st/xorg: fix composite batching

quite a large performance optimization (text demo from 1.6fps to 9fps)

3 files changed, 18 insertions, 5 deletions

diff --git a/src/gallium/state_trackers/xorg/xorg_composite.c b/src/gallium/state_trackers/xorg/xorg_composite.c
index 8947d0a67c4..4f4b02cdbc8 100644
--- a/src/gallium/state_trackers/xorg/xorg_composite.c
+++ b/src/gallium/state_trackers/xorg/xorg_composite.c
@@ -287,6 +287,14 @@ bind_samplers(struct exa_context *exa, int op,
 
    exa->num_bound_samplers = 0;
 
+#if 0
+   if ((pSrc && (exa->pipe->is_texture_referenced(exa->pipe, pSrc->tex, 0, 0) &
+                 PIPE_REFERENCED_FOR_WRITE)) ||
+       (pMask && (exa->pipe->is_texture_referenced(exa->pipe, pMask->tex, 0, 0) &
+        PIPE_REFERENCED_FOR_WRITE)))
+      xorg_exa_flush(exa, PIPE_FLUSH_RENDER_CACHE, NULL);
+#endif
+
    memset(&src_sampler, 0, sizeof(struct pipe_sampler_state));
    memset(&mask_sampler, 0, sizeof(struct pipe_sampler_state));
 
@@ -461,7 +469,7 @@ void xorg_composite(struct exa_context *exa,
       if (exa->transform.has_mask)
          mask_matrix = exa->transform.mask;
 
-#if 1
+#if 0
       renderer_draw_textures(exa->renderer,
                              pos, width, height,
                              exa->bound_textures,
diff --git a/src/gallium/state_trackers/xorg/xorg_renderer.c b/src/gallium/state_trackers/xorg/xorg_renderer.c
index 947f4ca5313..c7a04836a5e 100644
--- a/src/gallium/state_trackers/xorg/xorg_renderer.c
+++ b/src/gallium/state_trackers/xorg/xorg_renderer.c
@@ -21,6 +21,8 @@ enum AxisOrientation {
 #define floatsEqual(x, y) (fabs(x - y) <= 0.00001f * MIN2(fabs(x), fabs(y)))
 #define floatIsZero(x) (floatsEqual((x) + 1, 1))
 
+#define NUM_COMPONENTS 4
+
 static INLINE boolean is_affine(float *matrix)
 {
    return floatIsZero(matrix[2]) && floatIsZero(matrix[5])
@@ -62,6 +64,7 @@ renderer_draw(struct xorg_renderer *r)
 {
    struct pipe_context *pipe = r->pipe;
    struct pipe_buffer *buf = 0;
+   int num_verts = r->num_vertices/(r->num_attributes * NUM_COMPONENTS);
 
    if (!r->num_vertices)
       return;
@@ -72,7 +75,7 @@ renderer_draw(struct xorg_renderer *r)
    if (buf) {
       util_draw_vertex_buffer(pipe, buf, 0,
                               PIPE_PRIM_QUADS,
-                              4,  /* verts */
+                              num_verts,  /* verts */
                               r->num_attributes); /* attribs/vert */
 
       pipe_buffer_reference(&buf, NULL);
@@ -84,8 +87,9 @@ renderer_draw_conditional(struct xorg_renderer *r,
                           int next_batch)
 {
    if (r->num_vertices + next_batch >= BUF_SIZE ||
-       (next_batch == 0 && r->num_vertices))
+       (next_batch == 0 && r->num_vertices)) {
       renderer_draw(r);
+   }
 }
 
 static void
@@ -892,6 +896,7 @@ void renderer_begin_textures(struct xorg_renderer *r,
                              int num_textures)
 {
    r->num_attributes = 1 + num_textures;
+   r->num_vertices = 0;
 }
 
 void renderer_texture(struct xorg_renderer *r,
diff --git a/src/gallium/state_trackers/xorg/xorg_renderer.h b/src/gallium/state_trackers/xorg/xorg_renderer.h
index 780d97fe85a..4cd929342ea 100644
--- a/src/gallium/state_trackers/xorg/xorg_renderer.h
+++ b/src/gallium/state_trackers/xorg/xorg_renderer.h
@@ -11,9 +11,9 @@ struct exa_pixmap_priv;
  * max number of attributes per vertex *
  * max number of components per attribute
  *
- * currently the max is 5 quads
+ * currently the max is 100 quads
  */
-#define BUF_SIZE (20 * 3 * 4)
+#define BUF_SIZE (100 * 4 * 3 * 4)
 
 struct xorg_renderer {
    struct pipe_context *pipe;