#define __gen_address_offset anv_address_add
#include "common/gen_mi_builder.h"
+static void genX(flush_pipeline_select)(struct anv_cmd_buffer *cmd_buffer,
+ uint32_t pipeline);
+
static void
emit_lri(struct anv_batch *batch, uint32_t reg, uint32_t imm)
{
genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
{
struct anv_device *device = cmd_buffer->device;
+ UNUSED const struct gen_device_info *devinfo = &device->info;
uint32_t mocs = device->isl_dev.mocs.internal;
/* If we are emitting a new state base address we probably need to re-emit
pc.CommandStreamerStallEnable = true;
#if GEN_GEN >= 12
pc.TileCacheFlushEnable = true;
+#endif
+#if GEN_GEN == 12
+ /* GEN:BUG:1606662791:
+ *
+ * Software must program PIPE_CONTROL command with "HDC Pipeline
+ * Flush" prior to programming of the below two non-pipeline state :
+ * * STATE_BASE_ADDRESS
+ * * 3DSTATE_BINDING_TABLE_POOL_ALLOC
+ */
+ if (devinfo->revision == 0 /* A0 */)
+ pc.HDCPipelineFlushEnable = true;
#endif
}
+#if GEN_GEN == 12
+ /* GEN:BUG:1607854226:
+ *
+ * Workaround the non pipelined state not applying in MEDIA/GPGPU pipeline
+ * mode by putting the pipeline temporarily in 3D mode.
+ */
+ uint32_t gen12_wa_pipeline = cmd_buffer->state.current_pipeline;
+ genX(flush_pipeline_select_3d)(cmd_buffer);
+#endif
+
anv_batch_emit(&cmd_buffer->batch, GENX(STATE_BASE_ADDRESS), sba) {
sba.GeneralStateBaseAddress = (struct anv_address) { NULL, 0 };
sba.GeneralStateMOCS = mocs;
* these fields. However, since we will be growing the BO's live, we
* just set them all to the maximum.
*/
- sba.GeneralStateBufferSize = 0xfffff;
+ sba.GeneralStateBufferSize = 0xfffff;
+ sba.IndirectObjectBufferSize = 0xfffff;
+ if (device->physical->use_softpin) {
+ /* With softpin, we use fixed addresses so we actually know how big
+ * our base addresses are.
+ */
+ sba.DynamicStateBufferSize = DYNAMIC_STATE_POOL_SIZE / 4096;
+ sba.InstructionBufferSize = INSTRUCTION_STATE_POOL_SIZE / 4096;
+ } else {
+ sba.DynamicStateBufferSize = 0xfffff;
+ sba.InstructionBufferSize = 0xfffff;
+ }
sba.GeneralStateBufferSizeModifyEnable = true;
- sba.DynamicStateBufferSize = 0xfffff;
- sba.DynamicStateBufferSizeModifyEnable = true;
- sba.IndirectObjectBufferSize = 0xfffff;
sba.IndirectObjectBufferSizeModifyEnable = true;
- sba.InstructionBufferSize = 0xfffff;
+ sba.DynamicStateBufferSizeModifyEnable = true;
sba.InstructionBuffersizeModifyEnable = true;
# else
/* On gen7, we have upper bounds instead. According to the docs,
sba.InstructionAccessUpperBoundModifyEnable = true;
# endif
# if (GEN_GEN >= 9)
- if (cmd_buffer->device->instance->physicalDevice.use_softpin) {
+ if (cmd_buffer->device->physical->use_softpin) {
sba.BindlessSurfaceStateBaseAddress = (struct anv_address) {
.bo = device->surface_state_pool.block_pool.bo,
.offset = 0,
# endif
}
+#if GEN_GEN == 12
+ /* GEN:BUG:1607854226:
+ *
+ * Put the pipeline back into its current mode.
+ */
+ if (gen12_wa_pipeline != UINT32_MAX)
+ genX(flush_pipeline_select)(cmd_buffer, gen12_wa_pipeline);
+#endif
+
/* After re-setting the surface state base address, we have to do some
* cache flusing so that the sampler engine will pick up the new
* SURFACE_STATE objects and binding tables. From the Broadwell PRM,
}
}
+static bool
+isl_color_value_requires_conversion(union isl_color_value color,
+ const struct isl_surf *surf,
+ const struct isl_view *view)
+{
+ if (surf->format == view->format && isl_swizzle_is_identity(view->swizzle))
+ return false;
+
+ uint32_t surf_pack[4] = { 0, 0, 0, 0 };
+ isl_color_value_pack(&color, surf->format, surf_pack);
+
+ uint32_t view_pack[4] = { 0, 0, 0, 0 };
+ union isl_color_value swiz_color =
+ isl_color_value_swizzle_inv(color, view->swizzle);
+ isl_color_value_pack(&swiz_color, view->format, view_pack);
+
+ return memcmp(surf_pack, view_pack, sizeof(surf_pack)) != 0;
+}
+
static void
color_attachment_compute_aux_usage(struct anv_device * device,
struct anv_cmd_state * cmd_state,
- uint32_t att, VkRect2D render_area,
- union isl_color_value *fast_clear_color)
+ uint32_t att, VkRect2D render_area)
{
struct anv_attachment_state *att_state = &cmd_state->attachments[att];
struct anv_image_view *iview = cmd_state->attachments[att].image_view;
att_state->aux_usage =
anv_layout_to_aux_usage(&device->info, iview->image,
VK_IMAGE_ASPECT_COLOR_BIT,
+ VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
/* If we don't have aux, then we should have returned early in the layer
*/
if (cmd_state->pass->attachments[att].first_subpass_layout ==
VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL) {
- anv_perf_warn(device->instance, iview->image,
+ anv_perf_warn(device, iview->image,
"Not temporarily enabling CCS_E.");
}
} else {
union isl_color_value clear_color = {};
anv_clear_color_from_att_state(&clear_color, att_state, iview);
- att_state->clear_color_is_zero_one =
+ const bool clear_color_is_zero_one =
isl_color_value_is_zero_one(clear_color, iview->planes[0].isl.format);
att_state->clear_color_is_zero =
isl_color_value_is_zero(clear_color, iview->planes[0].isl.format);
att_state->fast_clear = false;
/* On Broadwell and earlier, we can only handle 0/1 clear colors */
- if (GEN_GEN <= 8 && !att_state->clear_color_is_zero_one)
+ if (GEN_GEN <= 8 && !clear_color_is_zero_one)
att_state->fast_clear = false;
+ /* If the clear color is one that would require non-trivial format
+ * conversion on resolve, we don't bother with the fast clear. This
+ * shouldn't be common as most clear colors are 0/1 and the most common
+ * format re-interpretation is for sRGB.
+ */
+ if (isl_color_value_requires_conversion(clear_color,
+ &iview->image->planes[0].surface.isl,
+ &iview->planes[0].isl)) {
+ anv_perf_warn(device, iview,
+ "Cannot fast-clear to colors which would require "
+ "format conversion on resolve");
+ att_state->fast_clear = false;
+ }
+
/* We only allow fast clears to the first slice of an image (level 0,
* layer 0) and only for the entire slice. This guarantees us that, at
* any given time, there is only one clear color on any given image at
if (att_state->fast_clear &&
(iview->planes[0].isl.base_level > 0 ||
iview->planes[0].isl.base_array_layer > 0)) {
- anv_perf_warn(device->instance, iview->image,
+ anv_perf_warn(device, iview->image,
"Rendering with multi-lod or multi-layer framebuffer "
"with LOAD_OP_LOAD and baseMipLevel > 0 or "
"baseArrayLayer > 0. Not fast clearing.");
att_state->fast_clear = false;
} else if (att_state->fast_clear && cmd_state->framebuffer->layers > 1) {
- anv_perf_warn(device->instance, iview->image,
+ anv_perf_warn(device, iview->image,
"Rendering to a multi-layer framebuffer with "
"LOAD_OP_CLEAR. Only fast-clearing the first slice");
}
-
- if (att_state->fast_clear)
- *fast_clear_color = clear_color;
} else {
att_state->fast_clear = false;
}
}
-static void
-depth_stencil_attachment_compute_aux_usage(struct anv_device *device,
- struct anv_cmd_state *cmd_state,
- uint32_t att, VkRect2D render_area)
+static bool
+anv_can_hiz_clear_ds_view(struct anv_device *device,
+ struct anv_image_view *iview,
+ VkImageLayout layout,
+ VkImageAspectFlags clear_aspects,
+ float depth_clear_value,
+ VkRect2D render_area)
{
- struct anv_render_pass_attachment *pass_att =
- &cmd_state->pass->attachments[att];
- struct anv_attachment_state *att_state = &cmd_state->attachments[att];
- struct anv_image_view *iview = cmd_state->attachments[att].image_view;
-
- /* These will be initialized after the first subpass transition. */
- att_state->aux_usage = ISL_AUX_USAGE_NONE;
- att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
-
- if (GEN_GEN == 7) {
- /* We don't do any HiZ or depth fast-clears on gen7 yet */
- att_state->fast_clear = false;
- return;
- }
-
- if (!(att_state->pending_clear_aspects & VK_IMAGE_ASPECT_DEPTH_BIT)) {
- /* If we're just clearing stencil, we can always HiZ clear */
- att_state->fast_clear = true;
- return;
- }
+ /* We don't do any HiZ or depth fast-clears on gen7 yet */
+ if (GEN_GEN == 7)
+ return false;
- /* Default to false for now */
- att_state->fast_clear = false;
+ /* If we're just clearing stencil, we can always HiZ clear */
+ if (!(clear_aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
+ return true;
/* We must have depth in order to have HiZ */
if (!(iview->image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
- return;
+ return false;
- const enum isl_aux_usage first_subpass_aux_usage =
+ const enum isl_aux_usage clear_aux_usage =
anv_layout_to_aux_usage(&device->info, iview->image,
VK_IMAGE_ASPECT_DEPTH_BIT,
- pass_att->first_subpass_layout);
+ VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
+ layout);
if (!blorp_can_hiz_clear_depth(&device->info,
&iview->image->planes[0].surface.isl,
- first_subpass_aux_usage,
+ clear_aux_usage,
iview->planes[0].isl.base_level,
iview->planes[0].isl.base_array_layer,
render_area.offset.x,
render_area.extent.width,
render_area.offset.y +
render_area.extent.height))
- return;
+ return false;
- if (att_state->clear_value.depthStencil.depth != ANV_HZ_FC_VAL)
- return;
+ if (depth_clear_value != ANV_HZ_FC_VAL)
+ return false;
- if (GEN_GEN == 8 && anv_can_sample_with_hiz(&device->info, iview->image)) {
- /* Only gen9+ supports returning ANV_HZ_FC_VAL when sampling a
- * fast-cleared portion of a HiZ buffer. Testing has revealed that Gen8
- * only supports returning 0.0f. Gens prior to gen8 do not support this
- * feature at all.
- */
- return;
- }
+ /* Only gen9+ supports returning ANV_HZ_FC_VAL when sampling a fast-cleared
+ * portion of a HiZ buffer. Testing has revealed that Gen8 only supports
+ * returning 0.0f. Gens prior to gen8 do not support this feature at all.
+ */
+ if (GEN_GEN == 8 && anv_can_sample_with_hiz(&device->info, iview->image))
+ return false;
/* If we got here, then we can fast clear */
- att_state->fast_clear = true;
+ return true;
}
-static bool
-need_input_attachment_state(const struct anv_render_pass_attachment *att)
+#define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
+
+#if GEN_GEN == 12
+static void
+anv_image_init_aux_tt(struct anv_cmd_buffer *cmd_buffer,
+ const struct anv_image *image,
+ VkImageAspectFlagBits aspect,
+ uint32_t base_level, uint32_t level_count,
+ uint32_t base_layer, uint32_t layer_count)
{
- if (!(att->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT))
- return false;
+ uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
+
+ uint64_t base_address =
+ anv_address_physical(image->planes[plane].address);
+
+ const struct isl_surf *isl_surf = &image->planes[plane].surface.isl;
+ uint64_t format_bits = gen_aux_map_format_bits_for_isl_surf(isl_surf);
- /* We only allocate input attachment states for color surfaces. Compression
- * is not yet enabled for depth textures and stencil doesn't allow
- * compression so we can just use the texture surface state from the view.
+ /* We're about to live-update the AUX-TT. We really don't want anyone else
+ * trying to read it while we're doing this. We could probably get away
+ * with not having this stall in some cases if we were really careful but
+ * it's better to play it safe. Full stall the GPU.
*/
- return vk_format_is_color(att->format);
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_END_OF_PIPE_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ for (uint32_t a = 0; a < layer_count; a++) {
+ const uint32_t layer = base_layer + a;
+
+ uint64_t start_offset_B = UINT64_MAX, end_offset_B = 0;
+ for (uint32_t l = 0; l < level_count; l++) {
+ const uint32_t level = base_level + l;
+
+ uint32_t logical_array_layer, logical_z_offset_px;
+ if (image->type == VK_IMAGE_TYPE_3D) {
+ logical_array_layer = 0;
+
+ /* If the given miplevel does not have this layer, then any higher
+ * miplevels won't either because miplevels only get smaller the
+ * higher the LOD.
+ */
+ assert(layer < image->extent.depth);
+ if (layer >= anv_minify(image->extent.depth, level))
+ break;
+ logical_z_offset_px = layer;
+ } else {
+ assert(layer < image->array_size);
+ logical_array_layer = layer;
+ logical_z_offset_px = 0;
+ }
+
+ uint32_t slice_start_offset_B, slice_end_offset_B;
+ isl_surf_get_image_range_B_tile(isl_surf, level,
+ logical_array_layer,
+ logical_z_offset_px,
+ &slice_start_offset_B,
+ &slice_end_offset_B);
+
+ start_offset_B = MIN2(start_offset_B, slice_start_offset_B);
+ end_offset_B = MAX2(end_offset_B, slice_end_offset_B);
+ }
+
+ /* Aux operates 64K at a time */
+ start_offset_B = align_down_u64(start_offset_B, 64 * 1024);
+ end_offset_B = align_u64(end_offset_B, 64 * 1024);
+
+ for (uint64_t offset = start_offset_B;
+ offset < end_offset_B; offset += 64 * 1024) {
+ uint64_t address = base_address + offset;
+
+ uint64_t aux_entry_addr64, *aux_entry_map;
+ aux_entry_map = gen_aux_map_get_entry(cmd_buffer->device->aux_map_ctx,
+ address, &aux_entry_addr64);
+
+ assert(cmd_buffer->device->physical->use_softpin);
+ struct anv_address aux_entry_address = {
+ .bo = NULL,
+ .offset = aux_entry_addr64,
+ };
+
+ const uint64_t old_aux_entry = READ_ONCE(*aux_entry_map);
+ uint64_t new_aux_entry =
+ (old_aux_entry & GEN_AUX_MAP_ADDRESS_MASK) | format_bits;
+
+ if (isl_aux_usage_has_ccs(image->planes[plane].aux_usage))
+ new_aux_entry |= GEN_AUX_MAP_ENTRY_VALID_BIT;
+
+ gen_mi_store(&b, gen_mi_mem64(aux_entry_address),
+ gen_mi_imm(new_aux_entry));
+ }
+ }
+
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT;
}
+#endif /* GEN_GEN == 12 */
/* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
* the initial layout is undefined, the HiZ buffer and depth buffer will
static void
transition_depth_buffer(struct anv_cmd_buffer *cmd_buffer,
const struct anv_image *image,
+ uint32_t base_layer, uint32_t layer_count,
VkImageLayout initial_layout,
VkImageLayout final_layout)
{
- const bool hiz_enabled = ISL_AUX_USAGE_HIZ ==
- anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
- VK_IMAGE_ASPECT_DEPTH_BIT, initial_layout);
- const bool enable_hiz = ISL_AUX_USAGE_HIZ ==
- anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
- VK_IMAGE_ASPECT_DEPTH_BIT, final_layout);
-
- enum isl_aux_op hiz_op;
- if (hiz_enabled && !enable_hiz) {
- hiz_op = ISL_AUX_OP_FULL_RESOLVE;
- } else if (!hiz_enabled && enable_hiz) {
- hiz_op = ISL_AUX_OP_AMBIGUATE;
- } else {
- assert(hiz_enabled == enable_hiz);
- /* If the same buffer will be used, no resolves are necessary. */
- hiz_op = ISL_AUX_OP_NONE;
+ uint32_t depth_plane =
+ anv_image_aspect_to_plane(image->aspects, VK_IMAGE_ASPECT_DEPTH_BIT);
+ if (image->planes[depth_plane].aux_usage == ISL_AUX_USAGE_NONE)
+ return;
+
+#if GEN_GEN == 12
+ if ((initial_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
+ initial_layout == VK_IMAGE_LAYOUT_PREINITIALIZED) &&
+ cmd_buffer->device->physical->has_implicit_ccs &&
+ cmd_buffer->device->info.has_aux_map) {
+ anv_image_init_aux_tt(cmd_buffer, image, VK_IMAGE_ASPECT_DEPTH_BIT,
+ 0, 1, 0, 1);
}
+#endif
+
+ const enum isl_aux_state initial_state =
+ anv_layout_to_aux_state(&cmd_buffer->device->info, image,
+ VK_IMAGE_ASPECT_DEPTH_BIT,
+ initial_layout);
+ const enum isl_aux_state final_state =
+ anv_layout_to_aux_state(&cmd_buffer->device->info, image,
+ VK_IMAGE_ASPECT_DEPTH_BIT,
+ final_layout);
+
+ const bool initial_depth_valid =
+ isl_aux_state_has_valid_primary(initial_state);
+ const bool initial_hiz_valid =
+ isl_aux_state_has_valid_aux(initial_state);
+ const bool final_needs_depth =
+ isl_aux_state_has_valid_primary(final_state);
+ const bool final_needs_hiz =
+ isl_aux_state_has_valid_aux(final_state);
+
+ /* Getting into the pass-through state for Depth is tricky and involves
+ * both a resolve and an ambiguate. We don't handle that state right now
+ * as anv_layout_to_aux_state never returns it.
+ */
+ assert(final_state != ISL_AUX_STATE_PASS_THROUGH);
- if (hiz_op != ISL_AUX_OP_NONE)
+ if (final_needs_depth && !initial_depth_valid) {
+ assert(initial_hiz_valid);
anv_image_hiz_op(cmd_buffer, image, VK_IMAGE_ASPECT_DEPTH_BIT,
- 0, 0, 1, hiz_op);
+ 0, base_layer, layer_count, ISL_AUX_OP_FULL_RESOLVE);
+ } else if (final_needs_hiz && !initial_hiz_valid) {
+ assert(initial_depth_valid);
+ anv_image_hiz_op(cmd_buffer, image, VK_IMAGE_ASPECT_DEPTH_BIT,
+ 0, base_layer, layer_count, ISL_AUX_OP_AMBIGUATE);
+ }
}
static inline bool
anv_cmd_predicated_ccs_resolve(struct anv_cmd_buffer *cmd_buffer,
const struct anv_image *image,
enum isl_format format,
+ struct isl_swizzle swizzle,
VkImageAspectFlagBits aspect,
uint32_t level, uint32_t array_layer,
enum isl_aux_op resolve_op,
* to do a partial resolve on a CCS_D surface.
*/
if (resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE &&
- image->planes[plane].aux_usage == ISL_AUX_USAGE_NONE)
+ image->planes[plane].aux_usage == ISL_AUX_USAGE_CCS_D)
resolve_op = ISL_AUX_OP_FULL_RESOLVE;
- anv_image_ccs_op(cmd_buffer, image, format, aspect, level,
- array_layer, 1, resolve_op, NULL, true);
+ anv_image_ccs_op(cmd_buffer, image, format, swizzle, aspect,
+ level, array_layer, 1, resolve_op, NULL, true);
}
static void
anv_cmd_predicated_mcs_resolve(struct anv_cmd_buffer *cmd_buffer,
const struct anv_image *image,
enum isl_format format,
+ struct isl_swizzle swizzle,
VkImageAspectFlagBits aspect,
uint32_t array_layer,
enum isl_aux_op resolve_op,
aspect, 0, array_layer,
resolve_op, fast_clear_supported);
- anv_image_mcs_op(cmd_buffer, image, format, aspect,
+ anv_image_mcs_op(cmd_buffer, image, format, swizzle, aspect,
array_layer, 1, resolve_op, NULL, true);
#else
unreachable("MCS resolves are unsupported on Ivybridge and Bay Trail");
VkImageLayout initial_layout,
VkImageLayout final_layout)
{
- const struct gen_device_info *devinfo = &cmd_buffer->device->info;
+ struct anv_device *device = cmd_buffer->device;
+ const struct gen_device_info *devinfo = &device->info;
/* Validate the inputs. */
assert(cmd_buffer);
assert(image && image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);
if (base_layer >= anv_image_aux_layers(image, aspect, base_level))
return;
- assert(image->tiling == VK_IMAGE_TILING_OPTIMAL);
+ assert(image->planes[plane].surface.isl.tiling != ISL_TILING_LINEAR);
if (initial_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
initial_layout == VK_IMAGE_LAYOUT_PREINITIALIZED) {
+#if GEN_GEN == 12
+ if (device->physical->has_implicit_ccs && devinfo->has_aux_map) {
+ anv_image_init_aux_tt(cmd_buffer, image, aspect,
+ base_level, level_count,
+ base_layer, layer_count);
+ }
+#else
+ assert(!(device->physical->has_implicit_ccs && devinfo->has_aux_map));
+#endif
+
/* A subresource in the undefined layout may have been aliased and
* populated with any arrangement of bits. Therefore, we must initialize
* the related aux buffer and clear buffer entry with desirable values.
anv_image_ccs_op(cmd_buffer, image,
image->planes[plane].surface.isl.format,
+ ISL_SWIZZLE_IDENTITY,
aspect, level, base_layer, level_layer_count,
ISL_AUX_OP_AMBIGUATE, NULL, false);
}
} else {
if (image->samples == 4 || image->samples == 16) {
- anv_perf_warn(cmd_buffer->device->instance, image,
+ anv_perf_warn(cmd_buffer->device, image,
"Doing a potentially unnecessary fast-clear to "
"define an MCS buffer.");
}
assert(base_level == 0 && level_count == 1);
anv_image_mcs_op(cmd_buffer, image,
image->planes[plane].surface.isl.format,
+ ISL_SWIZZLE_IDENTITY,
aspect, base_layer, layer_count,
ISL_AUX_OP_FAST_CLEAR, NULL, false);
}
}
const enum isl_aux_usage initial_aux_usage =
- anv_layout_to_aux_usage(devinfo, image, aspect, initial_layout);
+ anv_layout_to_aux_usage(devinfo, image, aspect, 0, initial_layout);
const enum isl_aux_usage final_aux_usage =
- anv_layout_to_aux_usage(devinfo, image, aspect, final_layout);
+ anv_layout_to_aux_usage(devinfo, image, aspect, 0, final_layout);
/* The current code assumes that there is no mixing of CCS_E and CCS_D.
* We can handle transitions between CCS_D/E to and from NONE. What we
* we do any more rendering or clearing.
*/
cmd_buffer->state.pending_pipe_bits |=
- ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_CS_STALL_BIT;
+ ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_END_OF_PIPE_SYNC_BIT;
for (uint32_t l = 0; l < level_count; l++) {
uint32_t level = base_level + l;
if (image->samples == 1) {
anv_cmd_predicated_ccs_resolve(cmd_buffer, image,
image->planes[plane].surface.isl.format,
+ ISL_SWIZZLE_IDENTITY,
aspect, level, array_layer, resolve_op,
final_fast_clear);
} else {
anv_cmd_predicated_mcs_resolve(cmd_buffer, image,
image->planes[plane].surface.isl.format,
+ ISL_SWIZZLE_IDENTITY,
aspect, array_layer, resolve_op,
final_fast_clear);
}
}
cmd_buffer->state.pending_pipe_bits |=
- ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_CS_STALL_BIT;
+ ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_END_OF_PIPE_SYNC_BIT;
}
-/**
- * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
- */
static VkResult
genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
- struct anv_render_pass *pass,
+ const struct anv_render_pass *pass,
+ const struct anv_framebuffer *framebuffer,
const VkRenderPassBeginInfo *begin)
{
- const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
struct anv_cmd_state *state = &cmd_buffer->state;
- struct anv_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
vk_free(&cmd_buffer->pool->alloc, state->attachments);
if (pass->attachment_count > 0) {
- state->attachments = vk_alloc(&cmd_buffer->pool->alloc,
- pass->attachment_count *
- sizeof(state->attachments[0]),
- 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+ state->attachments = vk_zalloc(&cmd_buffer->pool->alloc,
+ pass->attachment_count *
+ sizeof(state->attachments[0]),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (state->attachments == NULL) {
/* Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
return anv_batch_set_error(&cmd_buffer->batch,
state->attachments = NULL;
}
- /* Reserve one for the NULL state. */
- unsigned num_states = 1;
- for (uint32_t i = 0; i < pass->attachment_count; ++i) {
- if (vk_format_is_color(pass->attachments[i].format))
- num_states++;
-
- if (need_input_attachment_state(&pass->attachments[i]))
- num_states++;
- }
-
- const uint32_t ss_stride = align_u32(isl_dev->ss.size, isl_dev->ss.align);
- state->render_pass_states =
- anv_state_stream_alloc(&cmd_buffer->surface_state_stream,
- num_states * ss_stride, isl_dev->ss.align);
-
- struct anv_state next_state = state->render_pass_states;
- next_state.alloc_size = isl_dev->ss.size;
-
- state->null_surface_state = next_state;
- next_state.offset += ss_stride;
- next_state.map += ss_stride;
-
- const VkRenderPassAttachmentBeginInfoKHR *begin_attachment =
+ const VkRenderPassAttachmentBeginInfoKHR *attach_begin =
vk_find_struct_const(begin, RENDER_PASS_ATTACHMENT_BEGIN_INFO_KHR);
-
- if (begin && !begin_attachment)
+ if (begin && !attach_begin)
assert(pass->attachment_count == framebuffer->attachment_count);
for (uint32_t i = 0; i < pass->attachment_count; ++i) {
- if (vk_format_is_color(pass->attachments[i].format)) {
- state->attachments[i].color.state = next_state;
- next_state.offset += ss_stride;
- next_state.map += ss_stride;
- }
-
- if (need_input_attachment_state(&pass->attachments[i])) {
- state->attachments[i].input.state = next_state;
- next_state.offset += ss_stride;
- next_state.map += ss_stride;
- }
-
- if (begin_attachment && begin_attachment->attachmentCount != 0) {
- assert(begin_attachment->attachmentCount == pass->attachment_count);
- ANV_FROM_HANDLE(anv_image_view, iview, begin_attachment->pAttachments[i]);
- cmd_buffer->state.attachments[i].image_view = iview;
+ if (attach_begin && attach_begin->attachmentCount != 0) {
+ assert(attach_begin->attachmentCount == pass->attachment_count);
+ ANV_FROM_HANDLE(anv_image_view, iview, attach_begin->pAttachments[i]);
+ state->attachments[i].image_view = iview;
} else if (framebuffer && i < framebuffer->attachment_count) {
- cmd_buffer->state.attachments[i].image_view = framebuffer->attachments[i];
+ state->attachments[i].image_view = framebuffer->attachments[i];
+ } else {
+ state->attachments[i].image_view = NULL;
}
}
- assert(next_state.offset == state->render_pass_states.offset +
- state->render_pass_states.alloc_size);
if (begin) {
- isl_null_fill_state(isl_dev, state->null_surface_state.map,
- isl_extent3d(framebuffer->width,
- framebuffer->height,
- framebuffer->layers));
-
for (uint32_t i = 0; i < pass->attachment_count; ++i) {
- struct anv_render_pass_attachment *att = &pass->attachments[i];
- VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
+ const struct anv_render_pass_attachment *pass_att = &pass->attachments[i];
+ struct anv_attachment_state *att_state = &state->attachments[i];
+ VkImageAspectFlags att_aspects = vk_format_aspects(pass_att->format);
VkImageAspectFlags clear_aspects = 0;
VkImageAspectFlags load_aspects = 0;
if (att_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
/* color attachment */
- if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ if (pass_att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
- } else if (att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
+ } else if (pass_att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
load_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
}
} else {
/* depthstencil attachment */
if (att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
- if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ if (pass_att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
- } else if (att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
+ } else if (pass_att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
load_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
}
}
if (att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
- if (att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ if (pass_att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
- } else if (att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
+ } else if (pass_att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
load_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
}
}
}
- state->attachments[i].current_layout = att->initial_layout;
- state->attachments[i].current_stencil_layout = att->stencil_initial_layout;
- state->attachments[i].pending_clear_aspects = clear_aspects;
- state->attachments[i].pending_load_aspects = load_aspects;
+ att_state->current_layout = pass_att->initial_layout;
+ att_state->current_stencil_layout = pass_att->stencil_initial_layout;
+ att_state->pending_clear_aspects = clear_aspects;
+ att_state->pending_load_aspects = load_aspects;
if (clear_aspects)
- state->attachments[i].clear_value = begin->pClearValues[i];
+ att_state->clear_value = begin->pClearValues[i];
- struct anv_image_view *iview = cmd_buffer->state.attachments[i].image_view;
- anv_assert(iview->vk_format == att->format);
+ struct anv_image_view *iview = state->attachments[i].image_view;
+ anv_assert(iview->vk_format == pass_att->format);
const uint32_t num_layers = iview->planes[0].isl.array_len;
- state->attachments[i].pending_clear_views = (1 << num_layers) - 1;
+ att_state->pending_clear_views = (1 << num_layers) - 1;
- union isl_color_value clear_color = { .u32 = { 0, } };
if (att_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
anv_assert(iview->n_planes == 1);
assert(att_aspects == VK_IMAGE_ASPECT_COLOR_BIT);
color_attachment_compute_aux_usage(cmd_buffer->device,
- state, i, begin->renderArea,
- &clear_color);
-
- anv_image_fill_surface_state(cmd_buffer->device,
- iview->image,
- VK_IMAGE_ASPECT_COLOR_BIT,
- &iview->planes[0].isl,
- ISL_SURF_USAGE_RENDER_TARGET_BIT,
- state->attachments[i].aux_usage,
- &clear_color,
- 0,
- &state->attachments[i].color,
- NULL);
-
- add_surface_state_relocs(cmd_buffer, state->attachments[i].color);
+ state, i, begin->renderArea);
} else {
- depth_stencil_attachment_compute_aux_usage(cmd_buffer->device,
- state, i,
- begin->renderArea);
- }
-
- if (need_input_attachment_state(&pass->attachments[i])) {
- anv_image_fill_surface_state(cmd_buffer->device,
- iview->image,
- VK_IMAGE_ASPECT_COLOR_BIT,
- &iview->planes[0].isl,
- ISL_SURF_USAGE_TEXTURE_BIT,
- state->attachments[i].input_aux_usage,
- &clear_color,
- 0,
- &state->attachments[i].input,
- NULL);
-
- add_surface_state_relocs(cmd_buffer, state->attachments[i].input);
+ /* These will be initialized after the first subpass transition. */
+ att_state->aux_usage = ISL_AUX_USAGE_NONE;
+ att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
+ att_state->fast_clear =
+ anv_can_hiz_clear_ds_view(cmd_buffer->device, iview,
+ pass_att->first_subpass_layout,
+ clear_aspects,
+ att_state->clear_value.depthStencil.depth,
+ begin->renderArea);
}
}
}
return VK_SUCCESS;
}
+/**
+ * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
+ */
+static VkResult
+genX(cmd_buffer_alloc_att_surf_states)(struct anv_cmd_buffer *cmd_buffer,
+ const struct anv_render_pass *pass,
+ const struct anv_subpass *subpass)
+{
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+ struct anv_cmd_state *state = &cmd_buffer->state;
+
+ /* Reserve one for the NULL state. */
+ unsigned num_states = 1;
+ for (uint32_t i = 0; i < subpass->attachment_count; i++) {
+ uint32_t att = subpass->attachments[i].attachment;
+ if (att == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ assert(att < pass->attachment_count);
+ if (!vk_format_is_color(pass->attachments[att].format))
+ continue;
+
+ const VkImageUsageFlagBits att_usage = subpass->attachments[i].usage;
+ assert(util_bitcount(att_usage) == 1);
+
+ if (att_usage == VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT ||
+ att_usage == VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT)
+ num_states++;
+ }
+
+ const uint32_t ss_stride = align_u32(isl_dev->ss.size, isl_dev->ss.align);
+ state->attachment_states =
+ anv_state_stream_alloc(&cmd_buffer->surface_state_stream,
+ num_states * ss_stride, isl_dev->ss.align);
+ if (state->attachment_states.map == NULL) {
+ return anv_batch_set_error(&cmd_buffer->batch,
+ VK_ERROR_OUT_OF_DEVICE_MEMORY);
+ }
+
+ struct anv_state next_state = state->attachment_states;
+ next_state.alloc_size = isl_dev->ss.size;
+
+ state->null_surface_state = next_state;
+ next_state.offset += ss_stride;
+ next_state.map += ss_stride;
+
+ for (uint32_t i = 0; i < subpass->attachment_count; i++) {
+ uint32_t att = subpass->attachments[i].attachment;
+ if (att == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ assert(att < pass->attachment_count);
+ if (!vk_format_is_color(pass->attachments[att].format))
+ continue;
+
+ const VkImageUsageFlagBits att_usage = subpass->attachments[i].usage;
+ assert(util_bitcount(att_usage) == 1);
+
+ if (att_usage == VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT)
+ state->attachments[att].color.state = next_state;
+ else if (att_usage == VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT)
+ state->attachments[att].input.state = next_state;
+ else
+ continue;
+
+ state->attachments[att].color.state = next_state;
+ next_state.offset += ss_stride;
+ next_state.map += ss_stride;
+ }
+
+ assert(next_state.offset == state->attachment_states.offset +
+ state->attachment_states.alloc_size);
+
+ return VK_SUCCESS;
+}
+
VkResult
genX(BeginCommandBuffer)(
VkCommandBuffer commandBuffer,
*/
cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_VF_CACHE_INVALIDATE_BIT;
+ /* Re-emit the aux table register in every command buffer. This way we're
+ * ensured that we have the table even if this command buffer doesn't
+ * initialize any images.
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT;
+
/* We send an "Indirect State Pointers Disable" packet at
* EndCommandBuffer, so all push contant packets are ignored during a
* context restore. Documentation says after that command, we need to
if (cmd_buffer->usage_flags &
VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
assert(pBeginInfo->pInheritanceInfo);
- cmd_buffer->state.pass =
- anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
- cmd_buffer->state.subpass =
- &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ ANV_FROM_HANDLE(anv_render_pass, pass,
+ pBeginInfo->pInheritanceInfo->renderPass);
+ struct anv_subpass *subpass =
+ &pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ ANV_FROM_HANDLE(anv_framebuffer, framebuffer,
+ pBeginInfo->pInheritanceInfo->framebuffer);
+
+ cmd_buffer->state.pass = pass;
+ cmd_buffer->state.subpass = subpass;
/* This is optional in the inheritance info. */
- cmd_buffer->state.framebuffer =
- anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
+ cmd_buffer->state.framebuffer = framebuffer;
+
+ result = genX(cmd_buffer_setup_attachments)(cmd_buffer, pass,
+ framebuffer, NULL);
+ if (result != VK_SUCCESS)
+ return result;
- result = genX(cmd_buffer_setup_attachments)(cmd_buffer,
- cmd_buffer->state.pass, NULL);
+ result = genX(cmd_buffer_alloc_att_surf_states)(cmd_buffer, pass,
+ subpass);
+ if (result != VK_SUCCESS)
+ return result;
/* Record that HiZ is enabled if we can. */
if (cmd_buffer->state.framebuffer) {
enum isl_aux_usage aux_usage =
anv_layout_to_aux_usage(&cmd_buffer->device->info, iview->image,
- VK_IMAGE_ASPECT_DEPTH_BIT, layout);
+ VK_IMAGE_ASPECT_DEPTH_BIT,
+ VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
+ layout);
- cmd_buffer->state.hiz_enabled = aux_usage == ISL_AUX_USAGE_HIZ;
+ cmd_buffer->state.hiz_enabled = isl_aux_usage_has_hiz(aux_usage);
}
}
*/
struct anv_bo *ss_bo =
primary->device->surface_state_pool.block_pool.bo;
- struct anv_state src_state = primary->state.render_pass_states;
- struct anv_state dst_state = secondary->state.render_pass_states;
+ struct anv_state src_state = primary->state.attachment_states;
+ struct anv_state dst_state = secondary->state.attachment_states;
assert(src_state.alloc_size == dst_state.alloc_size);
genX(cmd_buffer_so_memcpy)(primary,
uint32_t l3cr;
anv_pack_struct(&l3cr, L3_ALLOCATION_REG,
-#if GEN_GEN < 12
+#if GEN_GEN < 11
.SLMEnable = has_slm,
#endif
#if GEN_GEN == 11
emit_lri(&cmd_buffer->batch, GENX(L3CNTLREG3_num), l3cr3);
#if GEN_IS_HASWELL
- if (cmd_buffer->device->instance->physicalDevice.cmd_parser_version >= 4) {
+ if (cmd_buffer->device->physical->cmd_parser_version >= 4) {
/* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
* them disabled to avoid crashing the system hard.
*/
void
genX(cmd_buffer_apply_pipe_flushes)(struct anv_cmd_buffer *cmd_buffer)
{
+ UNUSED const struct gen_device_info *devinfo = &cmd_buffer->device->info;
enum anv_pipe_bits bits = cmd_buffer->state.pending_pipe_bits;
- if (cmd_buffer->device->instance->physicalDevice.always_flush_cache)
+ if (cmd_buffer->device->physical->always_flush_cache)
bits |= ANV_PIPE_FLUSH_BITS | ANV_PIPE_INVALIDATE_BITS;
- /* Flushes are pipelined while invalidations are handled immediately.
- * Therefore, if we're flushing anything then we need to schedule a stall
- * before any invalidations can happen.
+ /*
+ * From Sandybridge PRM, volume 2, "1.7.2 End-of-Pipe Synchronization":
+ *
+ * Write synchronization is a special case of end-of-pipe
+ * synchronization that requires that the render cache and/or depth
+ * related caches are flushed to memory, where the data will become
+ * globally visible. This type of synchronization is required prior to
+ * SW (CPU) actually reading the result data from memory, or initiating
+ * an operation that will use as a read surface (such as a texture
+ * surface) a previous render target and/or depth/stencil buffer
+ *
+ *
+ * From Haswell PRM, volume 2, part 1, "End-of-Pipe Synchronization":
+ *
+ * Exercising the write cache flush bits (Render Target Cache Flush
+ * Enable, Depth Cache Flush Enable, DC Flush) in PIPE_CONTROL only
+ * ensures the write caches are flushed and doesn't guarantee the data
+ * is globally visible.
+ *
+ * SW can track the completion of the end-of-pipe-synchronization by
+ * using "Notify Enable" and "PostSync Operation - Write Immediate
+ * Data" in the PIPE_CONTROL command.
+ *
+ * In other words, flushes are pipelined while invalidations are handled
+ * immediately. Therefore, if we're flushing anything then we need to
+ * schedule an end-of-pipe sync before any invalidations can happen.
*/
if (bits & ANV_PIPE_FLUSH_BITS)
- bits |= ANV_PIPE_NEEDS_CS_STALL_BIT;
+ bits |= ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT;
- /* If we're going to do an invalidate and we have a pending CS stall that
- * has yet to be resolved, we do the CS stall now.
+
+ /* HSD 1209978178: docs say that before programming the aux table:
+ *
+ * "Driver must ensure that the engine is IDLE but ensure it doesn't
+ * add extra flushes in the case it knows that the engine is already
+ * IDLE."
+ */
+ if (GEN_GEN == 12 && (bits & ANV_PIPE_AUX_TABLE_INVALIDATE_BIT))
+ bits |= ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT;
+
+ /* If we're going to do an invalidate and we have a pending end-of-pipe
+ * sync that has yet to be resolved, we do the end-of-pipe sync now.
*/
if ((bits & ANV_PIPE_INVALIDATE_BITS) &&
- (bits & ANV_PIPE_NEEDS_CS_STALL_BIT)) {
- bits |= ANV_PIPE_CS_STALL_BIT;
- bits &= ~ANV_PIPE_NEEDS_CS_STALL_BIT;
+ (bits & ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT)) {
+ bits |= ANV_PIPE_END_OF_PIPE_SYNC_BIT;
+ bits &= ~ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT;
}
if (GEN_GEN >= 12 &&
bits |= ANV_PIPE_TILE_CACHE_FLUSH_BIT;
}
+ /* GEN:BUG:1409226450, Wait for EU to be idle before pipe control which
+ * invalidates the instruction cache
+ */
+ if (GEN_GEN == 12 && (bits & ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT))
+ bits |= ANV_PIPE_CS_STALL_BIT | ANV_PIPE_STALL_AT_SCOREBOARD_BIT;
+
if ((GEN_GEN >= 8 && GEN_GEN <= 9) &&
(bits & ANV_PIPE_CS_STALL_BIT) &&
(bits & ANV_PIPE_VF_CACHE_INVALIDATE_BIT)) {
sizeof(cmd_buffer->state.gfx.ib_dirty_range));
}
- if (bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT)) {
+ /* Project: SKL / Argument: LRI Post Sync Operation [23]
+ *
+ * "PIPECONTROL command with “Command Streamer Stall Enable” must be
+ * programmed prior to programming a PIPECONTROL command with "LRI
+ * Post Sync Operation" in GPGPU mode of operation (i.e when
+ * PIPELINE_SELECT command is set to GPGPU mode of operation)."
+ *
+ * The same text exists a few rows below for Post Sync Op.
+ *
+ * On Gen12 this is GEN:BUG:1607156449.
+ */
+ if (bits & ANV_PIPE_POST_SYNC_BIT) {
+ if ((GEN_GEN == 9 || (GEN_GEN == 12 && devinfo->revision == 0 /* A0 */)) &&
+ cmd_buffer->state.current_pipeline == GPGPU)
+ bits |= ANV_PIPE_CS_STALL_BIT;
+ bits &= ~ANV_PIPE_POST_SYNC_BIT;
+ }
+
+ if (bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT |
+ ANV_PIPE_END_OF_PIPE_SYNC_BIT)) {
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
#if GEN_GEN >= 12
pipe.TileCacheFlushEnable = bits & ANV_PIPE_TILE_CACHE_FLUSH_BIT;
pipe.RenderTargetCacheFlushEnable =
bits & ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT;
+ /* GEN:BUG:1409600907: "PIPE_CONTROL with Depth Stall Enable bit must
+ * be set with any PIPE_CONTROL with Depth Flush Enable bit set.
+ */
+#if GEN_GEN >= 12
+ pipe.DepthStallEnable =
+ pipe.DepthCacheFlushEnable || (bits & ANV_PIPE_DEPTH_STALL_BIT);
+#else
pipe.DepthStallEnable = bits & ANV_PIPE_DEPTH_STALL_BIT;
+#endif
+
pipe.CommandStreamerStallEnable = bits & ANV_PIPE_CS_STALL_BIT;
pipe.StallAtPixelScoreboard = bits & ANV_PIPE_STALL_AT_SCOREBOARD_BIT;
+ /* From Sandybridge PRM, volume 2, "1.7.3.1 Writing a Value to Memory":
+ *
+ * "The most common action to perform upon reaching a
+ * synchronization point is to write a value out to memory. An
+ * immediate value (included with the synchronization command) may
+ * be written."
+ *
+ *
+ * From Broadwell PRM, volume 7, "End-of-Pipe Synchronization":
+ *
+ * "In case the data flushed out by the render engine is to be
+ * read back in to the render engine in coherent manner, then the
+ * render engine has to wait for the fence completion before
+ * accessing the flushed data. This can be achieved by following
+ * means on various products: PIPE_CONTROL command with CS Stall
+ * and the required write caches flushed with Post-Sync-Operation
+ * as Write Immediate Data.
+ *
+ * Example:
+ * - Workload-1 (3D/GPGPU/MEDIA)
+ * - PIPE_CONTROL (CS Stall, Post-Sync-Operation Write
+ * Immediate Data, Required Write Cache Flush bits set)
+ * - Workload-2 (Can use the data produce or output by
+ * Workload-1)
+ */
+ if (bits & ANV_PIPE_END_OF_PIPE_SYNC_BIT) {
+ pipe.CommandStreamerStallEnable = true;
+ pipe.PostSyncOperation = WriteImmediateData;
+ pipe.Address = (struct anv_address) {
+ .bo = cmd_buffer->device->workaround_bo,
+ .offset = 0
+ };
+ }
+
/*
* According to the Broadwell documentation, any PIPE_CONTROL with the
* "Command Streamer Stall" bit set must also have another bit set,
* I chose "Stall at Pixel Scoreboard" since that's what we use in
* mesa and it seems to work fine. The choice is fairly arbitrary.
*/
- if ((bits & ANV_PIPE_CS_STALL_BIT) &&
- !(bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_DEPTH_STALL_BIT |
- ANV_PIPE_STALL_AT_SCOREBOARD_BIT)))
+ if (pipe.CommandStreamerStallEnable &&
+ !pipe.RenderTargetCacheFlushEnable &&
+ !pipe.DepthCacheFlushEnable &&
+ !pipe.StallAtPixelScoreboard &&
+ !pipe.PostSyncOperation &&
+ !pipe.DepthStallEnable &&
+ !pipe.DCFlushEnable)
pipe.StallAtPixelScoreboard = true;
}
if (bits & ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT)
bits &= ~(ANV_PIPE_RENDER_TARGET_BUFFER_WRITES);
- bits &= ~(ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT);
+ if (GEN_IS_HASWELL) {
+ /* Haswell needs addition work-arounds:
+ *
+ * From Haswell PRM, volume 2, part 1, "End-of-Pipe Synchronization":
+ *
+ * Option 1:
+ * PIPE_CONTROL command with the CS Stall and the required write
+ * caches flushed with Post-SyncOperation as Write Immediate Data
+ * followed by eight dummy MI_STORE_DATA_IMM (write to scratch
+ * spce) commands.
+ *
+ * Example:
+ * - Workload-1
+ * - PIPE_CONTROL (CS Stall, Post-Sync-Operation Write
+ * Immediate Data, Required Write Cache Flush bits set)
+ * - MI_STORE_DATA_IMM (8 times) (Dummy data, Scratch Address)
+ * - Workload-2 (Can use the data produce or output by
+ * Workload-1)
+ *
+ * Unfortunately, both the PRMs and the internal docs are a bit
+ * out-of-date in this regard. What the windows driver does (and
+ * this appears to actually work) is to emit a register read from the
+ * memory address written by the pipe control above.
+ *
+ * What register we load into doesn't matter. We choose an indirect
+ * rendering register because we know it always exists and it's one
+ * of the first registers the command parser allows us to write. If
+ * you don't have command parser support in your kernel (pre-4.2),
+ * this will get turned into MI_NOOP and you won't get the
+ * workaround. Unfortunately, there's just not much we can do in
+ * that case. This register is perfectly safe to write since we
+ * always re-load all of the indirect draw registers right before
+ * 3DPRIMITIVE when needed anyway.
+ */
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
+ lrm.RegisterAddress = 0x243C; /* GEN7_3DPRIM_START_INSTANCE */
+ lrm.MemoryAddress = (struct anv_address) {
+ .bo = cmd_buffer->device->workaround_bo,
+ .offset = 0
+ };
+ }
+ }
+
+ bits &= ~(ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT |
+ ANV_PIPE_END_OF_PIPE_SYNC_BIT);
}
if (bits & ANV_PIPE_INVALIDATE_BITS) {
}
}
+#if GEN_GEN == 12
+ if ((bits & ANV_PIPE_AUX_TABLE_INVALIDATE_BIT) &&
+ cmd_buffer->device->info.has_aux_map) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
+ lri.RegisterOffset = GENX(GFX_CCS_AUX_INV_num);
+ lri.DataDWord = 1;
+ }
+ }
+#endif
+
bits &= ~ANV_PIPE_INVALIDATE_BITS;
}
if (range->aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
+ base_layer, layer_count,
pImageMemoryBarriers[i].oldLayout,
pImageMemoryBarriers[i].newLayout);
}
cmd_buffer_alloc_push_constants(struct anv_cmd_buffer *cmd_buffer)
{
VkShaderStageFlags stages =
- cmd_buffer->state.gfx.base.pipeline->active_stages;
+ cmd_buffer->state.gfx.pipeline->active_stages;
/* In order to avoid thrash, we assume that vertex and fragment stages
* always exist. In the rare case where one is missing *and* the other
static VkResult
emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
- gl_shader_stage stage,
+ struct anv_cmd_pipeline_state *pipe_state,
+ struct anv_shader_bin *shader,
struct anv_state *bt_state)
{
struct anv_subpass *subpass = cmd_buffer->state.subpass;
- struct anv_cmd_pipeline_state *pipe_state;
- struct anv_pipeline *pipeline;
uint32_t state_offset;
- switch (stage) {
- case MESA_SHADER_COMPUTE:
- pipe_state = &cmd_buffer->state.compute.base;
- break;
- default:
- pipe_state = &cmd_buffer->state.gfx.base;
- break;
- }
- pipeline = pipe_state->pipeline;
-
- if (!anv_pipeline_has_stage(pipeline, stage)) {
- *bt_state = (struct anv_state) { 0, };
- return VK_SUCCESS;
- }
-
- struct anv_pipeline_bind_map *map = &pipeline->shaders[stage]->bind_map;
+ struct anv_pipeline_bind_map *map = &shader->bind_map;
if (map->surface_count == 0) {
*bt_state = (struct anv_state) { 0, };
return VK_SUCCESS;
* softpin then we always keep all user-allocated memory objects resident.
*/
const bool need_client_mem_relocs =
- !cmd_buffer->device->instance->physicalDevice.use_softpin;
+ !cmd_buffer->device->physical->use_softpin;
for (uint32_t s = 0; s < map->surface_count; s++) {
struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
case ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS:
/* Color attachment binding */
- assert(stage == MESA_SHADER_FRAGMENT);
+ assert(shader->stage == MESA_SHADER_FRAGMENT);
if (binding->index < subpass->color_count) {
const unsigned att =
subpass->color_attachments[binding->index].attachment;
surface_state = cmd_buffer->state.null_surface_state;
}
+ assert(surface_state.map);
bt_map[s] = surface_state.offset + state_offset;
break;
anv_cmd_buffer_alloc_surface_state(cmd_buffer);
struct anv_address constant_data = {
- .bo = pipeline->device->dynamic_state_pool.block_pool.bo,
- .offset = pipeline->shaders[stage]->constant_data.offset,
+ .bo = cmd_buffer->device->dynamic_state_pool.block_pool.bo,
+ .offset = shader->constant_data.offset,
};
- unsigned constant_data_size =
- pipeline->shaders[stage]->constant_data_size;
+ unsigned constant_data_size = shader->constant_data_size;
const enum isl_format format =
anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER);
surface_state, format,
constant_data, constant_data_size, 1);
+ assert(surface_state.map);
bt_map[s] = surface_state.offset + state_offset;
add_surface_reloc(cmd_buffer, surface_state, constant_data);
break;
case ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS: {
/* This is always the first binding for compute shaders */
- assert(stage == MESA_SHADER_COMPUTE && s == 0);
+ assert(shader->stage == MESA_SHADER_COMPUTE && s == 0);
struct anv_state surface_state =
anv_cmd_buffer_alloc_surface_state(cmd_buffer);
format,
cmd_buffer->state.compute.num_workgroups,
12, 1);
+
+ assert(surface_state.map);
bt_map[s] = surface_state.offset + state_offset;
if (need_client_mem_relocs) {
add_surface_reloc(cmd_buffer, surface_state,
break;
}
case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
- assert(stage == MESA_SHADER_FRAGMENT);
+ assert(shader->stage == MESA_SHADER_FRAGMENT);
if ((desc->image_view->aspect_mask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) == 0) {
/* For depth and stencil input attachments, we treat it like any
* old texture that a user may have bound.
case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
/* Compute the offset within the buffer */
struct anv_push_constants *push =
- &cmd_buffer->state.push_constants[stage];
+ &cmd_buffer->state.push_constants[shader->stage];
uint32_t dynamic_offset =
push->dynamic_offsets[binding->dynamic_offset_index];
/* Clamp the range to the buffer size */
uint32_t range = MIN2(desc->range, desc->buffer->size - offset);
+ /* Align the range for consistency */
+ if (desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC)
+ range = align_u32(range, ANV_UBO_BOUNDS_CHECK_ALIGNMENT);
+
struct anv_address address =
anv_address_add(desc->buffer->address, offset);
assert(!"Invalid descriptor type");
continue;
}
+ assert(surface_state.map);
bt_map[s] = surface_state.offset + state_offset;
break;
}
static VkResult
emit_samplers(struct anv_cmd_buffer *cmd_buffer,
- gl_shader_stage stage,
+ struct anv_cmd_pipeline_state *pipe_state,
+ struct anv_shader_bin *shader,
struct anv_state *state)
{
- struct anv_cmd_pipeline_state *pipe_state =
- stage == MESA_SHADER_COMPUTE ? &cmd_buffer->state.compute.base :
- &cmd_buffer->state.gfx.base;
- struct anv_pipeline *pipeline = pipe_state->pipeline;
-
- if (!anv_pipeline_has_stage(pipeline, stage)) {
- *state = (struct anv_state) { 0, };
- return VK_SUCCESS;
- }
-
- struct anv_pipeline_bind_map *map = &pipeline->shaders[stage]->bind_map;
+ struct anv_pipeline_bind_map *map = &shader->bind_map;
if (map->sampler_count == 0) {
*state = (struct anv_state) { 0, };
return VK_SUCCESS;
}
static uint32_t
-flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
+flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_cmd_pipeline_state *pipe_state,
+ struct anv_shader_bin **shaders,
+ uint32_t num_shaders)
{
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
-
- VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty &
- pipeline->active_stages;
+ const VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty;
+ VkShaderStageFlags flushed = 0;
VkResult result = VK_SUCCESS;
- anv_foreach_stage(s, dirty) {
- result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
+ for (uint32_t i = 0; i < num_shaders; i++) {
+ if (!shaders[i])
+ continue;
+
+ gl_shader_stage stage = shaders[i]->stage;
+ VkShaderStageFlags vk_stage = mesa_to_vk_shader_stage(stage);
+ if ((vk_stage & dirty) == 0)
+ continue;
+
+ result = emit_samplers(cmd_buffer, pipe_state, shaders[i],
+ &cmd_buffer->state.samplers[stage]);
if (result != VK_SUCCESS)
break;
- result = emit_binding_table(cmd_buffer, s,
- &cmd_buffer->state.binding_tables[s]);
+ result = emit_binding_table(cmd_buffer, pipe_state, shaders[i],
+ &cmd_buffer->state.binding_tables[stage]);
if (result != VK_SUCCESS)
break;
+
+ flushed |= vk_stage;
}
if (result != VK_SUCCESS) {
genX(cmd_buffer_emit_state_base_address)(cmd_buffer);
/* Re-emit all active binding tables */
- dirty |= pipeline->active_stages;
- anv_foreach_stage(s, dirty) {
- result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
+ flushed = 0;
+
+ for (uint32_t i = 0; i < num_shaders; i++) {
+ if (!shaders[i])
+ continue;
+
+ gl_shader_stage stage = shaders[i]->stage;
+
+ result = emit_samplers(cmd_buffer, pipe_state, shaders[i],
+ &cmd_buffer->state.samplers[stage]);
if (result != VK_SUCCESS) {
anv_batch_set_error(&cmd_buffer->batch, result);
return 0;
}
- result = emit_binding_table(cmd_buffer, s,
- &cmd_buffer->state.binding_tables[s]);
+ result = emit_binding_table(cmd_buffer, pipe_state, shaders[i],
+ &cmd_buffer->state.binding_tables[stage]);
if (result != VK_SUCCESS) {
anv_batch_set_error(&cmd_buffer->batch, result);
return 0;
}
+
+ flushed |= mesa_to_vk_shader_stage(stage);
}
}
- cmd_buffer->state.descriptors_dirty &= ~dirty;
+ cmd_buffer->state.descriptors_dirty &= ~flushed;
- return dirty;
+ return flushed;
}
static void
}
}
-#if GEN_GEN >= 8 || GEN_IS_HASWELL
static struct anv_address
get_push_range_address(struct anv_cmd_buffer *cmd_buffer,
gl_shader_stage stage,
struct anv_descriptor_set *set =
gfx_state->base.descriptors[range->index];
return anv_descriptor_set_address(cmd_buffer, set);
- break;
}
case ANV_DESCRIPTOR_SET_PUSH_CONSTANTS: {
.bo = cmd_buffer->device->dynamic_state_pool.block_pool.bo,
.offset = state.offset,
};
- break;
}
default: {
&set->descriptors[range->index];
if (desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER) {
- return desc->buffer_view->address;
+ return desc->buffer_view->address;
+ } else {
+ assert(desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC);
+ struct anv_push_constants *push =
+ &cmd_buffer->state.push_constants[stage];
+ uint32_t dynamic_offset =
+ push->dynamic_offsets[range->dynamic_offset_index];
+ return anv_address_add(desc->buffer->address,
+ desc->offset + dynamic_offset);
+ }
+ }
+ }
+}
+
+
+/** Returns the size in bytes of the bound buffer
+ *
+ * The range is relative to the start of the buffer, not the start of the
+ * range. The returned range may be smaller than
+ *
+ * (range->start + range->length) * 32;
+ */
+static uint32_t
+get_push_range_bound_size(struct anv_cmd_buffer *cmd_buffer,
+ gl_shader_stage stage,
+ const struct anv_push_range *range)
+{
+ assert(stage != MESA_SHADER_COMPUTE);
+ const struct anv_cmd_graphics_state *gfx_state = &cmd_buffer->state.gfx;
+ switch (range->set) {
+ case ANV_DESCRIPTOR_SET_DESCRIPTORS: {
+ struct anv_descriptor_set *set =
+ gfx_state->base.descriptors[range->index];
+ assert(range->start * 32 < set->desc_mem.alloc_size);
+ assert((range->start + range->length) * 32 <= set->desc_mem.alloc_size);
+ return set->desc_mem.alloc_size;
+ }
+
+ case ANV_DESCRIPTOR_SET_PUSH_CONSTANTS:
+ return (range->start + range->length) * 32;
+
+ default: {
+ assert(range->set < MAX_SETS);
+ struct anv_descriptor_set *set =
+ gfx_state->base.descriptors[range->set];
+ const struct anv_descriptor *desc =
+ &set->descriptors[range->index];
+
+ if (desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER) {
+ return desc->buffer_view->range;
} else {
assert(desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC);
+ /* Compute the offset within the buffer */
struct anv_push_constants *push =
&cmd_buffer->state.push_constants[stage];
uint32_t dynamic_offset =
push->dynamic_offsets[range->dynamic_offset_index];
- return anv_address_add(desc->buffer->address,
- desc->offset + dynamic_offset);
+ uint64_t offset = desc->offset + dynamic_offset;
+ /* Clamp to the buffer size */
+ offset = MIN2(offset, desc->buffer->size);
+ /* Clamp the range to the buffer size */
+ uint32_t bound_range = MIN2(desc->range, desc->buffer->size - offset);
+
+ /* Align the range for consistency */
+ bound_range = align_u32(bound_range, ANV_UBO_BOUNDS_CHECK_ALIGNMENT);
+
+ return bound_range;
}
}
}
}
-#endif
static void
cmd_buffer_emit_push_constant(struct anv_cmd_buffer *cmd_buffer,
- gl_shader_stage stage, unsigned buffer_count)
+ gl_shader_stage stage,
+ struct anv_address *buffers,
+ unsigned buffer_count)
{
const struct anv_cmd_graphics_state *gfx_state = &cmd_buffer->state.gfx;
- const struct anv_pipeline *pipeline = gfx_state->base.pipeline;
+ const struct anv_graphics_pipeline *pipeline = gfx_state->pipeline;
static const uint32_t push_constant_opcodes[] = {
[MESA_SHADER_VERTEX] = 21,
const struct anv_pipeline_bind_map *bind_map =
&pipeline->shaders[stage]->bind_map;
+#if GEN_GEN >= 12
+ c.MOCS = cmd_buffer->device->isl_dev.mocs.internal;
+#endif
+
#if GEN_GEN >= 8 || GEN_IS_HASWELL
/* The Skylake PRM contains the following restriction:
*
*/
assert((GEN_GEN >= 8 || GEN_IS_HASWELL) || i == 0);
- const struct anv_address addr =
- get_push_range_address(cmd_buffer, stage, range);
c.ConstantBody.ReadLength[i + shift] = range->length;
c.ConstantBody.Buffer[i + shift] =
- anv_address_add(addr, range->start * 32);
+ anv_address_add(buffers[i], range->start * 32);
}
#else
/* For Ivy Bridge, push constants are relative to dynamic state
* base address and we only ever push actual push constants.
*/
if (bind_map->push_ranges[0].length > 0) {
+ assert(buffer_count == 1);
assert(bind_map->push_ranges[0].set ==
ANV_DESCRIPTOR_SET_PUSH_CONSTANTS);
- struct anv_state state =
- anv_cmd_buffer_push_constants(cmd_buffer, stage);
+ assert(buffers[0].bo ==
+ cmd_buffer->device->dynamic_state_pool.block_pool.bo);
c.ConstantBody.ReadLength[0] = bind_map->push_ranges[0].length;
c.ConstantBody.Buffer[0].bo = NULL;
- c.ConstantBody.Buffer[0].offset = state.offset;
+ c.ConstantBody.Buffer[0].offset = buffers[0].offset;
}
assert(bind_map->push_ranges[1].length == 0);
assert(bind_map->push_ranges[2].length == 0);
#if GEN_GEN >= 12
static void
cmd_buffer_emit_push_constant_all(struct anv_cmd_buffer *cmd_buffer,
- uint32_t shader_mask, uint32_t count)
+ uint32_t shader_mask,
+ struct anv_address *buffers,
+ uint32_t buffer_count)
{
- if (count == 0) {
+ if (buffer_count == 0) {
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CONSTANT_ALL), c) {
c.ShaderUpdateEnable = shader_mask;
+ c.MOCS = cmd_buffer->device->isl_dev.mocs.internal;
}
return;
}
const struct anv_cmd_graphics_state *gfx_state = &cmd_buffer->state.gfx;
- const struct anv_pipeline *pipeline = gfx_state->base.pipeline;
+ const struct anv_graphics_pipeline *pipeline = gfx_state->pipeline;
static const uint32_t push_constant_opcodes[] = {
[MESA_SHADER_VERTEX] = 21,
&pipeline->shaders[stage]->bind_map;
uint32_t *dw;
- const uint32_t buffers = (1 << count) - 1;
- const uint32_t num_dwords = 2 + 2 * count;
+ const uint32_t buffer_mask = (1 << buffer_count) - 1;
+ const uint32_t num_dwords = 2 + 2 * buffer_count;
dw = anv_batch_emitn(&cmd_buffer->batch, num_dwords,
GENX(3DSTATE_CONSTANT_ALL),
.ShaderUpdateEnable = shader_mask,
- .PointerBufferMask = buffers);
+ .PointerBufferMask = buffer_mask,
+ .MOCS = cmd_buffer->device->isl_dev.mocs.internal);
- for (int i = 0; i < count; i++) {
+ for (int i = 0; i < buffer_count; i++) {
const struct anv_push_range *range = &bind_map->push_ranges[i];
- const struct anv_address addr =
- get_push_range_address(cmd_buffer, stage, range);
-
GENX(3DSTATE_CONSTANT_ALL_DATA_pack)(
&cmd_buffer->batch, dw + 2 + i * 2,
&(struct GENX(3DSTATE_CONSTANT_ALL_DATA)) {
- .PointerToConstantBuffer = anv_address_add(addr, range->start * 32),
+ .PointerToConstantBuffer =
+ anv_address_add(buffers[i], range->start * 32),
.ConstantBufferReadLength = range->length,
});
}
{
VkShaderStageFlags flushed = 0;
const struct anv_cmd_graphics_state *gfx_state = &cmd_buffer->state.gfx;
- const struct anv_pipeline *pipeline = gfx_state->base.pipeline;
+ const struct anv_graphics_pipeline *pipeline = gfx_state->pipeline;
#if GEN_GEN >= 12
uint32_t nobuffer_stages = 0;
anv_foreach_stage(stage, dirty_stages) {
unsigned buffer_count = 0;
flushed |= mesa_to_vk_shader_stage(stage);
- uint32_t max_push_range = 0;
+ UNUSED uint32_t max_push_range = 0;
+ struct anv_address buffers[4] = {};
if (anv_pipeline_has_stage(pipeline, stage)) {
const struct anv_pipeline_bind_map *bind_map =
&pipeline->shaders[stage]->bind_map;
+ struct anv_push_constants *push =
+ &cmd_buffer->state.push_constants[stage];
+
+ if (cmd_buffer->device->robust_buffer_access) {
+ push->push_reg_mask = 0;
+ /* Start of the current range in the shader, relative to the start
+ * of push constants in the shader.
+ */
+ unsigned range_start_reg = 0;
+ for (unsigned i = 0; i < 4; i++) {
+ const struct anv_push_range *range = &bind_map->push_ranges[i];
+ if (range->length == 0)
+ continue;
+
+ unsigned bound_size =
+ get_push_range_bound_size(cmd_buffer, stage, range);
+ if (bound_size >= range->start * 32) {
+ unsigned bound_regs =
+ MIN2(DIV_ROUND_UP(bound_size, 32) - range->start,
+ range->length);
+ assert(range_start_reg + bound_regs <= 64);
+ push->push_reg_mask |= BITFIELD64_RANGE(range_start_reg,
+ bound_regs);
+ }
+
+ cmd_buffer->state.push_constants_dirty |=
+ mesa_to_vk_shader_stage(stage);
+
+ range_start_reg += range->length;
+ }
+ }
+ /* We have to gather buffer addresses as a second step because the
+ * loop above puts data into the push constant area and the call to
+ * get_push_range_address is what locks our push constants and copies
+ * them into the actual GPU buffer. If we did the two loops at the
+ * same time, we'd risk only having some of the sizes in the push
+ * constant buffer when we did the copy.
+ */
for (unsigned i = 0; i < 4; i++) {
const struct anv_push_range *range = &bind_map->push_ranges[i];
- if (range->length > 0) {
- buffer_count++;
- if (GEN_GEN >= 12 && range->length > max_push_range)
- max_push_range = range->length;
- }
+ if (range->length == 0)
+ break;
+
+ buffers[i] = get_push_range_address(cmd_buffer, stage, range);
+ max_push_range = MAX2(max_push_range, range->length);
+ buffer_count++;
}
+
+ /* We have at most 4 buffers but they should be tightly packed */
+ for (unsigned i = buffer_count; i < 4; i++)
+ assert(bind_map->push_ranges[i].length == 0);
}
#if GEN_GEN >= 12
*/
if (max_push_range < 32) {
cmd_buffer_emit_push_constant_all(cmd_buffer, 1 << stage,
- buffer_count);
+ buffers, buffer_count);
continue;
}
#endif
- cmd_buffer_emit_push_constant(cmd_buffer, stage, buffer_count);
+ cmd_buffer_emit_push_constant(cmd_buffer, stage, buffers, buffer_count);
}
#if GEN_GEN >= 12
if (nobuffer_stages)
- cmd_buffer_emit_push_constant_all(cmd_buffer, nobuffer_stages, 0);
+ cmd_buffer_emit_push_constant_all(cmd_buffer, nobuffer_stages, NULL, 0);
#endif
cmd_buffer->state.push_constants_dirty &= ~flushed;
}
-#if GEN_GEN >= 12
-void
-genX(cmd_buffer_aux_map_state)(struct anv_cmd_buffer *cmd_buffer)
-{
- void *aux_map_ctx = cmd_buffer->device->aux_map_ctx;
- if (!aux_map_ctx)
- return;
- uint32_t aux_map_state_num = gen_aux_map_get_state_num(aux_map_ctx);
- if (cmd_buffer->state.last_aux_map_state != aux_map_state_num) {
- /* If the aux-map state number increased, then we need to rewrite the
- * register. Rewriting the register is used to both set the aux-map
- * translation table address, and also to invalidate any previously
- * cached translations.
- */
- uint64_t base_addr = gen_aux_map_get_base(aux_map_ctx);
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
- lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num);
- lri.DataDWord = base_addr & 0xffffffff;
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
- lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num) + 4;
- lri.DataDWord = base_addr >> 32;
- }
- cmd_buffer->state.last_aux_map_state = aux_map_state_num;
- }
-}
-#endif
-
void
genX(cmd_buffer_flush_state)(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
uint32_t *p;
- uint32_t vb_emit = cmd_buffer->state.gfx.vb_dirty & pipeline->vb_used;
- if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE)
- vb_emit |= pipeline->vb_used;
-
assert((pipeline->active_stages & VK_SHADER_STAGE_COMPUTE_BIT) == 0);
- genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);
+ genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->base.l3_config);
genX(cmd_buffer_emit_hashing_mode)(cmd_buffer, UINT_MAX, UINT_MAX, 1);
genX(flush_pipeline_select_3d)(cmd_buffer);
-#if GEN_GEN >= 12
- genX(cmd_buffer_aux_map_state)(cmd_buffer);
-#endif
+ /* Apply any pending pipeline flushes we may have. We want to apply them
+ * now because, if any of those flushes are for things like push constants,
+ * the GPU will read the state at weird times.
+ */
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
+ uint32_t vb_emit = cmd_buffer->state.gfx.vb_dirty & pipeline->vb_used;
+ if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE)
+ vb_emit |= pipeline->vb_used;
if (vb_emit) {
const uint32_t num_buffers = __builtin_popcount(vb_emit);
#endif
if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE) {
- anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
+ anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->base.batch);
/* If the pipeline changed, we may need to re-allocate push constant
* space in the URB.
* 3DSTATE_BINDING_TABLE_POINTER_* for the push constants to take effect.
*/
uint32_t dirty = 0;
- if (cmd_buffer->state.descriptors_dirty)
- dirty = flush_descriptor_sets(cmd_buffer);
+ if (cmd_buffer->state.descriptors_dirty) {
+ dirty = flush_descriptor_sets(cmd_buffer,
+ &cmd_buffer->state.gfx.base,
+ pipeline->shaders,
+ ARRAY_SIZE(pipeline->shaders));
+ }
if (dirty || cmd_buffer->state.push_constants_dirty) {
/* Because we're pushing UBOs, we have to push whenever either
update_dirty_vbs_for_gen8_vb_flush(struct anv_cmd_buffer *cmd_buffer,
uint32_t access_type)
{
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
uint64_t vb_used = pipeline->vb_used;
uint32_t firstInstance)
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
/* Our implementation of VK_KHR_multiview uses instancing to draw the
* different views. We need to multiply instanceCount by the view count.
*/
- instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
+ if (!pipeline->use_primitive_replication)
+ instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
uint32_t firstInstance)
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
/* Our implementation of VK_KHR_multiview uses instancing to draw the
* different views. We need to multiply instanceCount by the view count.
*/
- instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
+ if (!pipeline->use_primitive_replication)
+ instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
#if GEN_IS_HASWELL || GEN_GEN >= 8
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_buffer, counter_buffer, counterBuffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
/* firstVertex is always zero for this draw function */
/* Our implementation of VK_KHR_multiview uses instancing to draw the
* different views. We need to multiply instanceCount by the view count.
*/
- instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
+ if (!pipeline->use_primitive_replication)
+ instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
struct gen_mi_builder b;
gen_mi_builder_init(&b, &cmd_buffer->batch);
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_buffer->state.gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
}
#endif
-void genX(CmdDrawIndirectCountKHR)(
+void genX(CmdDrawIndirectCount)(
VkCommandBuffer commandBuffer,
VkBuffer _buffer,
VkDeviceSize offset,
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
ANV_FROM_HANDLE(anv_buffer, count_buffer, _countBuffer);
struct anv_cmd_state *cmd_state = &cmd_buffer->state;
- struct anv_pipeline *pipeline = cmd_state->gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_state->gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
}
}
-void genX(CmdDrawIndexedIndirectCountKHR)(
+void genX(CmdDrawIndexedIndirectCount)(
VkCommandBuffer commandBuffer,
VkBuffer _buffer,
VkDeviceSize offset,
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
ANV_FROM_HANDLE(anv_buffer, count_buffer, _countBuffer);
struct anv_cmd_state *cmd_state = &cmd_buffer->state;
- struct anv_pipeline *pipeline = cmd_state->gfx.base.pipeline;
+ struct anv_graphics_pipeline *pipeline = cmd_state->gfx.pipeline;
const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
if (anv_batch_has_error(&cmd_buffer->batch))
cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_XFB_ENABLE;
}
-static VkResult
-flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
-{
- struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
- struct anv_state surfaces = { 0, }, samplers = { 0, };
- VkResult result;
-
- result = emit_binding_table(cmd_buffer, MESA_SHADER_COMPUTE, &surfaces);
- if (result != VK_SUCCESS) {
- assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
- result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
- if (result != VK_SUCCESS)
- return result;
-
- /* Re-emit state base addresses so we get the new surface state base
- * address before we start emitting binding tables etc.
- */
- genX(cmd_buffer_emit_state_base_address)(cmd_buffer);
-
- result = emit_binding_table(cmd_buffer, MESA_SHADER_COMPUTE, &surfaces);
- if (result != VK_SUCCESS) {
- anv_batch_set_error(&cmd_buffer->batch, result);
- return result;
- }
- }
-
- result = emit_samplers(cmd_buffer, MESA_SHADER_COMPUTE, &samplers);
- if (result != VK_SUCCESS) {
- anv_batch_set_error(&cmd_buffer->batch, result);
- return result;
- }
-
- uint32_t iface_desc_data_dw[GENX(INTERFACE_DESCRIPTOR_DATA_length)];
- struct GENX(INTERFACE_DESCRIPTOR_DATA) desc = {
- .BindingTablePointer = surfaces.offset,
- .SamplerStatePointer = samplers.offset,
- };
- GENX(INTERFACE_DESCRIPTOR_DATA_pack)(NULL, iface_desc_data_dw, &desc);
-
- struct anv_state state =
- anv_cmd_buffer_merge_dynamic(cmd_buffer, iface_desc_data_dw,
- pipeline->interface_descriptor_data,
- GENX(INTERFACE_DESCRIPTOR_DATA_length),
- 64);
-
- uint32_t size = GENX(INTERFACE_DESCRIPTOR_DATA_length) * sizeof(uint32_t);
- anv_batch_emit(&cmd_buffer->batch,
- GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD), mid) {
- mid.InterfaceDescriptorTotalLength = size;
- mid.InterfaceDescriptorDataStartAddress = state.offset;
- }
-
- return VK_SUCCESS;
-}
-
void
genX(cmd_buffer_flush_compute_state)(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
- VkResult result;
+ struct anv_compute_pipeline *pipeline = cmd_buffer->state.compute.pipeline;
- assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
+ assert(pipeline->cs);
- genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);
+ genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->base.l3_config);
genX(flush_pipeline_select_gpgpu)(cmd_buffer);
-#if GEN_GEN >= 12
- genX(cmd_buffer_aux_map_state)(cmd_buffer);
-#endif
+ /* Apply any pending pipeline flushes we may have. We want to apply them
+ * now because, if any of those flushes are for things like push constants,
+ * the GPU will read the state at weird times.
+ */
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
if (cmd_buffer->state.compute.pipeline_dirty) {
/* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
- anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
+ anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->base.batch);
/* The workgroup size of the pipeline affects our push constant layout
* so flag push constants as dirty if we change the pipeline.
if ((cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_COMPUTE_BIT) ||
cmd_buffer->state.compute.pipeline_dirty) {
- /* FIXME: figure out descriptors for gen7 */
- result = flush_compute_descriptor_set(cmd_buffer);
- if (result != VK_SUCCESS)
- return;
+ flush_descriptor_sets(cmd_buffer,
+ &cmd_buffer->state.compute.base,
+ &pipeline->cs, 1);
+
+ uint32_t iface_desc_data_dw[GENX(INTERFACE_DESCRIPTOR_DATA_length)];
+ struct GENX(INTERFACE_DESCRIPTOR_DATA) desc = {
+ .BindingTablePointer =
+ cmd_buffer->state.binding_tables[MESA_SHADER_COMPUTE].offset,
+ .SamplerStatePointer =
+ cmd_buffer->state.samplers[MESA_SHADER_COMPUTE].offset,
+ };
+ GENX(INTERFACE_DESCRIPTOR_DATA_pack)(NULL, iface_desc_data_dw, &desc);
+
+ struct anv_state state =
+ anv_cmd_buffer_merge_dynamic(cmd_buffer, iface_desc_data_dw,
+ pipeline->interface_descriptor_data,
+ GENX(INTERFACE_DESCRIPTOR_DATA_length),
+ 64);
- cmd_buffer->state.descriptors_dirty &= ~VK_SHADER_STAGE_COMPUTE_BIT;
+ uint32_t size = GENX(INTERFACE_DESCRIPTOR_DATA_length) * sizeof(uint32_t);
+ anv_batch_emit(&cmd_buffer->batch,
+ GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD), mid) {
+ mid.InterfaceDescriptorTotalLength = size;
+ mid.InterfaceDescriptorDataStartAddress = state.offset;
+ }
}
if (cmd_buffer->state.push_constants_dirty & VK_SHADER_STAGE_COMPUTE_BIT) {
int required_version,
const char *function)
{
- if (device->instance->physicalDevice.cmd_parser_version < required_version) {
- return vk_errorf(device->instance, device->instance,
+ if (device->physical->cmd_parser_version < required_version) {
+ return vk_errorf(device, device->physical,
VK_ERROR_FEATURE_NOT_PRESENT,
"cmd parser version %d is required for %s",
required_version, function);
uint32_t groupCountZ)
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
+ struct anv_compute_pipeline *pipeline = cmd_buffer->state.compute.pipeline;
const struct brw_cs_prog_data *prog_data = get_cs_prog_data(pipeline);
anv_cmd_buffer_push_base_group_id(cmd_buffer, baseGroupX,
ggw.SIMDSize = prog_data->simd_size / 16;
ggw.ThreadDepthCounterMaximum = 0;
ggw.ThreadHeightCounterMaximum = 0;
- ggw.ThreadWidthCounterMaximum = prog_data->threads - 1;
+ ggw.ThreadWidthCounterMaximum = anv_cs_threads(pipeline) - 1;
ggw.ThreadGroupIDXDimension = groupCountX;
ggw.ThreadGroupIDYDimension = groupCountY;
ggw.ThreadGroupIDZDimension = groupCountZ;
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
- struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
+ struct anv_compute_pipeline *pipeline = cmd_buffer->state.compute.pipeline;
const struct brw_cs_prog_data *prog_data = get_cs_prog_data(pipeline);
struct anv_address addr = anv_address_add(buffer->address, offset);
struct anv_batch *batch = &cmd_buffer->batch;
ggw.SIMDSize = prog_data->simd_size / 16;
ggw.ThreadDepthCounterMaximum = 0;
ggw.ThreadHeightCounterMaximum = 0;
- ggw.ThreadWidthCounterMaximum = prog_data->threads - 1;
+ ggw.ThreadWidthCounterMaximum = anv_cs_threads(pipeline) - 1;
ggw.RightExecutionMask = pipeline->cs_right_mask;
ggw.BottomExecutionMask = 0xffffffff;
}
* really know why.
*/
const uint32_t subslices =
- MAX2(cmd_buffer->device->instance->physicalDevice.subslice_total, 1);
+ MAX2(cmd_buffer->device->physical->subslice_total, 1);
anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_VFE_STATE), vfe) {
vfe.MaximumNumberofThreads =
devinfo->max_cs_threads * subslices - 1;
pc.CommandStreamerStallEnable = true;
#if GEN_GEN >= 12
pc.TileCacheFlushEnable = true;
+
+ /* GEN:BUG:1409600907: "PIPE_CONTROL with Depth Stall Enable bit must be
+ * set with any PIPE_CONTROL with Depth Flush Enable bit set.
+ */
+ pc.DepthStallEnable = true;
#endif
}
uint32_t vb_size)
{
if (GEN_GEN < 8 || GEN_GEN > 9 ||
- !cmd_buffer->device->instance->physicalDevice.use_softpin)
+ !cmd_buffer->device->physical->use_softpin)
return;
struct anv_vb_cache_range *bound, *dirty;
uint64_t vb_used)
{
if (GEN_GEN < 8 || GEN_GEN > 9 ||
- !cmd_buffer->device->instance->physicalDevice.use_softpin)
+ !cmd_buffer->device->physical->use_softpin)
return;
if (access_type == RANDOM) {
const uint32_t ds =
cmd_buffer->state.subpass->depth_stencil_attachment->attachment;
info.hiz_usage = cmd_buffer->state.attachments[ds].aux_usage;
- if (info.hiz_usage == ISL_AUX_USAGE_HIZ) {
+ if (info.hiz_usage != ISL_AUX_USAGE_NONE) {
+ assert(isl_aux_usage_has_hiz(info.hiz_usage));
info.hiz_surf = &image->planes[depth_plane].aux_surface.isl;
info.hiz_address =
isl_emit_depth_stencil_hiz_s(&device->isl_dev, dw, &info);
if (GEN_GEN >= 12) {
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
/* GEN:BUG:1408224581
*
* Workaround: Gen12LP Astep only An additional pipe control with
(struct anv_address) { cmd_buffer->device->workaround_bo, 0 };
}
}
- cmd_buffer->state.hiz_enabled = info.hiz_usage == ISL_AUX_USAGE_HIZ;
+ cmd_buffer->state.hiz_enabled = isl_aux_usage_has_hiz(info.hiz_usage);
}
/**
uint32_t subpass_id)
{
struct anv_cmd_state *cmd_state = &cmd_buffer->state;
- struct anv_subpass *subpass = &cmd_state->pass->subpasses[subpass_id];
+ struct anv_render_pass *pass = cmd_state->pass;
+ struct anv_subpass *subpass = &pass->subpasses[subpass_id];
cmd_state->subpass = subpass;
cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
* color or auxiliary buffer usage isn't supported by the sampler.
*/
const bool input_needs_resolve =
- (att_state->fast_clear && !att_state->clear_color_is_zero_one) ||
+ (att_state->fast_clear && !att_state->clear_color_is_zero) ||
att_state->input_aux_usage != att_state->aux_usage;
- VkImageLayout target_layout, target_stencil_layout;
+ VkImageLayout target_layout;
if (iview->aspect_mask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV &&
!input_needs_resolve) {
/* Layout transitions before the final only help to enable sampling
target_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
} else {
target_layout = subpass->attachments[i].layout;
- target_stencil_layout = subpass->attachments[i].stencil_layout;
}
+ VkImageLayout target_stencil_layout =
+ subpass->attachments[i].stencil_layout;
+
uint32_t base_layer, layer_count;
if (image->type == VK_IMAGE_TYPE_3D) {
base_layer = 0;
if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
+ base_layer, layer_count,
att_state->current_layout, target_layout);
att_state->aux_usage =
anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
- VK_IMAGE_ASPECT_DEPTH_BIT, target_layout);
+ VK_IMAGE_ASPECT_DEPTH_BIT,
+ VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
+ target_layout);
}
if (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
if (iview->image->samples == 1) {
anv_image_ccs_op(cmd_buffer, image,
iview->planes[0].isl.format,
+ iview->planes[0].isl.swizzle,
VK_IMAGE_ASPECT_COLOR_BIT,
0, 0, 1, ISL_AUX_OP_FAST_CLEAR,
&clear_color,
} else {
anv_image_mcs_op(cmd_buffer, image,
iview->planes[0].isl.format,
+ iview->planes[0].isl.swizzle,
VK_IMAGE_ASPECT_COLOR_BIT,
0, 1, ISL_AUX_OP_FAST_CLEAR,
&clear_color,
} else if (att_state->pending_clear_aspects & (VK_IMAGE_ASPECT_DEPTH_BIT |
VK_IMAGE_ASPECT_STENCIL_BIT)) {
if (att_state->fast_clear && !is_multiview) {
- /* We currently only support HiZ for single-layer images */
+ /* We currently only support HiZ for single-LOD images */
if (att_state->pending_clear_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
- assert(iview->image->planes[0].aux_usage == ISL_AUX_USAGE_HIZ);
+ assert(isl_aux_usage_has_hiz(iview->image->planes[0].aux_usage));
assert(iview->planes[0].isl.base_level == 0);
- assert(iview->planes[0].isl.base_array_layer == 0);
- assert(fb->layers == 1);
}
anv_image_hiz_clear(cmd_buffer, image,
assert(att_state->pending_clear_aspects == 0);
}
- if (GEN_GEN < 10 &&
- (att_state->pending_load_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) &&
- image->planes[0].aux_surface.isl.size_B > 0 &&
- iview->planes[0].isl.base_level == 0 &&
- iview->planes[0].isl.base_array_layer == 0) {
- if (att_state->aux_usage != ISL_AUX_USAGE_NONE) {
- genX(copy_fast_clear_dwords)(cmd_buffer, att_state->color.state,
- image, VK_IMAGE_ASPECT_COLOR_BIT,
- false /* copy to ss */);
- }
-
- if (need_input_attachment_state(&cmd_state->pass->attachments[a]) &&
- att_state->input_aux_usage != ISL_AUX_USAGE_NONE) {
- genX(copy_fast_clear_dwords)(cmd_buffer, att_state->input.state,
- image, VK_IMAGE_ASPECT_COLOR_BIT,
- false /* copy to ss */);
- }
- }
-
- if (subpass->attachments[i].usage ==
- VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
- /* We assume that if we're starting a subpass, we're going to do some
- * rendering so we may end up with compressed data.
- */
- genX(cmd_buffer_mark_image_written)(cmd_buffer, iview->image,
- VK_IMAGE_ASPECT_COLOR_BIT,
- att_state->aux_usage,
- iview->planes[0].isl.base_level,
- iview->planes[0].isl.base_array_layer,
- fb->layers);
- } else if (subpass->attachments[i].usage ==
- VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
- /* We may be writing depth or stencil so we need to mark the surface.
- * Unfortunately, there's no way to know at this point whether the
- * depth or stencil tests used will actually write to the surface.
- *
- * Even though stencil may be plane 1, it always shares a base_level
- * with depth.
- */
- const struct isl_view *ds_view = &iview->planes[0].isl;
- if (iview->aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) {
- genX(cmd_buffer_mark_image_written)(cmd_buffer, image,
- VK_IMAGE_ASPECT_DEPTH_BIT,
- att_state->aux_usage,
- ds_view->base_level,
- ds_view->base_array_layer,
- fb->layers);
- }
- if (iview->aspect_mask & VK_IMAGE_ASPECT_STENCIL_BIT) {
- /* Even though stencil may be plane 1, it always shares a
- * base_level with depth.
- */
- genX(cmd_buffer_mark_image_written)(cmd_buffer, image,
- VK_IMAGE_ASPECT_STENCIL_BIT,
- ISL_AUX_USAGE_NONE,
- ds_view->base_level,
- ds_view->base_array_layer,
- fb->layers);
- }
- }
-
/* If multiview is enabled, then we are only done clearing when we no
* longer have pending layers to clear, or when we have processed the
* last subpass that uses this attachment.
att_state->pending_load_aspects = 0;
}
- cmd_buffer_emit_depth_stencil(cmd_buffer);
+ /* We've transitioned all our images possibly fast clearing them. Now we
+ * can fill out the surface states that we will use as render targets
+ * during actual subpass rendering.
+ */
+ VkResult result = genX(cmd_buffer_alloc_att_surf_states)(cmd_buffer,
+ pass, subpass);
+ if (result != VK_SUCCESS)
+ return;
+
+ isl_null_fill_state(&cmd_buffer->device->isl_dev,
+ cmd_state->null_surface_state.map,
+ isl_extent3d(fb->width, fb->height, fb->layers));
+
+ for (uint32_t i = 0; i < subpass->attachment_count; ++i) {
+ const uint32_t att = subpass->attachments[i].attachment;
+ if (att == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ assert(att < cmd_state->pass->attachment_count);
+ struct anv_render_pass_attachment *pass_att = &pass->attachments[att];
+ struct anv_attachment_state *att_state = &cmd_state->attachments[att];
+ struct anv_image_view *iview = att_state->image_view;
+
+ if (!vk_format_is_color(pass_att->format))
+ continue;
+
+ const VkImageUsageFlagBits att_usage = subpass->attachments[i].usage;
+ assert(util_bitcount(att_usage) == 1);
+
+ struct anv_surface_state *surface_state;
+ isl_surf_usage_flags_t isl_surf_usage;
+ enum isl_aux_usage isl_aux_usage;
+ if (att_usage == VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+ surface_state = &att_state->color;
+ isl_surf_usage = ISL_SURF_USAGE_RENDER_TARGET_BIT;
+ isl_aux_usage = att_state->aux_usage;
+ } else if (att_usage == VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT) {
+ surface_state = &att_state->input;
+ isl_surf_usage = ISL_SURF_USAGE_TEXTURE_BIT;
+ isl_aux_usage = att_state->input_aux_usage;
+ } else {
+ continue;
+ }
+
+ /* We had better have a surface state when we get here */
+ assert(surface_state->state.map);
+
+ union isl_color_value clear_color = { .u32 = { 0, } };
+ if (pass_att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR &&
+ att_state->fast_clear)
+ anv_clear_color_from_att_state(&clear_color, att_state, iview);
+
+ anv_image_fill_surface_state(cmd_buffer->device,
+ iview->image,
+ VK_IMAGE_ASPECT_COLOR_BIT,
+ &iview->planes[0].isl,
+ isl_surf_usage,
+ isl_aux_usage,
+ &clear_color,
+ 0,
+ surface_state,
+ NULL);
+
+ add_surface_state_relocs(cmd_buffer, *surface_state);
+
+ if (GEN_GEN < 10 &&
+ pass_att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD &&
+ iview->image->planes[0].aux_usage != ISL_AUX_USAGE_NONE &&
+ iview->planes[0].isl.base_level == 0 &&
+ iview->planes[0].isl.base_array_layer == 0) {
+ genX(copy_fast_clear_dwords)(cmd_buffer, surface_state->state,
+ iview->image,
+ VK_IMAGE_ASPECT_COLOR_BIT,
+ false /* copy to ss */);
+ }
+ }
#if GEN_GEN >= 11
/* The PIPE_CONTROL command description says:
* is set due to new association of BTI, PS Scoreboard Stall bit must
* be set in this packet."
*/
- anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
- pc.RenderTargetCacheFlushEnable = true;
- pc.StallAtPixelScoreboard = true;
-#if GEN_GEN >= 12
- pc.TileCacheFlushEnable = true;
+ cmd_buffer->state.pending_pipe_bits |=
+ ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT |
+ ANV_PIPE_STALL_AT_SCOREBOARD_BIT;
#endif
- }
+
+#if GEN_GEN == 12
+ /* GEN:BUG:14010455700
+ *
+ * ISL will change some CHICKEN registers depending on the depth surface
+ * format, along with emitting the depth and stencil packets. In that case,
+ * we want to do a depth flush and stall, so the pipeline is not using these
+ * settings while we change the registers.
+ */
+ cmd_buffer->state.pending_pipe_bits |=
+ ANV_PIPE_DEPTH_CACHE_FLUSH_BIT |
+ ANV_PIPE_DEPTH_STALL_BIT |
+ ANV_PIPE_END_OF_PIPE_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
#endif
+
+ cmd_buffer_emit_depth_stencil(cmd_buffer);
}
static enum blorp_filter
uint32_t subpass_id = anv_get_subpass_id(&cmd_buffer->state);
struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
+ /* We are done with the previous subpass and all rendering directly to that
+ * subpass is now complete. Zero out all the surface states so we don't
+ * accidentally use them between now and the next subpass.
+ */
+ for (uint32_t i = 0; i < cmd_state->pass->attachment_count; ++i) {
+ memset(&cmd_state->attachments[i].color, 0,
+ sizeof(cmd_state->attachments[i].color));
+ memset(&cmd_state->attachments[i].input, 0,
+ sizeof(cmd_state->attachments[i].input));
+ }
+ cmd_state->null_surface_state = ANV_STATE_NULL;
+ cmd_state->attachment_states = ANV_STATE_NULL;
+
+ for (uint32_t i = 0; i < subpass->attachment_count; ++i) {
+ const uint32_t a = subpass->attachments[i].attachment;
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ assert(a < cmd_state->pass->attachment_count);
+ struct anv_attachment_state *att_state = &cmd_state->attachments[a];
+ struct anv_image_view *iview = att_state->image_view;
+
+ assert(util_bitcount(subpass->attachments[i].usage) == 1);
+ if (subpass->attachments[i].usage ==
+ VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+ /* We assume that if we're ending a subpass, we did do some rendering
+ * so we may end up with compressed data.
+ */
+ genX(cmd_buffer_mark_image_written)(cmd_buffer, iview->image,
+ VK_IMAGE_ASPECT_COLOR_BIT,
+ att_state->aux_usage,
+ iview->planes[0].isl.base_level,
+ iview->planes[0].isl.base_array_layer,
+ fb->layers);
+ } else if (subpass->attachments[i].usage ==
+ VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+ /* We may be writing depth or stencil so we need to mark the surface.
+ * Unfortunately, there's no way to know at this point whether the
+ * depth or stencil tests used will actually write to the surface.
+ *
+ * Even though stencil may be plane 1, it always shares a base_level
+ * with depth.
+ */
+ const struct isl_view *ds_view = &iview->planes[0].isl;
+ if (iview->aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) {
+ genX(cmd_buffer_mark_image_written)(cmd_buffer, iview->image,
+ VK_IMAGE_ASPECT_DEPTH_BIT,
+ att_state->aux_usage,
+ ds_view->base_level,
+ ds_view->base_array_layer,
+ fb->layers);
+ }
+ if (iview->aspect_mask & VK_IMAGE_ASPECT_STENCIL_BIT) {
+ /* Even though stencil may be plane 1, it always shares a
+ * base_level with depth.
+ */
+ genX(cmd_buffer_mark_image_written)(cmd_buffer, iview->image,
+ VK_IMAGE_ASPECT_STENCIL_BIT,
+ ISL_AUX_USAGE_NONE,
+ ds_view->base_level,
+ ds_view->base_array_layer,
+ fb->layers);
+ }
+ }
+ }
+
if (subpass->has_color_resolve) {
/* We are about to do some MSAA resolves. We need to flush so that the
* result of writes to the MSAA color attachments show up in the sampler
* able to handle.
*/
transition_depth_buffer(cmd_buffer, src_iview->image,
+ src_iview->planes[0].isl.base_array_layer,
+ fb->layers,
src_state->current_layout,
- VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+ VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
src_state->aux_usage =
anv_layout_to_aux_usage(&cmd_buffer->device->info, src_iview->image,
VK_IMAGE_ASPECT_DEPTH_BIT,
- VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
- src_state->current_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+ VK_IMAGE_USAGE_TRANSFER_SRC_BIT,
+ VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
+ src_state->current_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
/* MSAA resolves write to the resolve attachment as if it were any
* other transfer op. Transition the resolve attachment accordingly.
dst_initial_layout = VK_IMAGE_LAYOUT_UNDEFINED;
transition_depth_buffer(cmd_buffer, dst_iview->image,
+ dst_iview->planes[0].isl.base_array_layer,
+ fb->layers,
dst_initial_layout,
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
dst_state->aux_usage =
anv_layout_to_aux_usage(&cmd_buffer->device->info, dst_iview->image,
VK_IMAGE_ASPECT_DEPTH_BIT,
+ VK_IMAGE_USAGE_TRANSFER_DST_BIT,
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
dst_state->current_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
if ((src_iview->image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
subpass->stencil_resolve_mode != VK_RESOLVE_MODE_NONE_KHR) {
- src_state->current_stencil_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+ src_state->current_stencil_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
dst_state->current_stencil_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
enum isl_aux_usage src_aux_usage = ISL_AUX_USAGE_NONE;
struct anv_image_view *iview = cmd_state->attachments[a].image_view;
const struct anv_image *image = iview->image;
- if ((image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) &&
- image->vk_format != iview->vk_format) {
- enum anv_fast_clear_type fast_clear_type =
- anv_layout_to_fast_clear_type(&cmd_buffer->device->info,
- image, VK_IMAGE_ASPECT_COLOR_BIT,
- att_state->current_layout);
-
- /* If any clear color was used, flush it down the aux surfaces. If we
- * don't do it now using the view's format we might use the clear
- * color incorrectly in the following resolves (for example with an
- * SRGB view & a UNORM image).
- */
- if (fast_clear_type != ANV_FAST_CLEAR_NONE) {
- anv_perf_warn(cmd_buffer->device->instance, iview,
- "Doing a partial resolve to get rid of clear color at the "
- "end of a renderpass due to an image/view format mismatch");
-
- uint32_t base_layer, layer_count;
- if (image->type == VK_IMAGE_TYPE_3D) {
- base_layer = 0;
- layer_count = anv_minify(iview->image->extent.depth,
- iview->planes[0].isl.base_level);
- } else {
- base_layer = iview->planes[0].isl.base_array_layer;
- layer_count = fb->layers;
- }
-
- for (uint32_t a = 0; a < layer_count; a++) {
- uint32_t array_layer = base_layer + a;
- if (image->samples == 1) {
- anv_cmd_predicated_ccs_resolve(cmd_buffer, image,
- iview->planes[0].isl.format,
- VK_IMAGE_ASPECT_COLOR_BIT,
- iview->planes[0].isl.base_level,
- array_layer,
- ISL_AUX_OP_PARTIAL_RESOLVE,
- ANV_FAST_CLEAR_NONE);
- } else {
- anv_cmd_predicated_mcs_resolve(cmd_buffer, image,
- iview->planes[0].isl.format,
- VK_IMAGE_ASPECT_COLOR_BIT,
- base_layer,
- ISL_AUX_OP_PARTIAL_RESOLVE,
- ANV_FAST_CLEAR_NONE);
- }
- }
- }
- }
-
/* Transition the image into the final layout for this render pass */
VkImageLayout target_layout =
cmd_state->pass->attachments[a].final_layout;
if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
+ base_layer, layer_count,
att_state->current_layout, target_layout);
}
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_render_pass, pass, pRenderPassBegin->renderPass);
ANV_FROM_HANDLE(anv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
+ VkResult result;
cmd_buffer->state.framebuffer = framebuffer;
cmd_buffer->state.pass = pass;
cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
- VkResult result =
- genX(cmd_buffer_setup_attachments)(cmd_buffer, pass, pRenderPassBegin);
- /* If we failed to setup the attachments we should not try to go further */
+ result = genX(cmd_buffer_setup_attachments)(cmd_buffer, pass,
+ framebuffer,
+ pRenderPassBegin);
if (result != VK_SUCCESS) {
assert(anv_batch_has_error(&cmd_buffer->batch));
return;
cmd_buffer_begin_subpass(cmd_buffer, 0);
}
-void genX(CmdBeginRenderPass2KHR)(
+void genX(CmdBeginRenderPass2)(
VkCommandBuffer commandBuffer,
const VkRenderPassBeginInfo* pRenderPassBeginInfo,
const VkSubpassBeginInfoKHR* pSubpassBeginInfo)
cmd_buffer_begin_subpass(cmd_buffer, prev_subpass + 1);
}
-void genX(CmdNextSubpass2KHR)(
+void genX(CmdNextSubpass2)(
VkCommandBuffer commandBuffer,
const VkSubpassBeginInfoKHR* pSubpassBeginInfo,
const VkSubpassEndInfoKHR* pSubpassEndInfo)
cmd_buffer->state.subpass = NULL;
}
-void genX(CmdEndRenderPass2KHR)(
+void genX(CmdEndRenderPass2)(
VkCommandBuffer commandBuffer,
const VkSubpassEndInfoKHR* pSubpassEndInfo)
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_event, event, _event);
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
pc.StallAtPixelScoreboard = true;
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_event, event, _event);
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
pc.StallAtPixelScoreboard = true;
projects / mesa.git / blobdiff