Cell: compute min index referenced in draw command, use it to reduce size of vertex data payload

3 files changed, 29 insertions, 6 deletions

diff --git a/src/mesa/pipe/cell/common.h b/src/mesa/pipe/cell/common.h
index f0d48ff403..90aa46a534 100644
--- a/src/mesa/pipe/cell/common.h
+++ b/src/mesa/pipe/cell/common.h
@@ -124,6 +124,8 @@ struct cell_command_render
    uint num_indexes;
    uint vertex_buf;  /**< which cell->buffer[] contains the vertex data */
    float xmin, dummy2, ymin, xmax, ymax;  /* XXX another dummy field */
+   uint dummy3;
+   uint min_index;
    boolean inline_verts;
 } ALIGN16_ATTRIB;
 
diff --git a/src/mesa/pipe/cell/ppu/cell_vbuf.c b/src/mesa/pipe/cell/ppu/cell_vbuf.c
index 9f737287ad..e63b34cf52 100644
--- a/src/mesa/pipe/cell/ppu/cell_vbuf.c
+++ b/src/mesa/pipe/cell/ppu/cell_vbuf.c
@@ -138,17 +138,24 @@ cell_vbuf_draw(struct vbuf_render *vbr,
    struct cell_context *cell = cvbr->cell;
    float xmin, ymin, xmax, ymax;
    uint i;
-   uint nr_vertices = 0;
+   uint nr_vertices = 0, min_index = ~0;
    const void *vertices = cvbr->vertex_buffer;
    const uint vertex_size = cvbr->vertex_size;
 
    for (i = 0; i < nr_indices; i++) {
       if (indices[i] > nr_vertices)
          nr_vertices = indices[i];
+      if (indices[i] < min_index)
+         min_index = indices[i];
    }
    nr_vertices++;
 
 #if 0
+   /*if (min_index > 0)*/
+      printf("%s min_index = %u\n", __FUNCTION__, min_index);
+#endif
+
+#if 0
    printf("cell_vbuf_draw() nr_indices = %u nr_verts = %u\n",
           nr_indices, nr_vertices);
    printf("  ");
@@ -169,7 +176,7 @@ cell_vbuf_draw(struct vbuf_render *vbr,
    /* compute x/y bounding box */
    xmin = ymin = 1e50;
    xmax = ymax = -1e50;
-   for (i = 0; i < nr_vertices; i++) {
+   for (i = min_index; i < nr_vertices; i++) {
       const float *v = (float *) ((ubyte *) vertices + i * vertex_size);
       if (v[0] < xmin)
          xmin = v[0];
@@ -204,6 +211,7 @@ cell_vbuf_draw(struct vbuf_render *vbr,
       render->prim_type = cvbr->prim;
 
       render->num_indexes = nr_indices;
+      render->min_index = min_index;
 
       /* append indices after render command */
       memcpy(render + 1, indices, nr_indices * 2);
@@ -214,6 +222,7 @@ cell_vbuf_draw(struct vbuf_render *vbr,
       render->vertex_size = 4 * cell->vertex_info.size;
       render->num_verts = nr_vertices;
       if (ALLOW_INLINE_VERTS &&
+          min_index == 0 &&
           vertex_bytes <= cell_batch_free_space(cell)) {
          /* vertex data inlined, after indices */
          void *dst = cell_batch_alloc(cell, vertex_bytes);
diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
index 5a5b17dd89..3c9efb4741 100644
--- a/src/mesa/pipe/cell/spu/spu_main.c
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -253,7 +253,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
    /* we'll DMA into these buffers */
    ubyte vertex_data[CELL_BUFFER_SIZE] ALIGN16_ATTRIB;
    const uint vertex_size = render->vertex_size; /* in bytes */
-   const uint total_vertex_bytes = render->num_verts * vertex_size;
+   /*const*/ uint total_vertex_bytes = render->num_verts * vertex_size;
    const ubyte *vertices;
    const ushort *indexes;
    uint i, j;
@@ -289,9 +289,21 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
    }
    else {
       /* Begin DMA fetch of vertex buffer */
-      void *src = spu.init.buffers[render->vertex_buf];
-      mfc_get(vertex_data,  /* dest */
-              (unsigned int) src,
+      ubyte *src = spu.init.buffers[render->vertex_buf];
+      ubyte *dest = vertex_data;
+
+      /* skip vertex data we won't use */
+#if 01
+      src += render->min_index * vertex_size;
+      dest += render->min_index * vertex_size;
+      total_vertex_bytes -= render->min_index * vertex_size;
+#endif
+      ASSERT(total_vertex_bytes % 16 == 0);
+      ASSERT_ALIGN16(dest);
+      ASSERT_ALIGN16(src);
+
+      mfc_get(dest,   /* in vertex_data[] array */
+              (unsigned int) src,  /* src in main memory */
               total_vertex_bytes,  /* size */
               TAG_VERTEX_BUFFER,
               0, /* tid */