Cell: after sending a batch, wait for a DMA completion signal.

This fixes sporadic rendering glitches.
Using a mailbox msg for now, until spe_mfcio_tag_status_read() or similar
is found to work.

src/mesa/pipe/cell/common.h

@@ -65,6 +65,8 @@
 #define CELL_NUM_BATCH_BUFFERS 2
 #define CELL_BATCH_BUFFER_SIZE 1024  /**< 16KB would be the max */
 
+#define CELL_BATCH_FINISHED 0x1234   /**< mbox message */
+
 
 /**
  * Tell SPUs about the framebuffer size, location

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

@@ -43,16 +43,27 @@ cell_batch_flush(struct cell_context *cell)
 
    assert(batch < CELL_NUM_BATCH_BUFFERS);
 
-   /*printf("cell_batch_dispatch: buf %u, size %u\n", batch, size);*/
-          
+   /*
+   printf("cell_batch_dispatch: buf %u at %p, size %u\n",
+          batch, &cell->batch_buffer[batch][0], size);
+   */
+     
    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);
    }
 
-   /* XXX wait on DMA xfer of prev buffer to complete */
+   /* 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...
+    */
+   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 */

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

@@ -46,7 +46,7 @@ helpful headers:
 /opt/ibm/cell-sdk/prototype/sysroot/usr/include/libmisc.h
 */
 
-static boolean Debug = TRUE;
+static boolean Debug = FALSE;
 
 struct spu_global spu;
 
@@ -106,7 +106,7 @@ really_clear_tiles(uint surfaceIndex)
       }
    }
 
-#if 01
+#if 0
    wait_on_mask(1 << TAG_SURFACE_CLEAR);
 #endif
 }
@@ -165,6 +165,9 @@ cmd_clear_surface(const struct cell_command_clear_surface *clear)
 #if 0
    wait_on_mask(1 << TAG_SURFACE_CLEAR);
 #endif
+
+   if (Debug)
+      printf("SPU %u: CLEAR SURF done\n", spu.init.id);
 }
 
 
@@ -222,8 +225,10 @@ cmd_render(const struct cell_command_render *render)
              render->prim_type,
              render->num_verts,
              render->num_indexes);
+      /*
       printf("       bound: %g, %g .. %g, %g\n",
              render->xmin, render->ymin, render->xmax, render->ymax);
+      */
    }
 
    ASSERT_ALIGN16(render->vertex_data);
@@ -244,6 +249,7 @@ cmd_render(const struct cell_command_render *render)
           render->index_data, render->vertex_data, vertex_bytes, index_bytes);
    */
 
+   ASSERT(vertex_bytes % 16 == 0);
    /* get vertex data from main memory */
    mfc_get(vertex_data,  /* dest */
            (unsigned int) render->vertex_data,  /* src */
@@ -252,6 +258,8 @@ cmd_render(const struct cell_command_render *render)
            0, /* tid */
            0  /* rid */);
 
+   ASSERT(index_bytes % 16 == 0);
+
    /* get index data from main memory */
    mfc_get(indexes,  /* dest */
            (unsigned int) render->index_data,  /* src */
@@ -330,6 +338,10 @@ cmd_render(const struct cell_command_render *render)
          wait_on_mask(1 << TAG_WRITE_TILE_Z);
       }
    }
+
+   if (Debug)
+      printf("SPU %u: RENDER done\n",
+             spu.init.id);
 }
 
 
@@ -406,6 +418,9 @@ cmd_batch(uint opcode)
 
    size = (size + 0xf) & ~0xf;
 
+   ASSERT(size % 16 == 0);
+   ASSERT((unsigned int) spu.init.batch_buffers[buf] % 16 == 0);
+
    mfc_get(buffer,  /* dest */
            (unsigned int) spu.init.batch_buffers[buf],  /* src */
            size,
@@ -414,6 +429,10 @@ 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);
+
    for (pos = 0; pos < usize; /* no incr */) {
       switch (buffer[pos]) {
       case CELL_CMD_FRAMEBUFFER:

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

@@ -77,9 +77,11 @@ put_tile(uint tx, uint ty, const uint *tile, int tag, int zBuf)
    ASSERT(ty < spu.fb.height_tiles);
    ASSERT_ALIGN16(tile);
    /*
-   printf("put_tile:  src: %p  dst: 0x%x  size: %d\n",
+   printf("SPU %u: put_tile:  src: %p  dst: 0x%x  size: %d\n",
+          spu.init.id,
           tile, (unsigned int) dst, bytesPerTile);
    */
+
    mfc_put((void *) tile,  /* src in local memory */
            (unsigned int) dst,  /* dst in main memory */
            bytesPerTile,