Cell: improve "finished copying batch buffer" signalling.

When the SPU is done copying a batch buffer to local store, use an mfc_put()
to write a "done" message back to the buffer status array in main memory.
We were previously using a mailbox message for synchronization.

src/mesa/pipe/cell/common.h

@@ -62,10 +62,12 @@
 #define CELL_CMD_STATE_DEPTH_STENCIL 7
 
 
-#define CELL_NUM_BATCH_BUFFERS 2
+#define CELL_NUM_BATCH_BUFFERS 3
 #define CELL_BATCH_BUFFER_SIZE 1024  /**< 16KB would be the max */
 
-#define CELL_BATCH_FINISHED 0x1234   /**< mbox message */
+#define CELL_BUFFER_STATUS_FREE 10
+#define CELL_BUFFER_STATUS_USED 20
+
 
 
 /**
@@ -122,6 +124,7 @@ struct cell_init_info
    unsigned num_spus;
    struct cell_command *cmd;
    ubyte *batch_buffers[CELL_NUM_BATCH_BUFFERS];
+   uint *buffer_status;  /**< points at cell_context->buffer_status */
 } ALIGN16_ATTRIB;
 
 

src/mesa/pipe/cell/ppu/cell_batch.c

@@ -34,9 +34,9 @@
 void
 cell_batch_flush(struct cell_context *cell)
 {
-   const uint batch = cell->cur_batch;
+   uint batch = cell->cur_batch;
    const uint size = cell->batch_buffer_size[batch];
-   uint i, cmd_word;
+   uint spu, cmd_word;
 
    if (size == 0)
       return;
@@ -48,25 +48,44 @@ cell_batch_flush(struct cell_context *cell)
           batch, &cell->batch_buffer[batch][0], size);
    */
      
+   /*
+    * Build "BATCH" command and sent to all SPUs.
+    */
    cmd_word = CELL_CMD_BATCH | (batch << 8) | (size << 16);
 
-   for (i = 0; i < cell->num_spus; i++) {
-      send_mbox_message(cell_global.spe_contexts[i], cmd_word);
+   for (spu = 0; spu < cell->num_spus; spu++) {
+      assert(cell->buffer_status[spu][batch][0] == CELL_BUFFER_STATUS_USED);
+      send_mbox_message(cell_global.spe_contexts[spu], cmd_word);
    }
 
-   /* XXX wait for the DMX xfer to finish.
-    * Using mailboxes here is temporary.
-    * Ideally, we want to use a PPE-side DMA status check function...
+   /* When the SPUs are done copying the buffer into their locals stores
+    * they'll write a BUFFER_STATUS_FREE message into the buffer_status[]
+    * array indicating that the PPU can re-use the buffer.
     */
-   for (i = 0; i < cell->num_spus; i++) {
-      uint k = wait_mbox_message(cell_global.spe_contexts[i]);
-      assert(k == CELL_BATCH_FINISHED);
-   }
 
-   /* next buffer */
-   cell->cur_batch = (batch + 1) % CELL_NUM_BATCH_BUFFERS;
 
-   cell->batch_buffer_size[cell->cur_batch] = 0;  /* empty */
+   /* 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->cur_batch = batch;
 }
 
 

src/mesa/pipe/cell/ppu/cell_context.c

@@ -160,7 +160,7 @@ struct pipe_context *
 cell_create_context(struct pipe_winsys *winsys, struct cell_winsys *cws)
 {
    struct cell_context *cell;
-   uint i;
+   uint spu, buf;
 
    /* some fields need to be 16-byte aligned, so align the whole object */
    cell = (struct cell_context*) align_malloc(sizeof(struct cell_context), 16);
@@ -248,13 +248,30 @@ cell_create_context(struct pipe_winsys *winsys, struct cell_winsys *cws)
 
    cell_start_spus(cell);
 
-   for (i = 0; i < CELL_NUM_BATCH_BUFFERS; i++) {
-      cell->batch_buffer_size[i] = 0;
+   for (buf = 0; buf < CELL_NUM_BATCH_BUFFERS; buf++) {
+      cell->batch_buffer_size[buf] = 0;
+
+      /* init batch buffer status values,
+       * mark 0th buffer as used, rest as free.
+       */
+      for (spu = 0; spu < cell->num_spus; spu++) {
+         if (buf == 0)
+            cell->buffer_status[spu][buf][0] = CELL_BUFFER_STATUS_USED;
+         else
+            cell->buffer_status[spu][buf][0] = CELL_BUFFER_STATUS_FREE;
+      }
    }
 
+
 #if 0
    test_spus(cell);
 #endif
 
    return &cell->pipe;
 }
+
+
+#if 0
+/** [4] to ensure 16-byte alignment for each status word */
+uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BATCH_BUFFERS][4] ALIGN16_ATTRIB;
+#endif

src/mesa/pipe/cell/ppu/cell_context.h

@@ -38,6 +38,9 @@
 #include "pipe/cell/common.h"
 
 
+#define CELL_MAX_SPUS 6
+
+
 struct cell_vbuf_render;
 
 struct cell_vertex_shader_state
@@ -103,10 +106,14 @@ struct cell_context
    ubyte batch_buffer[CELL_NUM_BATCH_BUFFERS][CELL_BATCH_BUFFER_SIZE] ALIGN16_ATTRIB;
    int cur_batch;  /**< which batch buffer is being filled */
 
+   /** [4] to ensure 16-byte alignment for each status word */
+   uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BATCH_BUFFERS][4] ALIGN16_ATTRIB;
+
 };
 
 
 
+
 static INLINE struct cell_context *
 cell_context(struct pipe_context *pipe)
 {

src/mesa/pipe/cell/ppu/cell_spu.c

@@ -114,6 +114,7 @@ cell_start_spus(struct cell_context *cell)
       for (j = 0; j < CELL_NUM_BATCH_BUFFERS; j++) {
          cell_global.inits[i].batch_buffers[j] = cell->batch_buffer[j];
       }
+      cell_global.inits[i].buffer_status = &cell->buffer_status[0][0][0];
 
       cell_global.spe_contexts[i] = spe_context_create(0, NULL);
       if (!cell_global.spe_contexts[i]) {

src/mesa/pipe/cell/spu/spu_main.c

@@ -395,6 +395,38 @@ cmd_finish(void)
 }
 
 
+/**
+ * Tell the PPU that this SPU has finished copying a batch 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[]).
+ */
+static void
+release_batch_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);
+   uint *dst = spu.init.buffer_status + index;
+
+   ASSERT(buffer < CELL_NUM_BATCH_BUFFERS);
+
+   /*
+   printf("SPU %u: Set batch status buf=%u, index %u, at %p to FREE\n",
+          spu.init.id, buffer, index, dst);
+   */
+
+   mfc_put((void *) &status,    /* src in local memory */
+           (unsigned int) dst,  /* dst in main memory */
+           sizeof(status),      /* size */
+           TAG_MISC,            /* tag is unimportant */
+           0, /* tid */
+           0  /* rid */);
+}
+
+
 /**
  * Execute a batch of commands
  * The opcode param encodes the location of the buffer and its size.
@@ -429,9 +461,9 @@ cmd_batch(uint opcode)
            0  /* rid */);
    wait_on_mask(1 << TAG_BATCH_BUFFER);
 
-   /* send mbox message to indicate DMA completed */
-   /* XXX temporary */
-   spu_write_out_mbox(CELL_BATCH_FINISHED);
+   /* Tell PPU we're done copying the buffer to local store */
+   release_batch_buffer(buf);
+
 
    for (pos = 0; pos < usize; /* no incr */) {
       switch (buffer[pos]) {

src/mesa/pipe/cell/spu/spu_main.h

@@ -76,6 +76,8 @@ extern struct spu_global spu;
 #define TAG_WRITE_TILE_Z      15
 #define TAG_INDEX_BUFFER      16
 #define TAG_BATCH_BUFFER      17
+#define TAG_MISC              18
+
 
 /** The standard assert macro doesn't seem to work on SPUs */
 #define ASSERT(x) \