#define FILE_DEBUG_FLAG DEBUG_MIPTREE
+static bool
+intel_miptree_alloc_mcs(struct brw_context *brw,
+ struct intel_mipmap_tree *mt,
+ GLuint num_samples);
+
/**
* Determine which MSAA layout should be used by the MSAA surface being
* created, based on the chip generation and the surface type.
*/
static enum intel_msaa_layout
-compute_msaa_layout(struct brw_context *brw, mesa_format format, GLenum target)
+compute_msaa_layout(struct brw_context *brw, mesa_format format, GLenum target,
+ bool disable_aux_buffers)
{
/* Prior to Gen7, all MSAA surfaces used IMS layout. */
if (brw->gen < 7)
*/
if (brw->gen == 7 && _mesa_get_format_datatype(format) == GL_INT) {
return INTEL_MSAA_LAYOUT_UMS;
+ } else if (disable_aux_buffers) {
+ /* We can't use the CMS layout because it uses an aux buffer, the MCS
+ * buffer. So fallback to UMS, which is identical to CMS without the
+ * MCS. */
+ return INTEL_MSAA_LAYOUT_UMS;
} else {
return INTEL_MSAA_LAYOUT_CMS;
}
if (brw->gen < 7)
return false;
+ if (mt->disable_aux_buffers)
+ return false;
+
/* MCS is only supported for color buffers */
switch (_mesa_get_format_base_format(mt->format)) {
case GL_DEPTH_COMPONENT:
return false;
if (mt->cpp != 4 && mt->cpp != 8 && mt->cpp != 16)
return false;
- if (mt->first_level != 0 || mt->last_level != 0)
+ if (mt->first_level != 0 || mt->last_level != 0) {
+ if (brw->gen >= 8) {
+ perf_debug("Multi-LOD fast clear - giving up (%dx%dx%d).\n",
+ mt->logical_width0, mt->logical_height0, mt->last_level);
+ }
+
return false;
- if (mt->physical_depth0 != 1)
+ }
+ if (mt->physical_depth0 != 1) {
+ if (brw->gen >= 8) {
+ perf_debug("Layered fast clear - giving up. (%dx%d%d)\n",
+ mt->logical_width0, mt->logical_height0,
+ mt->physical_depth0);
+ }
+
return false;
+ }
/* There's no point in using an MCS buffer if the surface isn't in a
* renderable format.
* intel_miptree_create_for_bo(). If true, then do not create
* \c stencil_mt.
*/
-struct intel_mipmap_tree *
+static struct intel_mipmap_tree *
intel_miptree_create_layout(struct brw_context *brw,
GLenum target,
mesa_format format,
GLuint depth0,
bool for_bo,
GLuint num_samples,
- bool force_all_slices_at_each_lod)
+ enum intel_miptree_tiling_mode requested,
+ bool force_all_slices_at_each_lod,
+ bool disable_aux_buffers)
{
struct intel_mipmap_tree *mt = calloc(sizeof(*mt), 1);
if (!mt)
return NULL;
- DBG("%s target %s format %s level %d..%d slices %d <-- %p\n", __FUNCTION__,
+ DBG("%s target %s format %s level %d..%d slices %d <-- %p\n", __func__,
_mesa_lookup_enum_by_nr(target),
_mesa_get_format_name(format),
first_level, last_level, depth0, mt);
mt->logical_height0 = height0;
mt->logical_depth0 = depth0;
mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_NO_MCS;
+ mt->disable_aux_buffers = disable_aux_buffers;
exec_list_make_empty(&mt->hiz_map);
/* The cpp is bytes per (1, blockheight)-sized block for compressed
if (num_samples > 1) {
/* Adjust width/height/depth for MSAA */
- mt->msaa_layout = compute_msaa_layout(brw, format, mt->target);
+ mt->msaa_layout = compute_msaa_layout(brw, format,
+ mt->target, mt->disable_aux_buffers);
if (mt->msaa_layout == INTEL_MSAA_LAYOUT_IMS) {
/* From the Ivybridge PRM, Volume 1, Part 1, page 108:
* "If the surface is multisampled and it is a depth or stencil
}
}
- /* Set array_layout to ALL_SLICES_AT_EACH_LOD when gen7+ array_spacing_lod0
- * can be used. array_spacing_lod0 is only used for non-IMS MSAA surfaces.
+ /* Set array_layout to ALL_SLICES_AT_EACH_LOD when array_spacing_lod0 can
+ * be used. array_spacing_lod0 is only used for non-IMS MSAA surfaces on
+ * Gen 7 and 8. On Gen 8 and 9 this layout is not available but it is still
+ * used on Gen8 to make it pick a qpitch value which doesn't include space
+ * for the mipmaps. On Gen9 this is not necessary because it will
+ * automatically pick a packed qpitch value whenever mt->first_level ==
+ * mt->last_level.
* TODO: can we use it elsewhere?
+ * TODO: also disable this on Gen8 and pick the qpitch value like Gen9
*/
- switch (mt->msaa_layout) {
- case INTEL_MSAA_LAYOUT_NONE:
- case INTEL_MSAA_LAYOUT_IMS:
+ if (brw->gen >= 9) {
mt->array_layout = ALL_LOD_IN_EACH_SLICE;
- break;
- case INTEL_MSAA_LAYOUT_UMS:
- case INTEL_MSAA_LAYOUT_CMS:
- mt->array_layout = ALL_SLICES_AT_EACH_LOD;
- break;
+ } else {
+ switch (mt->msaa_layout) {
+ case INTEL_MSAA_LAYOUT_NONE:
+ case INTEL_MSAA_LAYOUT_IMS:
+ mt->array_layout = ALL_LOD_IN_EACH_SLICE;
+ break;
+ case INTEL_MSAA_LAYOUT_UMS:
+ case INTEL_MSAA_LAYOUT_CMS:
+ mt->array_layout = ALL_SLICES_AT_EACH_LOD;
+ break;
+ }
}
if (target == GL_TEXTURE_CUBE_MAP) {
if (!for_bo &&
_mesa_get_format_base_format(format) == GL_DEPTH_STENCIL &&
(brw->must_use_separate_stencil ||
- (brw->has_separate_stencil && brw_is_hiz_depth_format(brw, format)))) {
+ (brw->has_separate_stencil &&
+ intel_miptree_wants_hiz_buffer(brw, mt)))) {
const bool force_all_slices_at_each_lod = brw->gen == 6;
mt->stencil_mt = intel_miptree_create(brw,
mt->target,
if (force_all_slices_at_each_lod)
mt->array_layout = ALL_SLICES_AT_EACH_LOD;
- brw_miptree_layout(brw, mt);
+ brw_miptree_layout(brw, for_bo, requested, mt);
- return mt;
-}
-
-/**
- * \brief Helper function for intel_miptree_create().
- */
-static uint32_t
-intel_miptree_choose_tiling(struct brw_context *brw,
- mesa_format format,
- uint32_t width0,
- uint32_t num_samples,
- enum intel_miptree_tiling_mode requested,
- struct intel_mipmap_tree *mt)
-{
- if (format == MESA_FORMAT_S_UINT8) {
- /* 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 (mt->disable_aux_buffers)
+ assert(mt->msaa_layout != INTEL_MSAA_LAYOUT_CMS);
- /* Some usages may want only one type of tiling, like depth miptrees (Y
- * tiled), or temporary BOs for uploading data once (linear).
- */
- switch (requested) {
- case INTEL_MIPTREE_TILING_ANY:
- break;
- case INTEL_MIPTREE_TILING_Y:
- return I915_TILING_Y;
- case INTEL_MIPTREE_TILING_NONE:
- return I915_TILING_NONE;
- }
-
- 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 (base_format == GL_DEPTH_COMPONENT ||
- base_format == GL_DEPTH_STENCIL_EXT)
- return I915_TILING_Y;
-
- /* 1D textures (and 1D array textures) don't get any benefit from tiling,
- * in fact it leads to a less efficient use of memory space and bandwidth
- * due to tile alignment.
- */
- if (mt->logical_height0 == 1)
- return I915_TILING_NONE;
-
- int minimum_pitch = mt->total_width * mt->cpp;
-
- /* If the width is much smaller than a tile, don't bother tiling. */
- if (minimum_pitch < 64)
- return I915_TILING_NONE;
-
- if (ALIGN(minimum_pitch, 512) >= 32768 ||
- mt->total_width >= 32768 || mt->total_height >= 32768) {
- perf_debug("%dx%d miptree too large to blit, falling back to untiled",
- mt->total_width, mt->total_height);
- return I915_TILING_NONE;
- }
-
- /* Pre-gen6 doesn't have BLORP to handle Y-tiling, so use X-tiling. */
- if (brw->gen < 6)
- return I915_TILING_X;
-
- /* From the Sandybridge PRM, Volume 1, Part 2, page 32:
- * "NOTE: 128BPE Format Color Buffer ( render target ) MUST be either TileX
- * or Linear."
- * 128 bits per pixel translates to 16 bytes per pixel. This is necessary
- * all the way back to 965, but is permitted on Gen7+.
- */
- if (brw->gen < 7 && mt->cpp >= 16)
- return I915_TILING_X;
-
- /* From the Ivy Bridge PRM, Vol4 Part1 2.12.2.1 (SURFACE_STATE for most
- * messages), on p64, under the heading "Surface Vertical Alignment":
- *
- * This field must be set to VALIGN_4 for all tiled Y Render Target
- * surfaces.
- *
- * So if the surface is renderable and uses a vertical alignment of 2,
- * force it to be X tiled. This is somewhat conservative (it's possible
- * that the client won't ever render to this surface), but it's difficult
- * to know that ahead of time. And besides, since we use a vertical
- * alignment of 4 as often as we can, this shouldn't happen very often.
- */
- if (brw->gen == 7 && mt->align_h == 2 &&
- brw->format_supported_as_render_target[format]) {
- return I915_TILING_X;
- }
-
- return I915_TILING_Y | I915_TILING_X;
+ return mt;
}
struct intel_mipmap_tree *
intel_miptree_create(struct brw_context *brw,
- GLenum target,
- mesa_format format,
- GLuint first_level,
- GLuint last_level,
- GLuint width0,
- GLuint height0,
- GLuint depth0,
- bool expect_accelerated_upload,
+ GLenum target,
+ mesa_format format,
+ GLuint first_level,
+ GLuint last_level,
+ GLuint width0,
+ GLuint height0,
+ GLuint depth0,
+ bool expect_accelerated_upload,
GLuint num_samples,
enum intel_miptree_tiling_mode requested_tiling,
bool force_all_slices_at_each_lod)
first_level, last_level, width0,
height0, depth0,
false, num_samples,
- force_all_slices_at_each_lod);
- /*
- * pitch == 0 || height == 0 indicates the null texture
- */
- if (!mt || !mt->total_width || !mt->total_height) {
- intel_miptree_release(&mt);
+ requested_tiling,
+ force_all_slices_at_each_lod,
+ false /*disable_aux_buffers*/);
+
+ if (!mt)
return NULL;
- }
total_width = mt->total_width;
total_height = mt->total_height;
total_height = ALIGN(total_height, 64);
}
- uint32_t tiling = intel_miptree_choose_tiling(brw, format, width0,
- num_samples, requested_tiling,
- mt);
bool y_or_x = false;
- if (tiling == (I915_TILING_Y | I915_TILING_X)) {
+ if (mt->tiling == (I915_TILING_Y | I915_TILING_X)) {
y_or_x = true;
mt->tiling = I915_TILING_Y;
- } else {
- mt->tiling = tiling;
}
unsigned long pitch;
uint32_t width,
uint32_t height,
uint32_t depth,
- int pitch)
+ int pitch,
+ bool disable_aux_buffers)
{
struct intel_mipmap_tree *mt;
uint32_t tiling, swizzle;
target = depth > 1 ? GL_TEXTURE_2D_ARRAY : GL_TEXTURE_2D;
+ /* 'requested' parameter of intel_miptree_create_layout() is relevant
+ * only for non bo miptree. Tiling for bo is already computed above.
+ * So, the tiling requested (INTEL_MIPTREE_TILING_ANY) below is
+ * just a place holder and will not make any change to the miptree
+ * tiling format.
+ */
mt = intel_miptree_create_layout(brw, target, format,
0, 0,
width, height, depth,
- true, 0, false);
- if (!mt) {
- free(mt);
- return mt;
- }
+ true, 0,
+ INTEL_MIPTREE_TILING_ANY,
+ false,
+ disable_aux_buffers);
+ if (!mt)
+ return NULL;
drm_intel_bo_reference(bo);
mt->bo = bo;
width,
height,
1,
- pitch);
+ pitch,
+ false);
if (!singlesample_mt)
goto fail;
if (!mt)
goto fail;
- if (brw_is_hiz_depth_format(brw, format)) {
+ if (intel_miptree_wants_hiz_buffer(brw, mt)) {
ok = intel_miptree_alloc_hiz(brw, mt);
if (!ok)
goto fail;
if (src) {
src->refcount++;
- DBG("%s %p refcount now %d\n", __FUNCTION__, src, src->refcount);
+ DBG("%s %p refcount now %d\n", __func__, src, src->refcount);
}
*dst = src;
if (!*mt)
return;
- DBG("%s %p refcount will be %d\n", __FUNCTION__, *mt, (*mt)->refcount - 1);
+ DBG("%s %p refcount will be %d\n", __func__, *mt, (*mt)->refcount - 1);
if (--(*mt)->refcount <= 0) {
GLuint i;
- DBG("%s deleting %p\n", __FUNCTION__, *mt);
+ DBG("%s deleting %p\n", __func__, *mt);
drm_intel_bo_unreference((*mt)->bo);
intel_miptree_release(&(*mt)->stencil_mt);
mt->level[level].level_x = x;
mt->level[level].level_y = y;
- DBG("%s level %d, depth %d, offset %d,%d\n", __FUNCTION__,
+ DBG("%s level %d, depth %d, offset %d,%d\n", __func__,
level, d, x, y);
assert(mt->level[level].slice == NULL);
mt->level[level].slice[img].y_offset = mt->level[level].level_y + y;
DBG("%s level %d img %d pos %d,%d\n",
- __FUNCTION__, level, img,
+ __func__, level, img,
mt->level[level].slice[img].x_offset,
mt->level[level].slice[img].y_offset);
}
intel_obj->needs_validate = true;
}
-bool
+static bool
intel_miptree_alloc_mcs(struct brw_context *brw,
struct intel_mipmap_tree *mt,
GLuint num_samples)
{
assert(brw->gen >= 7); /* MCS only used on Gen7+ */
assert(mt->mcs_mt == NULL);
+ assert(!mt->disable_aux_buffers);
/* Choose the correct format for the MCS buffer. All that really matters
* is that we allocate the right buffer size, since we'll always be
struct intel_mipmap_tree *mt)
{
assert(mt->mcs_mt == NULL);
+ assert(!mt->disable_aux_buffers);
/* The format of the MCS buffer is opaque to the driver; all that matters
* is that we get its size and pitch right. We'll pretend that the format
}
+/**
+ * Helper for intel_miptree_alloc_hiz() that determines the required hiz
+ * buffer dimensions and allocates a bo for the hiz buffer.
+ */
+static struct intel_miptree_aux_buffer *
+intel_gen8_hiz_buf_create(struct brw_context *brw,
+ struct intel_mipmap_tree *mt)
+{
+ unsigned z_width = mt->logical_width0;
+ unsigned z_height = mt->logical_height0;
+ const unsigned z_depth = MAX2(mt->logical_depth0, 1);
+ unsigned hz_width, hz_height;
+ struct intel_miptree_aux_buffer *buf = calloc(sizeof(*buf), 1);
+
+ if (!buf)
+ return NULL;
+
+ /* Gen7 PRM Volume 2, Part 1, 11.5.3 "Hierarchical Depth Buffer" documents
+ * adjustments required for Z_Height and Z_Width based on multisampling.
+ */
+ switch (mt->num_samples) {
+ case 0:
+ case 1:
+ break;
+ case 2:
+ case 4:
+ z_width *= 2;
+ z_height *= 2;
+ break;
+ case 8:
+ z_width *= 4;
+ z_height *= 2;
+ break;
+ default:
+ unreachable("unsupported sample count");
+ }
+
+ const unsigned vertical_align = 8; /* 'j' in the docs */
+ const unsigned H0 = z_height;
+ const unsigned h0 = ALIGN(H0, vertical_align);
+ const unsigned h1 = ALIGN(minify(H0, 1), vertical_align);
+ const unsigned Z0 = z_depth;
+
+ /* HZ_Width (bytes) = ceiling(Z_Width / 16) * 16 */
+ hz_width = ALIGN(z_width, 16);
+
+ unsigned H_i = H0;
+ unsigned Z_i = Z0;
+ unsigned sum_h_i = 0;
+ unsigned hz_height_3d_sum = 0;
+ for (int level = mt->first_level; level <= mt->last_level; ++level) {
+ unsigned i = level - mt->first_level;
+ unsigned h_i = ALIGN(H_i, vertical_align);
+ /* sum(i=2 to m; h_i) */
+ if (i >= 2) {
+ sum_h_i += h_i;
+ }
+ /* sum(i=0 to m; h_i * max(1, floor(Z_Depth/2**i))) */
+ hz_height_3d_sum += h_i * Z_i;
+ H_i = minify(H_i, 1);
+ Z_i = minify(Z_i, 1);
+ }
+ /* HZ_QPitch = h0 + max(h1, sum(i=2 to m; h_i)) */
+ buf->qpitch = h0 + MAX2(h1, sum_h_i);
+
+ if (mt->target == GL_TEXTURE_3D) {
+ /* (1/2) * sum(i=0 to m; h_i * max(1, floor(Z_Depth/2**i))) */
+ hz_height = DIV_ROUND_UP(hz_height_3d_sum, 2);
+ } else {
+ /* HZ_Height (rows) = ceiling( (HZ_QPitch/2)/8) *8 * Z_Depth */
+ hz_height = DIV_ROUND_UP(buf->qpitch, 2 * 8) * 8 * Z0;
+ if (mt->target == GL_TEXTURE_CUBE_MAP_ARRAY ||
+ mt->target == GL_TEXTURE_CUBE_MAP) {
+ /* HZ_Height (rows) = ceiling( (HZ_QPitch/2)/8) *8 * 6 * Z_Depth
+ *
+ * We can can just take our hz_height calculation from above, and
+ * multiply by 6 for the cube map and cube map array types.
+ */
+ hz_height *= 6;
+ }
+ }
+
+ unsigned long pitch;
+ uint32_t tiling = I915_TILING_Y;
+ buf->bo = drm_intel_bo_alloc_tiled(brw->bufmgr, "hiz",
+ hz_width, hz_height, 1,
+ &tiling, &pitch,
+ BO_ALLOC_FOR_RENDER);
+ if (!buf->bo) {
+ free(buf);
+ return NULL;
+ } else if (tiling != I915_TILING_Y) {
+ drm_intel_bo_unreference(buf->bo);
+ free(buf);
+ return NULL;
+ }
+
+ buf->pitch = pitch;
+
+ return buf;
+}
+
+
static struct intel_miptree_aux_buffer *
intel_hiz_miptree_buf_create(struct brw_context *brw,
struct intel_mipmap_tree *mt)
return buf;
}
+bool
+intel_miptree_wants_hiz_buffer(struct brw_context *brw,
+ struct intel_mipmap_tree *mt)
+{
+ if (!brw->has_hiz)
+ return false;
+
+ if (mt->hiz_buf != NULL)
+ return false;
+
+ if (mt->disable_aux_buffers)
+ return false;
+
+ switch (mt->format) {
+ case MESA_FORMAT_Z_FLOAT32:
+ case MESA_FORMAT_Z32_FLOAT_S8X24_UINT:
+ case MESA_FORMAT_Z24_UNORM_X8_UINT:
+ case MESA_FORMAT_Z24_UNORM_S8_UINT:
+ case MESA_FORMAT_Z_UNORM16:
+ return true;
+ default:
+ return false;
+ }
+}
bool
intel_miptree_alloc_hiz(struct brw_context *brw,
struct intel_mipmap_tree *mt)
{
assert(mt->hiz_buf == NULL);
+ assert(!mt->disable_aux_buffers);
if (brw->gen == 7) {
mt->hiz_buf = intel_gen7_hiz_buf_create(brw, mt);
+ } else if (brw->gen >= 8) {
+ mt->hiz_buf = intel_gen8_hiz_buf_create(brw, mt);
} else {
mt->hiz_buf = intel_hiz_miptree_buf_create(brw, mt);
}
}
DBG("%s: %d,%d %dx%d from mt %p (%s) "
- "%"PRIiPTR",%"PRIiPTR" = %p/%d\n", __FUNCTION__,
+ "%"PRIiPTR",%"PRIiPTR" = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, _mesa_get_format_name(mt->format),
x, y, map->ptr, map->stride);
map->ptr = intel_miptree_map_raw(brw, map->mt);
- DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __FUNCTION__,
+ DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, _mesa_get_format_name(mt->format),
level, slice, map->ptr, map->stride);
assert(map->mode & GL_MAP_READ_BIT);
assert(!(map->mode & GL_MAP_WRITE_BIT));
- DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __FUNCTION__,
+ DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, _mesa_get_format_name(mt->format),
level, slice, map->ptr, map->stride);
intel_miptree_unmap_raw(brw, mt);
- DBG("%s: %d,%d %dx%d from mt %p %d,%d = %p/%d\n", __FUNCTION__,
+ DBG("%s: %d,%d %dx%d from mt %p %d,%d = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, map->x + image_x, map->y + image_y, map->ptr, map->stride);
} else {
- DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __FUNCTION__,
+ DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, map->ptr, map->stride);
}
intel_miptree_unmap_raw(brw, z_mt);
DBG("%s: %d,%d %dx%d from z mt %p %d,%d, s mt %p %d,%d = %p/%d\n",
- __FUNCTION__,
+ __func__,
map->x, map->y, map->w, map->h,
z_mt, map->x + z_image_x, map->y + z_image_y,
s_mt, map->x + s_image_x, map->y + s_image_y,
map->ptr, map->stride);
} else {
- DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __FUNCTION__,
+ DBG("%s: %d,%d %dx%d from mt %p = %p/%d\n", __func__,
map->x, map->y, map->w, map->h,
mt, map->ptr, map->stride);
}
intel_miptree_unmap_raw(brw, z_mt);
DBG("%s: %d,%d %dx%d from z mt %p (%s) %d,%d, s mt %p %d,%d = %p/%d\n",
- __FUNCTION__,
+ __func__,
map->x, map->y, map->w, map->h,
z_mt, _mesa_get_format_name(z_mt->format),
map->x + z_image_x, map->y + z_image_y,
if (!map)
return;
- DBG("%s: mt %p (%s) level %d slice %d\n", __FUNCTION__,
+ DBG("%s: mt %p (%s) level %d slice %d\n", __func__,
mt, _mesa_get_format_name(mt->format), level, slice);
if (mt->format == MESA_FORMAT_S_UINT8) {
projects / mesa.git / blobdiff