mt->physical_depth0 = depth0;
if (!for_region &&
- _mesa_is_depthstencil_format(_mesa_get_format_base_format(format)) &&
+ _mesa_get_format_base_format(format) == GL_DEPTH_STENCIL &&
(intel->must_use_separate_stencil ||
(intel->has_separate_stencil &&
intel->vtbl.is_hiz_depth_format(intel, format)))) {
return mt;
}
+/**
+ * \brief Helper function for intel_miptree_create().
+ */
+static uint32_t
+intel_miptree_choose_tiling(struct intel_context *intel,
+ gl_format format,
+ uint32_t width0,
+ uint32_t num_samples,
+ bool force_y_tiling,
+ struct intel_mipmap_tree *mt)
+{
+
+ if (format == MESA_FORMAT_S8) {
+ /* The stencil buffer is W tiled. However, we request from the kernel a
+ * non-tiled buffer because the GTT is incapable of W fencing.
+ */
+ return I915_TILING_NONE;
+ }
+
+ if (!intel->use_texture_tiling || _mesa_is_format_compressed(format))
+ return I915_TILING_NONE;
+
+ if (force_y_tiling)
+ return I915_TILING_Y;
+
+ if (num_samples > 1) {
+ /* From p82 of the Sandy Bridge PRM, dw3[1] of SURFACE_STATE ("Tiled
+ * Surface"):
+ *
+ * [DevSNB+]: For multi-sample render targets, this field must be
+ * 1. MSRTs can only be tiled.
+ *
+ * Our usual reason for preferring X tiling (fast blits using the
+ * blitting engine) doesn't apply to MSAA, since we'll generally be
+ * downsampling or upsampling when blitting between the MSAA buffer
+ * and another buffer, and the blitting engine doesn't support that.
+ * So use Y tiling, since it makes better use of the cache.
+ */
+ return I915_TILING_Y;
+ }
+
+ GLenum base_format = _mesa_get_format_base_format(format);
+ if (intel->gen >= 4 &&
+ (base_format == GL_DEPTH_COMPONENT ||
+ base_format == GL_DEPTH_STENCIL_EXT))
+ return I915_TILING_Y;
+
+ if (width0 >= 64) {
+ if (ALIGN(mt->total_width * mt->cpp, 512) < 32768)
+ return I915_TILING_X;
+
+ perf_debug("%dx%d miptree too large to blit, falling back to untiled",
+ mt->total_width, mt->total_height);
+ }
+
+ return I915_TILING_NONE;
+}
struct intel_mipmap_tree *
intel_miptree_create(struct intel_context *intel,
bool force_y_tiling)
{
struct intel_mipmap_tree *mt;
- uint32_t tiling = I915_TILING_NONE;
- GLenum base_format;
gl_format tex_format = format;
gl_format etc_format = MESA_FORMAT_NONE;
GLuint total_width, total_height;
}
etc_format = (format != tex_format) ? tex_format : MESA_FORMAT_NONE;
- base_format = _mesa_get_format_base_format(format);
-
- if (num_samples > 1) {
- /* From p82 of the Sandy Bridge PRM, dw3[1] of SURFACE_STATE ("Tiled
- * Surface"):
- *
- * [DevSNB+]: For multi-sample render targets, this field must be
- * 1. MSRTs can only be tiled.
- *
- * Our usual reason for preferring X tiling (fast blits using the
- * blitting engine) doesn't apply to MSAA, since we'll generally be
- * downsampling or upsampling when blitting between the MSAA buffer
- * and another buffer, and the blitting engine doesn't support that.
- * So use Y tiling, since it makes better use of the cache.
- */
- force_y_tiling = true;
- }
-
- if (intel->use_texture_tiling && !_mesa_is_format_compressed(format)) {
- if (intel->gen >= 4 &&
- (base_format == GL_DEPTH_COMPONENT ||
- base_format == GL_DEPTH_STENCIL_EXT))
- tiling = I915_TILING_Y;
- else if (force_y_tiling) {
- tiling = I915_TILING_Y;
- } else if (width0 >= 64)
- tiling = I915_TILING_X;
- }
mt = intel_miptree_create_layout(intel, target, format,
first_level, last_level, width0,
total_height = mt->total_height;
if (format == MESA_FORMAT_S8) {
- /* The stencil buffer is W tiled. However, we request from the kernel a
- * non-tiled buffer because the GTT is incapable of W fencing. So round
- * up the width and height to match the size of W tiles (64x64).
- */
- tiling = I915_TILING_NONE;
+ /* Align to size of W tile, 64x64. */
total_width = ALIGN(total_width, 64);
total_height = ALIGN(total_height, 64);
}
+ uint32_t tiling = intel_miptree_choose_tiling(intel, format, width0,
+ num_samples, force_y_tiling,
+ mt);
mt->etc_format = etc_format;
mt->region = intel_region_alloc(intel->intelScreen,
tiling,
/**
* Copies the image's current data to the given miptree, and associates that
* miptree with the image.
+ *
+ * If \c invalidate is true, then the actual image data does not need to be
+ * copied, but the image still needs to be associated to the new miptree (this
+ * is set to true if we're about to clear the image).
*/
void
intel_miptree_copy_teximage(struct intel_context *intel,
struct intel_texture_image *intelImage,
- struct intel_mipmap_tree *dst_mt)
+ struct intel_mipmap_tree *dst_mt,
+ bool invalidate)
{
struct intel_mipmap_tree *src_mt = intelImage->mt;
struct intel_texture_object *intel_obj =
int face = intelImage->base.Base.Face;
GLuint depth = intelImage->base.Base.Depth;
- for (int slice = 0; slice < depth; slice++) {
- intel_miptree_copy_slice(intel, dst_mt, src_mt, level, face, slice);
+ if (!invalidate) {
+ for (int slice = 0; slice < depth; slice++) {
+ intel_miptree_copy_slice(intel, dst_mt, src_mt, level, face, slice);
+ }
}
intel_miptree_reference(&intelImage->mt, dst_mt);
unsigned int level,
unsigned int slice)
{
- assert(!(map->mode & GL_MAP_WRITE_BIT));
-
+ struct gl_context *ctx = &intel->ctx;
drm_intel_bo_unmap(map->bo);
+
+ if (map->mode & GL_MAP_WRITE_BIT) {
+ unsigned int image_x, image_y;
+ int x = map->x;
+ int y = map->y;
+ intel_miptree_get_image_offset(mt, level, slice, &image_x, &image_y);
+ x += image_x;
+ y += image_y;
+
+ bool ok = intelEmitCopyBlit(intel,
+ mt->region->cpp,
+ map->stride, map->bo,
+ 0, I915_TILING_NONE,
+ mt->region->pitch, mt->region->bo,
+ mt->offset, mt->region->tiling,
+ 0, 0,
+ x, y,
+ map->w, map->h,
+ GL_COPY);
+ WARN_ONCE(!ok, "Failed to blit from linear temporary mapping");
+ }
+
drm_intel_bo_unreference(map->bo);
}
{
struct intel_miptree_map *map;
+ /* Estimate the size of the mappable aperture into the GTT. There's an
+ * ioctl to get the whole GTT size, but not one to get the mappable subset.
+ * It turns out it's basically always 256MB, though some ancient hardware
+ * was smaller.
+ */
+ uint32_t gtt_size = 256 * 1024 * 1024;
+ if (intel->gen == 2)
+ gtt_size = 128 * 1024 * 1024;
+
+ /* We don't want to map two objects such that a memcpy between them would
+ * just fault one mapping in and then the other over and over forever. So
+ * we would need to divide the GTT size by 2. Additionally, some GTT is
+ * taken up by things like the framebuffer and the ringbuffer and such, so
+ * be more conservative.
+ */
+ uint32_t max_gtt_map_object_size = gtt_size / 4;
+
assert(mt->num_samples <= 1);
map = intel_miptree_attach_map(mt, level, slice, x, y, w, h, mode);
mt->region->tiling == I915_TILING_X &&
mt->region->pitch < 32768) {
intel_miptree_map_blit(intel, mt, map, level, slice);
+ } else if (mt->region->tiling != I915_TILING_NONE &&
+ mt->region->bo->size >= max_gtt_map_object_size) {
+ assert(mt->region->pitch < 32768);
+ intel_miptree_map_blit(intel, mt, map, level, slice);
} else {
intel_miptree_map_gtt(intel, mt, map, level, slice);
}