llvmpipe: use single swizzled tile

Use a single swizzled tile per colorbuf (and per thread) to avoid
accumulating large amounts of cached swizzled data.

Now that the SSE3 code has been merged to master, the performance delta
of this change is minimal, the main benefit is reduced memory usage
due to no longer keeping swizzled copies of render targets.

It's clear from the performance of the in-place version of this code
that there is still quite a bit of time being spent swizzling &
unswizzling, but it's not clear exactly how to reduce that.

1 files changed, 88 insertions, 0 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_texture.c b/src/gallium/drivers/llvmpipe/lp_texture.c
index d236bad69d..bbd834519a 100644
--- a/src/gallium/drivers/llvmpipe/lp_texture.c
+++ b/src/gallium/drivers/llvmpipe/lp_texture.c
@@ -1209,6 +1209,94 @@ llvmpipe_get_texture_tile(struct llvmpipe_resource *lpr,
 
 
 /**
+ * Get pointer to tiled data for rendering.
+ * \return pointer to the tiled data at the given tile position
+ */
+void
+llvmpipe_unswizzle_cbuf_tile(struct llvmpipe_resource *lpr,
+                             unsigned face_slice, unsigned level,
+                             unsigned x, unsigned y,
+                             uint8_t *tile)
+{
+   struct llvmpipe_texture_image *linear_img = &lpr->linear[level];
+   const unsigned tx = x / TILE_SIZE, ty = y / TILE_SIZE;
+   uint8_t *linear_image;
+
+   assert(x % TILE_SIZE == 0);
+   assert(y % TILE_SIZE == 0);
+
+   if (!linear_img->data) {
+      /* allocate memory for the linear image now */
+      alloc_image_data(lpr, level, LP_TEX_LAYOUT_LINEAR);
+   }
+
+   /* compute address of the slice/face of the image that contains the tile */
+   linear_image = llvmpipe_get_texture_image_address(lpr, face_slice, level,
+                                                     LP_TEX_LAYOUT_LINEAR);
+
+   {
+      uint ii = x, jj = y;
+      uint tile_offset = jj / TILE_SIZE + ii / TILE_SIZE;
+      uint byte_offset = tile_offset * TILE_SIZE * TILE_SIZE * 4;
+      
+      /* Note that lp_tiled_to_linear expects the tile parameter to
+       * point at the first tile in a whole-image sized array.  In
+       * this code, we have only a single tile and have to do some
+       * pointer arithmetic to figure out where the "image" would have
+       * started.
+       */
+      lp_tiled_to_linear(tile - byte_offset, linear_image,
+                         x, y, TILE_SIZE, TILE_SIZE,
+                         lpr->base.format,
+                         lpr->row_stride[level],
+                         1);       /* tiles per row */
+   }
+
+   llvmpipe_set_texture_tile_layout(lpr, face_slice, level, tx, ty,
+                                    LP_TEX_LAYOUT_LINEAR);
+}
+
+
+/**
+ * Get pointer to tiled data for rendering.
+ * \return pointer to the tiled data at the given tile position
+ */
+void
+llvmpipe_swizzle_cbuf_tile(struct llvmpipe_resource *lpr,
+                           unsigned face_slice, unsigned level,
+                           unsigned x, unsigned y,
+                           uint8_t *tile)
+{
+   uint8_t *linear_image;
+
+   assert(x % TILE_SIZE == 0);
+   assert(y % TILE_SIZE == 0);
+
+   /* compute address of the slice/face of the image that contains the tile */
+   linear_image = llvmpipe_get_texture_image_address(lpr, face_slice, level,
+                                                     LP_TEX_LAYOUT_LINEAR);
+
+   if (linear_image) {
+      uint ii = x, jj = y;
+      uint tile_offset = jj / TILE_SIZE + ii / TILE_SIZE;
+      uint byte_offset = tile_offset * TILE_SIZE * TILE_SIZE * 4;
+
+      /* Note that lp_linear_to_tiled expects the tile parameter to
+       * point at the first tile in a whole-image sized array.  In
+       * this code, we have only a single tile and have to do some
+       * pointer arithmetic to figure out where the "image" would have
+       * started.
+       */
+      lp_linear_to_tiled(linear_image, tile - byte_offset,
+                         x, y, TILE_SIZE, TILE_SIZE,
+                         lpr->base.format,
+                         lpr->row_stride[level],
+                         1);       /* tiles per row */
+   }
+}
+
+
+/**
  * Return size of resource in bytes
  */
 unsigned