use a shadow buffer for vertex data to optimize memory access
authorChristian König <deathsimple@vodafone.de>
Sun, 28 Nov 2010 13:48:31 +0000 (14:48 +0100)
committerChristian König <deathsimple@vodafone.de>
Fri, 3 Dec 2010 18:04:00 +0000 (19:04 +0100)
src/gallium/auxiliary/Makefile
src/gallium/auxiliary/vl/vl_idct.c
src/gallium/auxiliary/vl/vl_idct.h
src/gallium/auxiliary/vl/vl_types.h
src/gallium/auxiliary/vl/vl_vertex_buffers.c [new file with mode: 0644]
src/gallium/auxiliary/vl/vl_vertex_buffers.h [new file with mode: 0644]

index bf90a704dd42a196d584a7f47978c9dee933d470..997478f06644b49daabc5548cc32a9cba2624992 100644 (file)
@@ -147,7 +147,8 @@ C_SOURCES = \
        vl/vl_mpeg12_mc_renderer.c \
        vl/vl_compositor.c \
        vl/vl_csc.c \
-        vl/vl_idct.c
+        vl/vl_idct.c \
+        vl/vl_vertex_buffers.c
 
 GALLIVM_SOURCES = \
         gallivm/lp_bld_arit.c \
index 2466d5b751be1a68b099d0c1b5db23ae1b72bcdb..22feff8d8d4284d13e64a1f90264ec70379f3057 100644 (file)
@@ -26,6 +26,7 @@
  **************************************************************************/
 
 #include "vl_idct.h"
+#include "vl_vertex_buffers.h"
 #include "util/u_draw.h"
 #include <assert.h>
 #include <pipe/p_context.h>
@@ -78,11 +79,6 @@ static const float const_matrix[8][8] = {
    {  0.0975451f, -0.2777850f,  0.4157350f, -0.4903930f,  0.4903930f, -0.4157350f,  0.277786f, -0.0975458f }
 };
 
-/* vertices for a quad covering a block */
-static const struct vertex2f const_quad[4] = {
-   {0.0f, 0.0f}, {1.0f, 0.0f}, {1.0f, 1.0f}, {0.0f, 1.0f}
-};
-
 static void *
 create_vert_shader(struct vl_idct *idct, bool calc_src_cords)
 {
@@ -409,11 +405,6 @@ init_buffers(struct vl_idct *idct)
    struct pipe_vertex_element vertex_elems[2];
    unsigned i;
 
-   idct->max_blocks =
-      align(idct->destination->width0, BLOCK_WIDTH) / BLOCK_WIDTH *
-      align(idct->destination->height0, BLOCK_HEIGHT) / BLOCK_HEIGHT *
-      idct->destination->depth0;
-
    memset(&template, 0, sizeof(struct pipe_resource));
    template.last_level = 0;
    template.depth0 = 1;
@@ -443,15 +434,7 @@ init_buffers(struct vl_idct *idct)
       idct->sampler_views.all[i] = idct->pipe->create_sampler_view(idct->pipe, idct->textures.all[i], &sampler_view);
    }
 
-   idct->vertex_bufs.individual.quad.stride = sizeof(struct vertex2f);
-   idct->vertex_bufs.individual.quad.max_index = 4 * idct->max_blocks - 1;
-   idct->vertex_bufs.individual.quad.buffer_offset = 0;
-   idct->vertex_bufs.individual.quad.buffer = pipe_buffer_create
-   (
-      idct->pipe->screen,
-      PIPE_BIND_VERTEX_BUFFER,
-      sizeof(struct vertex2f) * 4 * idct->max_blocks
-   );
+   idct->vertex_bufs.individual.quad = vl_vb_upload_quads(idct->pipe, idct->max_blocks);
 
    if(idct->vertex_bufs.individual.quad.buffer == NULL)
       return false;
@@ -503,36 +486,12 @@ cleanup_buffers(struct vl_idct *idct)
    pipe_resource_reference(&idct->vertex_bufs.individual.pos.buffer, NULL);
 }
 
-static void
-init_constants(struct vl_idct *idct)
-{
-   struct pipe_transfer *buf_transfer;
-   struct vertex2f *v;
-
-   unsigned i;
-
-   /* quad vectors */
-   v = pipe_buffer_map
-   (
-      idct->pipe,
-      idct->vertex_bufs.individual.quad.buffer,
-      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
-      &buf_transfer
-   );
-   for ( i = 0; i < idct->max_blocks; ++i)
-     memcpy(v + i * 4, &const_quad, sizeof(const_quad));
-   pipe_buffer_unmap(idct->pipe, idct->vertex_bufs.individual.quad.buffer, buf_transfer);
-}
-
 static void
 init_state(struct vl_idct *idct)
 {
    struct pipe_sampler_state sampler;
    unsigned i;
 
-   idct->num_blocks = 0;
-   idct->num_empty_blocks = 0;
-
    idct->viewport[0].scale[0] = idct->textures.individual.intermediate->width0;
    idct->viewport[0].scale[1] = idct->textures.individual.intermediate->height0;
 
@@ -674,26 +633,11 @@ xfer_buffers_map(struct vl_idct *idct)
    );
 
    idct->texels = idct->pipe->transfer_map(idct->pipe, idct->tex_transfer);
-
-   idct->vectors = pipe_buffer_map
-   (
-      idct->pipe,
-      idct->vertex_bufs.individual.pos.buffer,
-      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
-      &idct->vec_transfer
-   );
-
-   idct->next_empty_block.l_x = ~1;
-   idct->next_empty_block.l_y = ~1;
-   idct->next_empty_block.r_x = ~1;
-   idct->next_empty_block.r_y = ~1;
 }
 
 static void
 xfer_buffers_unmap(struct vl_idct *idct)
 {
-   pipe_buffer_unmap(idct->pipe, idct->vertex_bufs.individual.pos.buffer, idct->vec_transfer);
-
    idct->pipe->transfer_unmap(idct->pipe, idct->tex_transfer);
    idct->pipe->transfer_destroy(idct->pipe, idct->tex_transfer);
 }
@@ -708,6 +652,11 @@ vl_idct_init(struct vl_idct *idct, struct pipe_context *pipe, struct pipe_resour
    pipe_resource_reference(&idct->textures.individual.transpose, matrix);
    pipe_resource_reference(&idct->destination, dst);
 
+   idct->max_blocks =
+      align(idct->destination->width0, BLOCK_WIDTH) / BLOCK_WIDTH *
+      align(idct->destination->height0, BLOCK_HEIGHT) / BLOCK_HEIGHT *
+      idct->destination->depth0;
+
    if(!init_buffers(idct))
       return false;
 
@@ -716,9 +665,21 @@ vl_idct_init(struct vl_idct *idct, struct pipe_context *pipe, struct pipe_resour
       return false;
    }
 
+   if(!vl_vb_init(&idct->blocks, idct->max_blocks)) {
+      cleanup_shaders(idct);
+      cleanup_buffers(idct);
+      return false;
+   }
+
+   if(!vl_vb_init(&idct->empty_blocks, idct->max_blocks)) {
+      vl_vb_cleanup(&idct->blocks);
+      cleanup_shaders(idct);
+      cleanup_buffers(idct);
+      return false;
+   }
+
    init_state(idct);
 
-   init_constants(idct);
    xfer_buffers_map(idct);
 
    return true;
@@ -727,6 +688,8 @@ vl_idct_init(struct vl_idct *idct, struct pipe_context *pipe, struct pipe_resour
 void
 vl_idct_cleanup(struct vl_idct *idct)
 {
+   vl_vb_cleanup(&idct->blocks);
+   vl_vb_cleanup(&idct->empty_blocks);
    cleanup_shaders(idct);
    cleanup_buffers(idct);
 
@@ -735,43 +698,9 @@ vl_idct_cleanup(struct vl_idct *idct)
    pipe_resource_reference(&idct->destination, NULL);
 }
 
-static void
-flush_empty_block(struct vl_idct *idct, unsigned new_x, unsigned new_y)
-{
-   if (idct->next_empty_block.l_x == ~1 ||
-       idct->next_empty_block.l_y == ~1) {
-   
-      idct->next_empty_block.l_x = new_x;
-      idct->next_empty_block.l_y = new_y;
-
-   } else if (idct->next_empty_block.r_x != (new_x - 1) ||
-              idct->next_empty_block.r_y != new_y) {
-
-      struct vertex2f l, r, *v_dst;
-
-      v_dst = idct->vectors + (idct->max_blocks - idct->num_empty_blocks) * 4 - 4;
-
-      l.x = idct->next_empty_block.l_x;
-      l.y = idct->next_empty_block.l_y;
-      r.x = idct->next_empty_block.r_x;
-      r.y = idct->next_empty_block.r_y;
-      v_dst[0] = v_dst[3] = l;
-      v_dst[1] = v_dst[2] = r;
-
-      idct->next_empty_block.l_x = new_x;
-      idct->next_empty_block.l_y = new_y;
-      idct->num_empty_blocks++;
-   }
-
-   idct->next_empty_block.r_x = new_x;
-   idct->next_empty_block.r_y = new_y;
-}
-
 void
 vl_idct_add_block(struct vl_idct *idct, unsigned x, unsigned y, short *block)
 {
-   struct vertex2f v, *v_dst;
-
    unsigned tex_pitch;
    short *texels;
 
@@ -786,32 +715,38 @@ vl_idct_add_block(struct vl_idct *idct, unsigned x, unsigned y, short *block)
       for (i = 0; i < BLOCK_HEIGHT; ++i)
          memcpy(texels + i * tex_pitch, block + i * BLOCK_WIDTH, BLOCK_WIDTH * sizeof(short));
 
-      /* non empty blocks fills the vector buffer from left to right */
-      v_dst = idct->vectors + idct->num_blocks * 4;
-
-      idct->num_blocks++;
-
-      v.x = x;
-      v.y = y;
-
-      for (i = 0; i < 4; ++i) {
-         v_dst[i] = v;
-      }
-
+      vl_vb_add_block(&idct->blocks, false, x, y);
    } else {
 
-      /* while empty blocks fills the vector buffer from right to left */
-      flush_empty_block(idct, x, y);
+      vl_vb_add_block(&idct->empty_blocks, true, x, y);
    }
 }
 
 void
 vl_idct_flush(struct vl_idct *idct)
 {
-   flush_empty_block(idct, ~1, ~1);
+   struct pipe_transfer *vec_transfer;
+   struct quadf *vectors;
+   unsigned num_blocks, num_empty_blocks;
+
+   assert(idct);
+
+   vectors = pipe_buffer_map
+   (
+      idct->pipe,
+      idct->vertex_bufs.individual.pos.buffer,
+      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
+      &vec_transfer
+   );
+
+   num_blocks = vl_vb_upload(&idct->blocks, vectors);
+   num_empty_blocks = vl_vb_upload(&idct->empty_blocks, vectors + num_blocks);
+
+   pipe_buffer_unmap(idct->pipe, idct->vertex_bufs.individual.pos.buffer, vec_transfer);
+
    xfer_buffers_unmap(idct);
 
-   if(idct->num_blocks > 0) {
+   if(num_blocks > 0) {
 
       /* first stage */
       idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[0]);
@@ -824,7 +759,7 @@ vl_idct_flush(struct vl_idct *idct)
       idct->pipe->bind_vs_state(idct->pipe, idct->matrix_vs);
       idct->pipe->bind_fs_state(idct->pipe, idct->matrix_fs);
 
-      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, idct->num_blocks * 4);
+      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, num_blocks * 4);
 
       /* second stage */
       idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[1]);
@@ -837,10 +772,10 @@ vl_idct_flush(struct vl_idct *idct)
       idct->pipe->bind_vs_state(idct->pipe, idct->transpose_vs);
       idct->pipe->bind_fs_state(idct->pipe, idct->transpose_fs);
 
-      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, idct->num_blocks * 4);
+      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, 0, num_blocks * 4);
    }
 
-   if(idct->num_empty_blocks > 0) {
+   if(num_empty_blocks > 0) {
 
       /* empty block handling */
       idct->pipe->set_framebuffer_state(idct->pipe, &idct->fb_state[1]);
@@ -851,12 +786,8 @@ vl_idct_flush(struct vl_idct *idct)
       idct->pipe->bind_vs_state(idct->pipe, idct->eb_vs);
       idct->pipe->bind_fs_state(idct->pipe, idct->eb_fs);
 
-      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS,
-         (idct->max_blocks - idct->num_empty_blocks) * 4,
-         idct->num_empty_blocks * 4);
+      util_draw_arrays(idct->pipe, PIPE_PRIM_QUADS, num_blocks * 4, num_empty_blocks * 4);
    }
 
-   idct->num_blocks = 0;
-   idct->num_empty_blocks = 0;
    xfer_buffers_map(idct);
 }
index c26f5cb8c09592ef5e0bc96d05a86e4b34fe2396..94a5c73977f39102b5c0abe8cd6fbfdf0ce347f2 100644 (file)
@@ -29,6 +29,7 @@
 #define vl_idct_h
 
 #include <pipe/p_state.h>
+#include "vl_vertex_buffers.h"
 
 struct vl_idct
 {
@@ -82,20 +83,11 @@ struct vl_idct
       struct { struct pipe_vertex_buffer quad, pos; } individual;
    } vertex_bufs;
 
-   unsigned num_blocks;
-
-   struct
-   {
-      unsigned l_x, l_y, r_x, r_y;
-   } next_empty_block;
-
-   unsigned num_empty_blocks;
+   struct vl_vertex_buffer blocks;
+   struct vl_vertex_buffer empty_blocks;
 
    struct pipe_transfer *tex_transfer;
    short *texels;
-
-   struct pipe_transfer *vec_transfer;
-   struct vertex2f *vectors;
 };
 
 struct pipe_resource *vl_idct_upload_matrix(struct pipe_context *pipe);
index ce1755468942451d5f37175add8dc37e554d321a..eeabd43cb239cf59d4f8aaad0cdea5bf99cc6728 100644 (file)
@@ -38,4 +38,9 @@ struct vertex4f
    float x, y, z, w;
 };
 
+struct quadf
+{
+   struct vertex2f bl, tl, tr, br;
+};
+
 #endif /* vl_types_h */
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.c b/src/gallium/auxiliary/vl/vl_vertex_buffers.c
new file mode 100644 (file)
index 0000000..6df11db
--- /dev/null
@@ -0,0 +1,116 @@
+/**************************************************************************
+ *
+ * Copyright 2010 Christian König
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+#include <assert.h>
+#include <pipe/p_context.h>
+#include <pipe/p_screen.h>
+#include <util/u_memory.h>
+#include <util/u_inlines.h>
+#include "vl_vertex_buffers.h"
+#include "vl_types.h"
+
+/* vertices for a quad covering a block */
+static const struct quadf const_quad = {
+   {0.0f, 1.0f}, {0.0f, 0.0f}, {1.0f, 0.0f}, {1.0f, 1.0f}
+};
+
+struct pipe_vertex_buffer
+vl_vb_upload_quads(struct pipe_context *pipe, unsigned max_blocks)
+{
+   struct pipe_vertex_buffer quad;
+   struct pipe_transfer *buf_transfer;
+   struct quadf *v;
+
+   unsigned i;
+
+   assert(pipe);
+   assert(max_blocks);
+
+   /* create buffer */
+   quad.stride = sizeof(struct vertex2f);
+   quad.max_index = 4 * max_blocks - 1;
+   quad.buffer_offset = 0;
+   quad.buffer = pipe_buffer_create
+   (
+      pipe->screen,
+      PIPE_BIND_VERTEX_BUFFER,
+      sizeof(struct vertex2f) * 4 * max_blocks
+   );
+
+   if(!quad.buffer)
+      return quad;
+
+   /* and fill it */
+   v = pipe_buffer_map
+   (
+      pipe,
+      quad.buffer,
+      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
+      &buf_transfer
+   );
+
+   for ( i = 0; i < max_blocks; ++i)
+     memcpy(v + i, &const_quad, sizeof(const_quad));
+
+   pipe_buffer_unmap(pipe, quad.buffer, buf_transfer);
+
+   return quad;
+}
+
+bool
+vl_vb_init(struct vl_vertex_buffer *buffer, unsigned max_blocks)
+{
+   assert(buffer);
+
+   buffer->num_blocks = 0;
+   buffer->blocks = MALLOC(max_blocks * sizeof(struct quadf));
+   return buffer->blocks != NULL;
+}
+
+unsigned
+vl_vb_upload(struct vl_vertex_buffer *buffer, struct quadf *dst)
+{
+   unsigned todo;
+
+   assert(buffer);
+
+   todo = buffer->num_blocks;
+   buffer->num_blocks = 0;
+
+   if(todo)
+      memcpy(dst, buffer->blocks, sizeof(struct quadf) * todo);
+
+   return todo;
+}
+
+void
+vl_vb_cleanup(struct vl_vertex_buffer *buffer)
+{
+   assert(buffer);
+
+   FREE(buffer->blocks);
+}
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.h b/src/gallium/auxiliary/vl/vl_vertex_buffers.h
new file mode 100644 (file)
index 0000000..43ddc34
--- /dev/null
@@ -0,0 +1,75 @@
+/**************************************************************************
+ *
+ * Copyright 2010 Christian König
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+#ifndef vl_vertex_buffers_h
+#define vl_vertex_buffers_h
+
+#include <assert.h>
+#include <pipe/p_state.h>
+#include "vl_types.h"
+
+struct vl_vertex_buffer
+{
+   unsigned num_blocks;
+   struct quadf *blocks;
+};
+
+struct pipe_vertex_buffer vl_vb_upload_quads(struct pipe_context *pipe, unsigned max_blocks);
+
+bool vl_vb_init(struct vl_vertex_buffer *buffer, unsigned max_blocks);
+
+static inline bool
+vl_vb_add_block(struct vl_vertex_buffer *buffer, bool allow_merge, signed x, signed y)
+{
+   struct quadf *quad;
+
+   assert(buffer);
+
+   allow_merge &= buffer->num_blocks > 0;
+   if (allow_merge) {
+
+      quad = buffer->blocks + buffer->num_blocks - 1;
+      if(quad->tr.x == (x - 1) && quad->br.x == (x - 1) && 
+         quad->tr.y == y && quad->br.y == y) {
+
+         quad->tr.x = quad->br.x = x;
+         quad->tr.y = quad->br.y = y;
+         return true;
+      } 
+   }
+
+   quad = buffer->blocks + buffer->num_blocks;
+   quad->bl.x = quad->tl.x = quad->tr.x = quad->br.x = x;
+   quad->bl.y = quad->tl.y = quad->tr.y = quad->br.y = y;
+   buffer->num_blocks++;
+   return false;
+}
+
+unsigned vl_vb_upload(struct vl_vertex_buffer *buffer, struct quadf *dst);
+
+void vl_vb_cleanup(struct vl_vertex_buffer *buffer);
+
+#endif