svga: implement an index buffer translation cache

Some OpenGL apps, like Cinebench R15, have many glDrawElements(GL_QUADS)
calls.  Since we don't directly support quads we have to convert these
calls into GL_TRIANGLES which involves generating a new index buffer.

This patch saves the new/translated index buffer in the hope that it
can be reused for a later draw call.

Cinebench R15 increases by about 20% with this change.
The NobelClinician Viewer app also hits this code.
Tested with full piglit run.

Reviewed-by: Charmaine Lee <charmainel@vmware.com>

src/gallium/drivers/svga/svga_draw_elements.c

@@ -37,26 +37,66 @@
 #include "svga_hw_reg.h"
 
 
+/**
+ * Return a new index buffer which contains a translation of the original
+ * index buffer.  An example of a translation is converting from QUAD
+ * primitives to TRIANGLE primitives.  Each set of four indexes for a quad
+ * will be converted to six indices for two triangles.
+ *
+ * Before generating the new index buffer we'll check if the incoming
+ * buffer already has a translated buffer that can be re-used.
+ * This benefits demos like Cinebench R15 which has many
+ * glDrawElements(GL_QUADS) commands (we can't draw quads natively).
+ *
+ * \param offset  offset in bytes to first index to translate in src buffer
+ * \param orig_prim  original primitive type (like PIPE_PRIM_QUADS)
+ * \param gen_prim  new/generated primitive type (like PIPE_PRIM_TRIANGLES)
+ * \param orig_nr  number of indexes to translate in source buffer
+ * \param gen_nr  number of indexes to write into new/dest buffer
+ * \param index_size  bytes per index (2 or 4)
+ * \param translate  the translation function from the u_translate module
+ * \param out_buf  returns the new/translated index buffer
+ * \return error code to indicate success failure
+ */
 static enum pipe_error
 translate_indices(struct svga_hwtnl *hwtnl, struct pipe_resource *src,
-                  unsigned offset, enum pipe_prim_type prim, unsigned nr,
+                  unsigned offset,
+                  enum pipe_prim_type orig_prim, enum pipe_prim_type gen_prim,
+                  unsigned orig_nr, unsigned gen_nr,
                   unsigned index_size,
                   u_translate_func translate, struct pipe_resource **out_buf)
 {
    struct pipe_context *pipe = &hwtnl->svga->pipe;
+   struct svga_screen *screen = svga_screen(pipe->screen);
+   struct svga_buffer *src_sbuf = svga_buffer(src);
    struct pipe_transfer *src_transfer = NULL;
    struct pipe_transfer *dst_transfer = NULL;
-   unsigned size = index_size * nr;
+   unsigned size = index_size * gen_nr;
    const void *src_map = NULL;
    struct pipe_resource *dst = NULL;
    void *dst_map = NULL;
 
+   assert(index_size == 2 || index_size == 4);
+
+   if (!screen->debug.no_cache_index_buffers) {
+      /* Check if we already have a translated index buffer */
+      if (src_sbuf->translated_indices.buffer &&
+          src_sbuf->translated_indices.orig_prim == PIPE_PRIM_QUADS &&
+          src_sbuf->translated_indices.new_prim == gen_prim &&
+          src_sbuf->translated_indices.offset == offset &&
+          src_sbuf->translated_indices.count == orig_nr &&
+          src_sbuf->translated_indices.index_size == index_size) {
+         pipe_resource_reference(out_buf, src_sbuf->translated_indices.buffer);
+         return PIPE_OK;
+      }
+   }
+
    /* Need to trim vertex count to make sure we don't write too much data
     * to the dst buffer in the translate() call.
     */
-   u_trim_pipe_prim(prim, &nr);
+   u_trim_pipe_prim(gen_prim, &gen_nr);
 
-   size = index_size * nr;
+   size = index_size * gen_nr;
 
    dst = pipe_buffer_create(pipe->screen,
                             PIPE_BIND_INDEX_BUFFER, PIPE_USAGE_DEFAULT, size);
@@ -71,12 +111,25 @@ translate_indices(struct svga_hwtnl *hwtnl, struct pipe_resource *src,
    if (!dst_map)
       goto fail;
 
-   translate((const char *) src_map + offset, 0, 0, nr, 0, dst_map);
+   translate((const char *) src_map + offset, 0, 0, gen_nr, 0, dst_map);
 
    pipe_buffer_unmap(pipe, src_transfer);
    pipe_buffer_unmap(pipe, dst_transfer);
 
    *out_buf = dst;
+
+   if (!screen->debug.no_cache_index_buffers) {
+      /* Save the new, translated index buffer in the hope we can use it
+       * again in the future.
+       */
+      pipe_resource_reference(&src_sbuf->translated_indices.buffer, dst);
+      src_sbuf->translated_indices.orig_prim = PIPE_PRIM_QUADS;
+      src_sbuf->translated_indices.new_prim = gen_prim;
+      src_sbuf->translated_indices.offset = offset;
+      src_sbuf->translated_indices.count = orig_nr;
+      src_sbuf->translated_indices.index_size = index_size;
+   }
+
    return PIPE_OK;
 
  fail:
@@ -186,7 +239,9 @@ svga_hwtnl_draw_range_elements(struct svga_hwtnl *hwtnl,
       ret = translate_indices(hwtnl,
                               index_buffer,
                               start * index_size,
-                              gen_prim, gen_nr, gen_size, gen_func, &gen_buf);
+                              prim, gen_prim,
+                              count, gen_nr, gen_size,
+                              gen_func, &gen_buf);
       if (ret != PIPE_OK)
          goto done;
 

src/gallium/drivers/svga/svga_resource_buffer.c

@@ -98,6 +98,13 @@ svga_buffer_transfer_map(struct pipe_context *pipe,
    transfer->stride = 0;
    transfer->layer_stride = 0;
 
+   if (usage & PIPE_TRANSFER_WRITE) {
+      /* If we write to the buffer for any reason, free any saved translated
+       * vertices.
+       */
+      pipe_resource_reference(&sbuf->translated_indices.buffer, NULL);
+   }
+
    if ((usage & PIPE_TRANSFER_READ) && sbuf->dirty) {
       enum pipe_error ret;
 
@@ -361,6 +368,8 @@ svga_buffer_destroy( struct pipe_screen *screen,
    if (sbuf->swbuf && !sbuf->user)
       align_free(sbuf->swbuf);
 
+   pipe_resource_reference(&sbuf->translated_indices.buffer, NULL);
+
    ss->hud.total_resource_bytes -= sbuf->size;
    assert(ss->hud.num_resources > 0);
    if (ss->hud.num_resources > 0)

src/gallium/drivers/svga/svga_resource_buffer.h

@@ -192,6 +192,17 @@ struct svga_buffer
    unsigned size;  /**< Approximate size in bytes */
 
    boolean dirty;  /**< Need to do a readback before mapping? */
+
+   /** In some cases we try to keep the results of the translate_indices()
+    * function from svga_draw_elements.c
+    */
+   struct {
+      enum pipe_prim_type orig_prim, new_prim;
+      struct pipe_resource *buffer;
+      unsigned index_size;
+      unsigned offset;  /**< first index */
+      unsigned count;   /**< num indices */
+   } translated_indices;
 };
 
 

src/gallium/drivers/svga/svga_screen.c

@@ -944,6 +944,8 @@ svga_screen_create(struct svga_winsys_screen *sws)
       debug_get_bool_option("SVGA_NO_SURFACE_VIEW", FALSE);
    svgascreen->debug.no_sampler_view =
       debug_get_bool_option("SVGA_NO_SAMPLER_VIEW", FALSE);
+   svgascreen->debug.no_cache_index_buffers =
+      debug_get_bool_option("SVGA_NO_CACHE_INDEX_BUFFERS", FALSE);
 
    screen = &svgascreen->screen;
 

src/gallium/drivers/svga/svga_screen.h

@@ -62,6 +62,7 @@ struct svga_screen
       boolean no_surface_view;
       boolean force_sampler_view;
       boolean no_sampler_view;
+      boolean no_cache_index_buffers;
    } debug;
 
    unsigned texture_timestamp;