1 files changed, 58 insertions, 65 deletions

diff --git a/src/gallium/drivers/cell/ppu/cell_batch.c b/src/gallium/drivers/cell/ppu/cell_batch.c
index 16882c0129..fe144f8b84 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"
 
 
@@ -42,7 +43,9 @@
 uint
 cell_get_empty_buffer(struct cell_context *cell)
 {
-   uint buf = 0, tries = 0;
+   static uint prev_buffer = 0;
+   uint buf = (prev_buffer + 1) % CELL_NUM_BUFFERS;
+   uint tries = 0;
 
    /* Find a buffer that's marked as free by all SPUs */
    while (1) {
@@ -58,8 +61,13 @@ cell_get_empty_buffer(struct cell_context *cell)
                   cell->buffer_status[spu][buf][0] = CELL_BUFFER_STATUS_USED;
                }
                /*
-               printf("PPU: ALLOC BUFFER %u\n", buf);
+               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;
             }
          }
@@ -82,6 +90,38 @@ 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;
+   struct cell_fence *fence = &cell->fenced_buffers[batch].fence;
+   uint i;
+
+   /* set fence status to emitted, not yet signalled */
+   for (i = 0; i < cell->num_spus; i++) {
+      fence->status[i][0] = CELL_FENCE_EMITTED;
+   }
+
+   STATIC_ASSERT(sizeof(struct cell_command_fence) % 16 == 0);
+   ASSERT(size % 16 == 0);
+   ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE);
+
+   fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size);
+   fence_cmd->opcode[0] = CELL_CMD_FENCE;
+   fence_cmd->fence = fence;
+
+   /* update batch buffer size */
+   cell->buffer_size[batch] = size + sizeof(struct cell_command_fence);
+}
+
+
+/**
  * Flush the current batch buffer to the SPUs.
  * An empty buffer will be found and set as the new current batch buffer
  * for subsequent commands/data.
@@ -91,7 +131,7 @@ cell_batch_flush(struct cell_context *cell)
 {
    static boolean flushing = FALSE;
    uint batch = cell->cur_batch;
-   const uint size = cell->buffer_size[batch];
+   uint size = cell->buffer_size[batch];
    uint spu, cmd_word;
 
    assert(!flushing);
@@ -99,6 +139,14 @@ 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);
+      size = cell->buffer_size[batch];
+   }
+
    flushing = TRUE;
 
    assert(batch < CELL_NUM_BUFFERS);
@@ -139,74 +187,24 @@ 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;
 }
 
 
 /**
- * Append data to the current batch buffer.
- * \param data  address of block of bytes to append
- * \param bytes  size of block of bytes
- */
-void
-cell_batch_append(struct cell_context *cell, const void *data, uint bytes)
-{
-   uint size;
-
-   ASSERT(bytes % 8 == 0);
-   ASSERT(bytes <= CELL_BUFFER_SIZE);
-   ASSERT(cell->cur_batch >= 0);
-
-#ifdef ASSERT
-   {
-      uint spu;
-      for (spu = 0; spu < cell->num_spus; spu++) {
-         ASSERT(cell->buffer_status[spu][cell->cur_batch][0]
-                 == CELL_BUFFER_STATUS_USED);
-      }
-   }
-#endif
-
-   size = cell->buffer_size[cell->cur_batch];
-
-   if (size + bytes > CELL_BUFFER_SIZE) {
-      cell_batch_flush(cell);
-      size = 0;
-   }
-
-   ASSERT(size + bytes <= CELL_BUFFER_SIZE);
-
-   memcpy(cell->buffer[cell->cur_batch] + size, data, bytes);
-
-   cell->buffer_size[cell->cur_batch] = size + bytes;
-}
-
-
-/**
  * Allocate space in the current batch buffer for 'bytes' space.
+ * Bytes must be a multiple of 16 bytes.  Allocation will be 16 byte aligned.
  * \return address in batch buffer to put data
  */
 void *
-cell_batch_alloc(struct cell_context *cell, uint bytes)
-{
-   return cell_batch_alloc_aligned(cell, bytes, 1);
-}
-
-
-/**
- * Same as \sa cell_batch_alloc, but return an address at a particular
- * alignment.
- */
-void *
-cell_batch_alloc_aligned(struct cell_context *cell, uint bytes,
-                         uint alignment)
+cell_batch_alloc16(struct cell_context *cell, uint bytes)
 {
    void *pos;
-   uint size, padbytes;
+   uint size;
 
-   ASSERT(bytes % 8 == 0);
+   ASSERT(bytes % 16 == 0);
    ASSERT(bytes <= CELL_BUFFER_SIZE);
-   ASSERT(alignment > 0);
    ASSERT(cell->cur_batch >= 0);
 
 #ifdef ASSERT
@@ -221,17 +219,12 @@ cell_batch_alloc_aligned(struct cell_context *cell, uint bytes,
 
    size = cell->buffer_size[cell->cur_batch];
 
-   padbytes = (alignment - (size % alignment)) % alignment;
-
-   if (padbytes + size + bytes > CELL_BUFFER_SIZE) {
+   if (bytes > cell_batch_free_space(cell)) {
       cell_batch_flush(cell);
       size = 0;
    }
-   else {
-      size += padbytes;
-   }
 
-   ASSERT(size % alignment == 0);
+   ASSERT(size % 16 == 0);
    ASSERT(size + bytes <= CELL_BUFFER_SIZE);
 
    pos = (void *) (cell->buffer[cell->cur_batch] + size);