#include "util/hash_table.h"
#include "util/set.h"
#include "iris_context.h"
+#include "compiler/nir/nir.h"
-static void
-resolve_sampler_views(struct iris_batch *batch,
- struct iris_shader_state *shs)
+/**
+ * Disable auxiliary buffers if a renderbuffer is also bound as a texture
+ * or shader image. This causes a self-dependency, where both rendering
+ * and sampling may concurrently read or write the CCS buffer, causing
+ * incorrect pixels.
+ */
+static bool
+disable_rb_aux_buffer(struct iris_context *ice,
+ bool *draw_aux_buffer_disabled,
+ struct iris_resource *tex_res,
+ unsigned min_level, unsigned num_levels,
+ const char *usage)
{
- for (int i = 0; i < shs->num_textures; i++) {
- struct iris_sampler_view *isv = shs->textures[i];
- if (!isv)
+ struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer;
+ bool found = false;
+
+ /* We only need to worry about color compression and fast clears. */
+ if (tex_res->aux.usage != ISL_AUX_USAGE_CCS_D &&
+ tex_res->aux.usage != ISL_AUX_USAGE_CCS_E)
+ return false;
+
+ for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) {
+ struct iris_surface *surf = (void *) cso_fb->cbufs[i];
+ if (!surf)
continue;
+ struct iris_resource *rb_res = (void *) surf->base.texture;
+
+ if (rb_res->bo == tex_res->bo &&
+ surf->base.u.tex.level >= min_level &&
+ surf->base.u.tex.level < min_level + num_levels) {
+ found = draw_aux_buffer_disabled[i] = true;
+ }
+ }
+
+ if (found) {
+ perf_debug(&ice->dbg,
+ "Disabling CCS because a renderbuffer is also bound %s.\n",
+ usage);
+ }
+
+ return found;
+}
+
+static void
+resolve_sampler_views(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_shader_state *shs,
+ const struct shader_info *info,
+ bool *draw_aux_buffer_disabled,
+ bool consider_framebuffer)
+{
+ uint32_t views = info ? (shs->bound_sampler_views & info->textures_used) : 0;
+
+ while (views) {
+ const int i = u_bit_scan(&views);
+ struct iris_sampler_view *isv = shs->textures[i];
struct iris_resource *res = (void *) isv->base.texture;
- // XXX: aux tracking
+ if (res->base.target != PIPE_BUFFER) {
+ if (consider_framebuffer) {
+ disable_rb_aux_buffer(ice, draw_aux_buffer_disabled,
+ res, isv->view.base_level, isv->view.levels,
+ "for sampling");
+ }
+
+ iris_resource_prepare_texture(ice, batch, res, isv->view.format,
+ isv->view.base_level, isv->view.levels,
+ isv->view.base_array_layer,
+ isv->view.array_len);
+ }
+
+ iris_cache_flush_for_read(batch, res->bo);
+ }
+}
+
+static void
+resolve_image_views(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_shader_state *shs,
+ const struct shader_info *info,
+ bool *draw_aux_buffer_disabled,
+ bool consider_framebuffer)
+{
+ uint32_t views = info ? (shs->bound_image_views & info->images_used) : 0;
+
+ while (views) {
+ const int i = u_bit_scan(&views);
+ struct pipe_image_view *pview = &shs->image[i].base;
+ struct iris_resource *res = (void *) pview->resource;
+
+ if (res->base.target != PIPE_BUFFER) {
+ if (consider_framebuffer) {
+ disable_rb_aux_buffer(ice, draw_aux_buffer_disabled,
+ res, pview->u.tex.level, 1,
+ "as a shader image");
+ }
+
+ unsigned num_layers =
+ pview->u.tex.last_layer - pview->u.tex.first_layer + 1;
+
+ /* The data port doesn't understand any compression */
+ iris_resource_prepare_access(ice, batch, res,
+ pview->u.tex.level, 1,
+ pview->u.tex.first_layer, num_layers,
+ ISL_AUX_USAGE_NONE, false);
+ }
+
iris_cache_flush_for_read(batch, res->bo);
}
}
+
/**
* \brief Resolve buffers before drawing.
*
*/
void
iris_predraw_resolve_inputs(struct iris_context *ice,
- struct iris_batch *batch)
+ struct iris_batch *batch,
+ bool *draw_aux_buffer_disabled,
+ gl_shader_stage stage,
+ bool consider_framebuffer)
{
- for (gl_shader_stage stage = 0; stage < MESA_SHADER_STAGES; stage++) {
- struct iris_shader_state *shs = &ice->state.shaders[stage];
- resolve_sampler_views(batch, shs);
- }
+ struct iris_shader_state *shs = &ice->state.shaders[stage];
+ const struct shader_info *info = iris_get_shader_info(ice, stage);
- // XXX: storage images
+ uint64_t dirty = (IRIS_DIRTY_BINDINGS_VS << stage) |
+ (consider_framebuffer ? IRIS_DIRTY_BINDINGS_FS : 0);
+
+ if (ice->state.dirty & dirty) {
+ resolve_sampler_views(ice, batch, shs, info, draw_aux_buffer_disabled,
+ consider_framebuffer);
+ resolve_image_views(ice, batch, shs, info, draw_aux_buffer_disabled,
+ consider_framebuffer);
+ }
}
void
iris_predraw_resolve_framebuffer(struct iris_context *ice,
- struct iris_batch *batch)
+ struct iris_batch *batch,
+ bool *draw_aux_buffer_disabled)
{
struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer;
- struct pipe_surface *zs_surf = cso_fb->zsbuf;
+ struct iris_screen *screen = (void *) ice->ctx.screen;
+ struct gen_device_info *devinfo = &screen->devinfo;
+ struct iris_uncompiled_shader *ish =
+ ice->shaders.uncompiled[MESA_SHADER_FRAGMENT];
+ const nir_shader *nir = ish->nir;
- if (zs_surf) {
- // XXX: HiZ resolves
+ if (ice->state.dirty & IRIS_DIRTY_DEPTH_BUFFER) {
+ struct pipe_surface *zs_surf = cso_fb->zsbuf;
+
+ if (zs_surf) {
+ struct iris_resource *z_res, *s_res;
+ iris_get_depth_stencil_resources(zs_surf->texture, &z_res, &s_res);
+ unsigned num_layers =
+ zs_surf->u.tex.last_layer - zs_surf->u.tex.first_layer + 1;
+
+ if (z_res) {
+ iris_resource_prepare_depth(ice, batch, z_res,
+ zs_surf->u.tex.level,
+ zs_surf->u.tex.first_layer,
+ num_layers);
+ iris_cache_flush_for_depth(batch, z_res->bo);
+ }
+
+ if (s_res) {
+ iris_cache_flush_for_depth(batch, s_res->bo);
+ }
+ }
}
- for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) {
- struct iris_surface *surf = (void *) cso_fb->cbufs[i];
- if (!surf)
- continue;
+ if (devinfo->gen == 8 && nir->info.outputs_read != 0) {
+ for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) {
+ if (cso_fb->cbufs[i]) {
+ struct iris_surface *surf = (void *) cso_fb->cbufs[i];
+ struct iris_resource *res = (void *) cso_fb->cbufs[i]->texture;
- struct iris_resource *res = (void *) surf->base.texture;
+ iris_resource_prepare_texture(ice, batch, res, surf->view.format,
+ surf->view.base_level, 1,
+ surf->view.base_array_layer,
+ surf->view.array_len);
+ }
+ }
+ }
+
+ if (ice->state.dirty & (IRIS_DIRTY_BINDINGS_FS | IRIS_DIRTY_BLEND_STATE)) {
+ for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) {
+ struct iris_surface *surf = (void *) cso_fb->cbufs[i];
+ if (!surf)
+ continue;
+
+ struct iris_resource *res = (void *) surf->base.texture;
- // XXX: aux tracking
+ enum isl_aux_usage aux_usage =
+ iris_resource_render_aux_usage(ice, res, surf->view.format,
+ ice->state.blend_enables & (1u << i),
+ draw_aux_buffer_disabled[i]);
- iris_cache_flush_for_render(batch, res->bo, surf->view.format,
- ISL_AUX_USAGE_NONE);
+ if (ice->state.draw_aux_usage[i] != aux_usage) {
+ ice->state.draw_aux_usage[i] = aux_usage;
+ /* XXX: Need to track which bindings to make dirty */
+ ice->state.dirty |= IRIS_ALL_DIRTY_BINDINGS;
+ }
+
+ iris_resource_prepare_render(ice, batch, res, surf->view.base_level,
+ surf->view.base_array_layer,
+ surf->view.array_len,
+ aux_usage);
+
+ iris_cache_flush_for_render(batch, res->bo, surf->view.format,
+ aux_usage);
+ }
}
}
struct iris_batch *batch)
{
struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer;
- struct pipe_surface *zs_surf = cso_fb->zsbuf;
// XXX: front buffer drawing?
+ bool may_have_resolved_depth =
+ ice->state.dirty & (IRIS_DIRTY_DEPTH_BUFFER |
+ IRIS_DIRTY_WM_DEPTH_STENCIL);
+
+ struct pipe_surface *zs_surf = cso_fb->zsbuf;
if (zs_surf) {
struct iris_resource *z_res, *s_res;
iris_get_depth_stencil_resources(zs_surf->texture, &z_res, &s_res);
+ unsigned num_layers =
+ zs_surf->u.tex.last_layer - zs_surf->u.tex.first_layer + 1;
if (z_res) {
- // XXX: aux tracking
+ if (may_have_resolved_depth) {
+ iris_resource_finish_depth(ice, z_res, zs_surf->u.tex.level,
+ zs_surf->u.tex.first_layer, num_layers,
+ ice->state.depth_writes_enabled);
+ }
if (ice->state.depth_writes_enabled)
iris_depth_cache_add_bo(batch, z_res->bo);
}
if (s_res) {
- // XXX: aux tracking
+ if (may_have_resolved_depth && ice->state.stencil_writes_enabled) {
+ iris_resource_finish_write(ice, s_res, zs_surf->u.tex.level,
+ zs_surf->u.tex.first_layer, num_layers,
+ s_res->aux.usage);
+ }
if (ice->state.stencil_writes_enabled)
- iris_depth_cache_add_bo(batch, z_res->bo);
+ iris_depth_cache_add_bo(batch, s_res->bo);
}
}
+ bool may_have_resolved_color =
+ ice->state.dirty & (IRIS_DIRTY_BINDINGS_FS | IRIS_DIRTY_BLEND_STATE);
+
for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) {
struct iris_surface *surf = (void *) cso_fb->cbufs[i];
if (!surf)
continue;
struct iris_resource *res = (void *) surf->base.texture;
+ enum isl_aux_usage aux_usage = ice->state.draw_aux_usage[i];
- // XXX: aux tracking
iris_render_cache_add_bo(batch, res->bo, surf->view.format,
- ISL_AUX_USAGE_NONE);
+ aux_usage);
+
+ if (may_have_resolved_color) {
+ union pipe_surface_desc *desc = &surf->base.u;
+ unsigned num_layers =
+ desc->tex.last_layer - desc->tex.first_layer + 1;
+ iris_resource_finish_render(ice, res, desc->tex.level,
+ desc->tex.first_layer, num_layers,
+ aux_usage);
+ }
}
}
void
iris_cache_sets_clear(struct iris_batch *batch)
{
- struct hash_entry *render_entry;
hash_table_foreach(batch->cache.render, render_entry)
_mesa_hash_table_remove(batch->cache.render, render_entry);
- struct set_entry *depth_entry;
set_foreach(batch->cache.depth, depth_entry)
_mesa_set_remove(batch->cache.depth, depth_entry);
}
iris_flush_depth_and_render_caches(struct iris_batch *batch)
{
iris_emit_pipe_control_flush(batch,
+ "cache tracker: render-to-texture",
PIPE_CONTROL_DEPTH_CACHE_FLUSH |
PIPE_CONTROL_RENDER_TARGET_FLUSH |
PIPE_CONTROL_CS_STALL);
iris_emit_pipe_control_flush(batch,
+ "cache tracker: render-to-texture",
PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
PIPE_CONTROL_CONST_CACHE_INVALIDATE);
iris_cache_flush_for_read(struct iris_batch *batch,
struct iris_bo *bo)
{
- if (_mesa_hash_table_search(batch->cache.render, bo) ||
- _mesa_set_search(batch->cache.depth, bo))
+ if (_mesa_hash_table_search_pre_hashed(batch->cache.render, bo->hash, bo) ||
+ _mesa_set_search_pre_hashed(batch->cache.depth, bo->hash, bo))
iris_flush_depth_and_render_caches(batch);
}
enum isl_format format,
enum isl_aux_usage aux_usage)
{
- if (_mesa_set_search(batch->cache.depth, bo))
+ if (_mesa_set_search_pre_hashed(batch->cache.depth, bo->hash, bo))
iris_flush_depth_and_render_caches(batch);
/* Check to see if this bo has been used by a previous rendering operation
* and flush on format changes too. We can always relax this later if we
* find it to be a performance problem.
*/
- struct hash_entry *entry = _mesa_hash_table_search(batch->cache.render, bo);
+ struct hash_entry *entry =
+ _mesa_hash_table_search_pre_hashed(batch->cache.render, bo->hash, bo);
if (entry && entry->data != format_aux_tuple(format, aux_usage))
iris_flush_depth_and_render_caches(batch);
}
enum isl_aux_usage aux_usage)
{
#ifndef NDEBUG
- struct hash_entry *entry = _mesa_hash_table_search(batch->cache.render, bo);
+ struct hash_entry *entry =
+ _mesa_hash_table_search_pre_hashed(batch->cache.render, bo->hash, bo);
if (entry) {
/* Otherwise, someone didn't do a flush_for_render and that would be
* very bad indeed.
}
#endif
- _mesa_hash_table_insert(batch->cache.render, bo,
- format_aux_tuple(format, aux_usage));
+ _mesa_hash_table_insert_pre_hashed(batch->cache.render, bo->hash, bo,
+ format_aux_tuple(format, aux_usage));
}
void
iris_cache_flush_for_depth(struct iris_batch *batch,
struct iris_bo *bo)
{
- if (_mesa_hash_table_search(batch->cache.render, bo))
+ if (_mesa_hash_table_search_pre_hashed(batch->cache.render, bo->hash, bo))
iris_flush_depth_and_render_caches(batch);
}
void
iris_depth_cache_add_bo(struct iris_batch *batch, struct iris_bo *bo)
{
- _mesa_set_add(batch->cache.depth, bo);
+ _mesa_set_add_pre_hashed(batch->cache.depth, bo->hash, bo);
+}
+
+static void
+iris_resolve_color(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res,
+ unsigned level, unsigned layer,
+ enum isl_aux_op resolve_op)
+{
+ //DBG("%s to mt %p level %u layer %u\n", __FUNCTION__, mt, level, layer);
+
+ struct blorp_surf surf;
+ iris_blorp_surf_for_resource(&ice->vtbl, &batch->screen->isl_dev, &surf,
+ &res->base, res->aux.usage, level, true);
+
+ iris_batch_maybe_flush(batch, 1500);
+
+ /* Ivybridge PRM Vol 2, Part 1, "11.7 MCS Buffer for Render Target(s)":
+ *
+ * "Any transition from any value in {Clear, Render, Resolve} to a
+ * different value in {Clear, Render, Resolve} requires end of pipe
+ * synchronization."
+ *
+ * In other words, fast clear ops are not properly synchronized with
+ * other drawing. We need to use a PIPE_CONTROL to ensure that the
+ * contents of the previous draw hit the render target before we resolve
+ * and again afterwards to ensure that the resolve is complete before we
+ * do any more regular drawing.
+ */
+ iris_emit_end_of_pipe_sync(batch, "color resolve: pre-flush",
+ PIPE_CONTROL_RENDER_TARGET_FLUSH);
+
+ struct blorp_batch blorp_batch;
+ blorp_batch_init(&ice->blorp, &blorp_batch, batch, 0);
+ /* On Gen >= 12, Stencil buffer with lossless compression needs to be
+ * resolve with WM_HZ_OP packet.
+ */
+ if (res->aux.usage == ISL_AUX_USAGE_STC_CCS) {
+ blorp_hiz_stencil_op(&blorp_batch, &surf, level, layer,
+ 1, resolve_op);
+ } else {
+ blorp_ccs_resolve(&blorp_batch, &surf, level, layer, 1,
+ isl_format_srgb_to_linear(res->surf.format),
+ resolve_op);
+ }
+ blorp_batch_finish(&blorp_batch);
+
+ /* See comment above */
+ iris_emit_end_of_pipe_sync(batch, "color resolve: post-flush",
+ PIPE_CONTROL_RENDER_TARGET_FLUSH);
+}
+
+static void
+iris_mcs_partial_resolve(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res,
+ uint32_t start_layer,
+ uint32_t num_layers)
+{
+ //DBG("%s to mt %p layers %u-%u\n", __FUNCTION__, mt,
+ //start_layer, start_layer + num_layers - 1);
+
+ assert(isl_aux_usage_has_mcs(res->aux.usage));
+
+ struct blorp_surf surf;
+ iris_blorp_surf_for_resource(&ice->vtbl, &batch->screen->isl_dev, &surf,
+ &res->base, res->aux.usage, 0, true);
+
+ struct blorp_batch blorp_batch;
+ blorp_batch_init(&ice->blorp, &blorp_batch, batch, 0);
+ blorp_mcs_partial_resolve(&blorp_batch, &surf,
+ isl_format_srgb_to_linear(res->surf.format),
+ start_layer, num_layers);
+ blorp_batch_finish(&blorp_batch);
+}
+
+
+/**
+ * Return true if the format that will be used to access the resource is
+ * CCS_E-compatible with the resource's linear/non-sRGB format.
+ *
+ * Why use the linear format? Well, although the resourcemay be specified
+ * with an sRGB format, the usage of that color space/format can be toggled.
+ * Since our HW tends to support more linear formats than sRGB ones, we use
+ * this format variant for check for CCS_E compatibility.
+ */
+static bool
+format_ccs_e_compat_with_resource(const struct gen_device_info *devinfo,
+ const struct iris_resource *res,
+ enum isl_format access_format)
+{
+ assert(res->aux.usage == ISL_AUX_USAGE_CCS_E);
+
+ enum isl_format isl_format = isl_format_srgb_to_linear(res->surf.format);
+ return isl_formats_are_ccs_e_compatible(devinfo, isl_format, access_format);
+}
+
+bool
+iris_sample_with_depth_aux(const struct gen_device_info *devinfo,
+ const struct iris_resource *res)
+{
+ switch (res->aux.usage) {
+ case ISL_AUX_USAGE_HIZ:
+ if (devinfo->has_sample_with_hiz)
+ break;
+ return false;
+ case ISL_AUX_USAGE_HIZ_CCS:
+ return false;
+ case ISL_AUX_USAGE_HIZ_CCS_WT:
+ break;
+ default:
+ return false;
+ }
+
+ /* It seems the hardware won't fallback to the depth buffer if some of the
+ * mipmap levels aren't available in the HiZ buffer. So we need all levels
+ * of the texture to be HiZ enabled.
+ */
+ for (unsigned level = 0; level < res->surf.levels; ++level) {
+ if (!iris_resource_level_has_hiz(res, level))
+ return false;
+ }
+
+ /* If compressed multisampling is enabled, then we use it for the auxiliary
+ * buffer instead.
+ *
+ * From the BDW PRM (Volume 2d: Command Reference: Structures
+ * RENDER_SURFACE_STATE.AuxiliarySurfaceMode):
+ *
+ * "If this field is set to AUX_HIZ, Number of Multisamples must be
+ * MULTISAMPLECOUNT_1, and Surface Type cannot be SURFTYPE_3D.
+ *
+ * There is no such blurb for 1D textures, but there is sufficient evidence
+ * that this is broken on SKL+.
+ */
+ // XXX: i965 disables this for arrays too, is that reasonable?
+ return res->surf.samples == 1 && res->surf.dim == ISL_SURF_DIM_2D;
+}
+
+/**
+ * Perform a HiZ or depth resolve operation.
+ *
+ * For an overview of HiZ ops, see the following sections of the Sandy Bridge
+ * PRM, Volume 1, Part 2:
+ * - 7.5.3.1 Depth Buffer Clear
+ * - 7.5.3.2 Depth Buffer Resolve
+ * - 7.5.3.3 Hierarchical Depth Buffer Resolve
+ */
+void
+iris_hiz_exec(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res,
+ unsigned int level, unsigned int start_layer,
+ unsigned int num_layers, enum isl_aux_op op,
+ bool update_clear_depth)
+{
+ assert(iris_resource_level_has_hiz(res, level));
+ assert(op != ISL_AUX_OP_NONE);
+ UNUSED const char *name = NULL;
+
+ switch (op) {
+ case ISL_AUX_OP_FULL_RESOLVE:
+ name = "depth resolve";
+ break;
+ case ISL_AUX_OP_AMBIGUATE:
+ name = "hiz ambiguate";
+ break;
+ case ISL_AUX_OP_FAST_CLEAR:
+ name = "depth clear";
+ break;
+ case ISL_AUX_OP_PARTIAL_RESOLVE:
+ case ISL_AUX_OP_NONE:
+ unreachable("Invalid HiZ op");
+ }
+
+ //DBG("%s %s to mt %p level %d layers %d-%d\n",
+ //__func__, name, mt, level, start_layer, start_layer + num_layers - 1);
+
+ /* The following stalls and flushes are only documented to be required
+ * for HiZ clear operations. However, they also seem to be required for
+ * resolve operations.
+ *
+ * From the Ivybridge PRM, volume 2, "Depth Buffer Clear":
+ *
+ * "If other rendering operations have preceded this clear, a
+ * PIPE_CONTROL with depth cache flush enabled, Depth Stall bit
+ * enabled must be issued before the rectangle primitive used for
+ * the depth buffer clear operation."
+ *
+ * Same applies for Gen8 and Gen9.
+ *
+ * In addition, from the Ivybridge PRM, volume 2, 1.10.4.1
+ * PIPE_CONTROL, Depth Cache Flush Enable:
+ *
+ * "This bit must not be set when Depth Stall Enable bit is set in
+ * this packet."
+ *
+ * This is confirmed to hold for real, Haswell gets immediate gpu hangs.
+ *
+ * Therefore issue two pipe control flushes, one for cache flush and
+ * another for depth stall.
+ */
+ iris_emit_pipe_control_flush(batch,
+ "hiz op: pre-flushes (1/2)",
+ PIPE_CONTROL_DEPTH_CACHE_FLUSH |
+ PIPE_CONTROL_CS_STALL);
+
+ iris_emit_pipe_control_flush(batch, "hiz op: pre-flushes (2/2)",
+ PIPE_CONTROL_DEPTH_STALL);
+
+ assert(isl_aux_usage_has_hiz(res->aux.usage) && res->aux.bo);
+
+ iris_batch_maybe_flush(batch, 1500);
+
+ struct blorp_surf surf;
+ iris_blorp_surf_for_resource(&ice->vtbl, &batch->screen->isl_dev, &surf,
+ &res->base, res->aux.usage, level, true);
+
+ struct blorp_batch blorp_batch;
+ enum blorp_batch_flags flags = 0;
+ flags |= update_clear_depth ? 0 : BLORP_BATCH_NO_UPDATE_CLEAR_COLOR;
+ blorp_batch_init(&ice->blorp, &blorp_batch, batch, flags);
+ blorp_hiz_op(&blorp_batch, &surf, level, start_layer, num_layers, op);
+ blorp_batch_finish(&blorp_batch);
+
+ /* The following stalls and flushes are only documented to be required
+ * for HiZ clear operations. However, they also seem to be required for
+ * resolve operations.
+ *
+ * From the Broadwell PRM, volume 7, "Depth Buffer Clear":
+ *
+ * "Depth buffer clear pass using any of the methods (WM_STATE,
+ * 3DSTATE_WM or 3DSTATE_WM_HZ_OP) must be followed by a
+ * PIPE_CONTROL command with DEPTH_STALL bit and Depth FLUSH bits
+ * "set" before starting to render. DepthStall and DepthFlush are
+ * not needed between consecutive depth clear passes nor is it
+ * required if the depth clear pass was done with
+ * 'full_surf_clear' bit set in the 3DSTATE_WM_HZ_OP."
+ *
+ * TODO: Such as the spec says, this could be conditional.
+ */
+ iris_emit_pipe_control_flush(batch,
+ "hiz op: post flush",
+ PIPE_CONTROL_DEPTH_CACHE_FLUSH |
+ PIPE_CONTROL_DEPTH_STALL);
+}
+
+static bool
+level_has_aux(const struct iris_resource *res, uint32_t level)
+{
+ return isl_aux_usage_has_hiz(res->aux.usage) ?
+ iris_resource_level_has_hiz(res, level) :
+ res->aux.usage != ISL_AUX_USAGE_NONE;
+}
+
+/**
+ * Does the resource's slice have hiz enabled?
+ */
+bool
+iris_resource_level_has_hiz(const struct iris_resource *res, uint32_t level)
+{
+ iris_resource_check_level_layer(res, level, 0);
+ return res->aux.has_hiz & 1 << level;
+}
+
+/** \brief Assert that the level and layer are valid for the resource. */
+void
+iris_resource_check_level_layer(UNUSED const struct iris_resource *res,
+ UNUSED uint32_t level, UNUSED uint32_t layer)
+{
+ assert(level < res->surf.levels);
+ assert(layer < util_num_layers(&res->base, level));
+}
+
+static inline uint32_t
+miptree_level_range_length(const struct iris_resource *res,
+ uint32_t start_level, uint32_t num_levels)
+{
+ assert(start_level < res->surf.levels);
+
+ if (num_levels == INTEL_REMAINING_LAYERS)
+ num_levels = res->surf.levels;
+
+ /* Check for overflow */
+ assert(start_level + num_levels >= start_level);
+ assert(start_level + num_levels <= res->surf.levels);
+
+ return num_levels;
+}
+
+static inline uint32_t
+miptree_layer_range_length(const struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t num_layers)
+{
+ assert(level <= res->base.last_level);
+
+ const uint32_t total_num_layers = iris_get_num_logical_layers(res, level);
+ assert(start_layer < total_num_layers);
+ if (num_layers == INTEL_REMAINING_LAYERS)
+ num_layers = total_num_layers - start_layer;
+ /* Check for overflow */
+ assert(start_layer + num_layers >= start_layer);
+ assert(start_layer + num_layers <= total_num_layers);
+
+ return num_layers;
+}
+
+bool
+iris_has_color_unresolved(const struct iris_resource *res,
+ unsigned start_level, unsigned num_levels,
+ unsigned start_layer, unsigned num_layers)
+{
+ if (!res->aux.bo)
+ return false;
+
+ /* Clamp the level range to fit the resource */
+ num_levels = miptree_level_range_length(res, start_level, num_levels);
+
+ for (uint32_t l = 0; l < num_levels; l++) {
+ const uint32_t level = start_level + l;
+ const uint32_t level_layers =
+ miptree_layer_range_length(res, level, start_layer, num_layers);
+ for (unsigned a = 0; a < level_layers; a++) {
+ enum isl_aux_state aux_state =
+ iris_resource_get_aux_state(res, level, start_layer + a);
+ assert(aux_state != ISL_AUX_STATE_AUX_INVALID);
+ if (aux_state != ISL_AUX_STATE_PASS_THROUGH)
+ return true;
+ }
+ }
+
+ return false;
+}
+
+void
+iris_resource_prepare_access(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res,
+ uint32_t start_level, uint32_t num_levels,
+ uint32_t start_layer, uint32_t num_layers,
+ enum isl_aux_usage aux_usage,
+ bool fast_clear_supported)
+{
+ const uint32_t clamped_levels =
+ miptree_level_range_length(res, start_level, num_levels);
+ for (uint32_t l = 0; l < clamped_levels; l++) {
+ const uint32_t level = start_level + l;
+ if (!level_has_aux(res, level))
+ continue;
+
+ const uint32_t level_layers =
+ miptree_layer_range_length(res, level, start_layer, num_layers);
+ for (uint32_t a = 0; a < level_layers; a++) {
+ const uint32_t layer = start_layer + a;
+ const enum isl_aux_state aux_state =
+ iris_resource_get_aux_state(res, level, layer);
+ const enum isl_aux_op aux_op =
+ isl_aux_prepare_access(aux_state, aux_usage, fast_clear_supported);
+
+ if (aux_op == ISL_AUX_OP_NONE) {
+ /* Nothing to do here. */
+ } else if (isl_aux_usage_has_mcs(res->aux.usage)) {
+ assert(aux_op == ISL_AUX_OP_PARTIAL_RESOLVE);
+ iris_mcs_partial_resolve(ice, batch, res, layer, 1);
+ } else if (isl_aux_usage_has_hiz(res->aux.usage)) {
+ iris_hiz_exec(ice, batch, res, level, layer, 1, aux_op, false);
+ } else {
+ assert(isl_aux_usage_has_ccs(res->aux.usage));
+ iris_resolve_color(ice, batch, res, level, layer, aux_op);
+ }
+
+ const enum isl_aux_state new_state =
+ isl_aux_state_transition_aux_op(aux_state, res->aux.usage, aux_op);
+ iris_resource_set_aux_state(ice, res, level, layer, 1, new_state);
+ }
+ }
+}
+
+void
+iris_resource_finish_write(struct iris_context *ice,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t num_layers,
+ enum isl_aux_usage aux_usage)
+{
+ if (!level_has_aux(res, level))
+ return;
+
+ const uint32_t level_layers =
+ miptree_layer_range_length(res, level, start_layer, num_layers);
+
+ for (uint32_t a = 0; a < level_layers; a++) {
+ const uint32_t layer = start_layer + a;
+ const enum isl_aux_state aux_state =
+ iris_resource_get_aux_state(res, level, layer);
+ const enum isl_aux_state new_aux_state =
+ isl_aux_state_transition_write(aux_state, aux_usage, false);
+ iris_resource_set_aux_state(ice, res, level, layer, 1, new_aux_state);
+ }
+}
+
+enum isl_aux_state
+iris_resource_get_aux_state(const struct iris_resource *res,
+ uint32_t level, uint32_t layer)
+{
+ iris_resource_check_level_layer(res, level, layer);
+
+ if (res->surf.usage & ISL_SURF_USAGE_DEPTH_BIT) {
+ assert(iris_resource_level_has_hiz(res, level));
+ } else {
+ assert(res->surf.samples == 1 ||
+ res->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY);
+ }
+
+ return res->aux.state[level][layer];
+}
+
+void
+iris_resource_set_aux_state(struct iris_context *ice,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t num_layers,
+ enum isl_aux_state aux_state)
+{
+ num_layers = miptree_layer_range_length(res, level, start_layer, num_layers);
+
+ if (res->surf.usage & ISL_SURF_USAGE_DEPTH_BIT) {
+ assert(iris_resource_level_has_hiz(res, level));
+ } else {
+ assert(res->surf.samples == 1 ||
+ res->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY);
+ }
+
+ for (unsigned a = 0; a < num_layers; a++) {
+ if (res->aux.state[level][start_layer + a] != aux_state) {
+ res->aux.state[level][start_layer + a] = aux_state;
+ /* XXX: Need to track which bindings to make dirty */
+ ice->state.dirty |= IRIS_ALL_DIRTY_BINDINGS;
+ }
+ }
+}
+
+/* On Gen9 color buffers may be compressed by the hardware (lossless
+ * compression). There are, however, format restrictions and care needs to be
+ * taken that the sampler engine is capable for re-interpreting a buffer with
+ * format different the buffer was originally written with.
+ *
+ * For example, SRGB formats are not compressible and the sampler engine isn't
+ * capable of treating RGBA_UNORM as SRGB_ALPHA. In such a case the underlying
+ * color buffer needs to be resolved so that the sampling surface can be
+ * sampled as non-compressed (i.e., without the auxiliary MCS buffer being
+ * set).
+ */
+static bool
+can_texture_with_ccs(const struct gen_device_info *devinfo,
+ struct pipe_debug_callback *dbg,
+ const struct iris_resource *res,
+ enum isl_format view_format)
+{
+ if (res->aux.usage != ISL_AUX_USAGE_CCS_E)
+ return false;
+
+ if (!format_ccs_e_compat_with_resource(devinfo, res, view_format)) {
+ const struct isl_format_layout *res_fmtl =
+ isl_format_get_layout(res->surf.format);
+ const struct isl_format_layout *view_fmtl =
+ isl_format_get_layout(view_format);
+
+ perf_debug(dbg, "Incompatible sampling format (%s) for CCS (%s)\n",
+ view_fmtl->name, res_fmtl->name);
+
+ return false;
+ }
+
+ return true;
+}
+
+enum isl_aux_usage
+iris_resource_texture_aux_usage(struct iris_context *ice,
+ const struct iris_resource *res,
+ enum isl_format view_format)
+{
+ struct iris_screen *screen = (void *) ice->ctx.screen;
+ struct gen_device_info *devinfo = &screen->devinfo;
+
+ switch (res->aux.usage) {
+ case ISL_AUX_USAGE_HIZ:
+ if (iris_sample_with_depth_aux(devinfo, res))
+ return ISL_AUX_USAGE_HIZ;
+ break;
+
+ case ISL_AUX_USAGE_HIZ_CCS:
+ assert(!iris_sample_with_depth_aux(devinfo, res));
+ return ISL_AUX_USAGE_NONE;
+
+ case ISL_AUX_USAGE_HIZ_CCS_WT:
+ if (iris_sample_with_depth_aux(devinfo, res))
+ return ISL_AUX_USAGE_HIZ_CCS_WT;
+ break;
+
+ case ISL_AUX_USAGE_MCS:
+ case ISL_AUX_USAGE_MCS_CCS:
+ case ISL_AUX_USAGE_STC_CCS:
+ return res->aux.usage;
+
+ case ISL_AUX_USAGE_CCS_D:
+ case ISL_AUX_USAGE_CCS_E:
+ /* If we don't have any unresolved color, report an aux usage of
+ * ISL_AUX_USAGE_NONE. This way, texturing won't even look at the
+ * aux surface and we can save some bandwidth.
+ */
+ if (!iris_has_color_unresolved(res, 0, INTEL_REMAINING_LEVELS,
+ 0, INTEL_REMAINING_LAYERS))
+ return ISL_AUX_USAGE_NONE;
+
+ if (can_texture_with_ccs(devinfo, &ice->dbg, res, view_format))
+ return ISL_AUX_USAGE_CCS_E;
+ break;
+
+ default:
+ break;
+ }
+
+ return ISL_AUX_USAGE_NONE;
+}
+
+static bool
+isl_formats_are_fast_clear_compatible(enum isl_format a, enum isl_format b)
+{
+ /* On gen8 and earlier, the hardware was only capable of handling 0/1 clear
+ * values so sRGB curve application was a no-op for all fast-clearable
+ * formats.
+ *
+ * On gen9+, the hardware supports arbitrary clear values. For sRGB clear
+ * values, the hardware interprets the floats, not as what would be
+ * returned from the sampler (or written by the shader), but as being
+ * between format conversion and sRGB curve application. This means that
+ * we can switch between sRGB and UNORM without having to whack the clear
+ * color.
+ */
+ return isl_format_srgb_to_linear(a) == isl_format_srgb_to_linear(b);
+}
+
+void
+iris_resource_prepare_texture(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res,
+ enum isl_format view_format,
+ uint32_t start_level, uint32_t num_levels,
+ uint32_t start_layer, uint32_t num_layers)
+{
+ enum isl_aux_usage aux_usage =
+ iris_resource_texture_aux_usage(ice, res, view_format);
+
+ bool clear_supported = isl_aux_usage_has_fast_clears(aux_usage);
+
+ /* Clear color is specified as ints or floats and the conversion is done by
+ * the sampler. If we have a texture view, we would have to perform the
+ * clear color conversion manually. Just disable clear color.
+ */
+ if (!isl_formats_are_fast_clear_compatible(res->surf.format, view_format))
+ clear_supported = false;
+
+ iris_resource_prepare_access(ice, batch, res, start_level, num_levels,
+ start_layer, num_layers,
+ aux_usage, clear_supported);
+}
+
+enum isl_aux_usage
+iris_resource_render_aux_usage(struct iris_context *ice,
+ struct iris_resource *res,
+ enum isl_format render_format,
+ bool blend_enabled,
+ bool draw_aux_disabled)
+{
+ struct iris_screen *screen = (void *) ice->ctx.screen;
+ struct gen_device_info *devinfo = &screen->devinfo;
+
+ if (draw_aux_disabled)
+ return ISL_AUX_USAGE_NONE;
+
+ switch (res->aux.usage) {
+ case ISL_AUX_USAGE_MCS:
+ case ISL_AUX_USAGE_MCS_CCS:
+ return res->aux.usage;
+
+ case ISL_AUX_USAGE_CCS_D:
+ case ISL_AUX_USAGE_CCS_E:
+ /* Gen9+ hardware technically supports non-0/1 clear colors with sRGB
+ * formats. However, there are issues with blending where it doesn't
+ * properly apply the sRGB curve to the clear color when blending.
+ */
+ if (devinfo->gen >= 9 && blend_enabled &&
+ isl_format_is_srgb(render_format) &&
+ !isl_color_value_is_zero_one(res->aux.clear_color, render_format))
+ return ISL_AUX_USAGE_NONE;
+
+ if (res->aux.usage == ISL_AUX_USAGE_CCS_E &&
+ format_ccs_e_compat_with_resource(devinfo, res, render_format))
+ return ISL_AUX_USAGE_CCS_E;
+
+ /* Otherwise, we try to fall back to CCS_D */
+ if (isl_format_supports_ccs_d(devinfo, render_format))
+ return ISL_AUX_USAGE_CCS_D;
+
+ default:
+ return ISL_AUX_USAGE_NONE;
+ }
+}
+
+void
+iris_resource_prepare_render(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t layer_count,
+ enum isl_aux_usage aux_usage)
+{
+ iris_resource_prepare_access(ice, batch, res, level, 1, start_layer,
+ layer_count, aux_usage,
+ isl_aux_usage_has_fast_clears(aux_usage));
+}
+
+void
+iris_resource_finish_render(struct iris_context *ice,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t layer_count,
+ enum isl_aux_usage aux_usage)
+{
+ iris_resource_finish_write(ice, res, level, start_layer, layer_count,
+ aux_usage);
+}
+
+void
+iris_resource_prepare_depth(struct iris_context *ice,
+ struct iris_batch *batch,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t layer_count)
+{
+ iris_resource_prepare_access(ice, batch, res, level, 1, start_layer,
+ layer_count, res->aux.usage, !!res->aux.bo);
+}
+
+void
+iris_resource_finish_depth(struct iris_context *ice,
+ struct iris_resource *res, uint32_t level,
+ uint32_t start_layer, uint32_t layer_count,
+ bool depth_written)
+{
+ if (depth_written) {
+ iris_resource_finish_write(ice, res, level, start_layer, layer_count,
+ res->aux.usage);
+ }
}