos = 0; pos < usize; /* no incr */) {
       switch (buffer[pos]) {