#include <stdbool.h>
#include "anv_private.h"
+#include "vk_format_info.h"
#include "common/gen_l3_config.h"
#include "genxml/gen_macros.h"
sba.DynamicStateBaseAddress =
(struct anv_address) { &device->dynamic_state_block_pool.bo, 0 };
- sba.DynamicStateMemoryObjectControlState = GENX(MOCS),
- sba.DynamicStateBaseAddressModifyEnable = true,
+ sba.DynamicStateMemoryObjectControlState = GENX(MOCS);
+ sba.DynamicStateBaseAddressModifyEnable = true;
sba.IndirectObjectBaseAddress = (struct anv_address) { NULL, 0 };
sba.IndirectObjectMemoryObjectControlState = GENX(MOCS);
}
}
+static void
+add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_state state,
+ struct anv_bo *bo, uint32_t offset)
+{
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+
+ anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
+ state.offset + isl_dev->ss.addr_offset, bo, offset);
+}
+
+static void
+add_image_view_relocs(struct anv_cmd_buffer *cmd_buffer,
+ const struct anv_image_view *iview,
+ enum isl_aux_usage aux_usage,
+ struct anv_state state)
+{
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+
+ anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
+ state.offset + isl_dev->ss.addr_offset,
+ iview->bo, iview->offset);
+
+ if (aux_usage != ISL_AUX_USAGE_NONE) {
+ uint32_t aux_offset = iview->offset + iview->image->aux_surface.offset;
+
+ /* On gen7 and prior, the bottom 12 bits of the MCS base address are
+ * used to store other information. This should be ok, however, because
+ * surface buffer addresses are always 4K page alinged.
+ */
+ assert((aux_offset & 0xfff) == 0);
+ uint32_t *aux_addr_dw = state.map + isl_dev->ss.aux_addr_offset;
+ aux_offset += *aux_addr_dw & 0xfff;
+
+ anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
+ state.offset + isl_dev->ss.aux_addr_offset,
+ iview->bo, aux_offset);
+ }
+}
+
+static enum isl_aux_usage
+fb_attachment_get_aux_usage(struct anv_device *device,
+ struct anv_framebuffer *fb,
+ uint32_t attachment)
+{
+ struct anv_image_view *iview = fb->attachments[attachment];
+
+ if (iview->image->aux_surface.isl.size == 0)
+ return ISL_AUX_USAGE_NONE; /* No aux surface */
+
+ assert(iview->image->aux_surface.isl.usage & ISL_SURF_USAGE_CCS_BIT);
+
+ if (isl_format_supports_lossless_compression(&device->info,
+ iview->isl.format))
+ return ISL_AUX_USAGE_CCS_E;
+
+ return ISL_AUX_USAGE_NONE;
+}
+
+/**
+ * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
+ */
+static void
+genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_render_pass *pass,
+ struct anv_framebuffer *framebuffer,
+ const VkClearValue *clear_values)
+{
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+ struct anv_cmd_state *state = &cmd_buffer->state;
+
+ vk_free(&cmd_buffer->pool->alloc, state->attachments);
+
+ if (pass->attachment_count == 0) {
+ state->attachments = NULL;
+ return;
+ }
+
+ state->attachments = vk_alloc(&cmd_buffer->pool->alloc,
+ pass->attachment_count *
+ sizeof(state->attachments[0]),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+ if (state->attachments == NULL) {
+ /* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
+ abort();
+ }
+
+ bool need_null_state = false;
+ unsigned num_states = 0;
+ for (uint32_t i = 0; i < pass->attachment_count; ++i) {
+ if (vk_format_is_color(pass->attachments[i].format)) {
+ num_states++;
+ } else {
+ /* We need a null state for any depth-stencil-only subpasses.
+ * Importantly, this includes depth/stencil clears so we create one
+ * whenever we have depth or stencil
+ */
+ need_null_state = true;
+ }
+ }
+ num_states += need_null_state;
+
+ 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;
+
+ if (need_null_state) {
+ state->null_surface_state = next_state;
+ next_state.offset += ss_stride;
+ next_state.map += ss_stride;
+ }
+
+ for (uint32_t i = 0; i < pass->attachment_count; ++i) {
+ if (vk_format_is_color(pass->attachments[i].format)) {
+ state->attachments[i].color_rt_state = next_state;
+ next_state.offset += ss_stride;
+ next_state.map += ss_stride;
+ }
+ }
+ assert(next_state.offset == state->render_pass_states.offset +
+ state->render_pass_states.alloc_size);
+
+ if (framebuffer) {
+ assert(pass->attachment_count == framebuffer->attachment_count);
+
+ if (need_null_state) {
+ struct GENX(RENDER_SURFACE_STATE) null_ss = {
+ .SurfaceType = SURFTYPE_NULL,
+ .SurfaceArray = framebuffer->layers > 0,
+ .SurfaceFormat = ISL_FORMAT_R8G8B8A8_UNORM,
+#if GEN_GEN >= 8
+ .TileMode = YMAJOR,
+#else
+ .TiledSurface = true,
+#endif
+ .Width = framebuffer->width - 1,
+ .Height = framebuffer->height - 1,
+ .Depth = framebuffer->layers - 1,
+ .RenderTargetViewExtent = framebuffer->layers - 1,
+ };
+ GENX(RENDER_SURFACE_STATE_pack)(NULL, state->null_surface_state.map,
+ &null_ss);
+ }
+
+ 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);
+ VkImageAspectFlags clear_aspects = 0;
+
+ if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
+ /* color attachment */
+ if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
+ }
+ } else {
+ /* depthstencil attachment */
+ if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
+ att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
+ }
+ if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
+ att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+ clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
+ }
+ }
+
+ state->attachments[i].pending_clear_aspects = clear_aspects;
+ if (clear_aspects)
+ state->attachments[i].clear_value = clear_values[i];
+
+ struct anv_image_view *iview = framebuffer->attachments[i];
+ assert(iview->vk_format == att->format);
+
+ if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
+ state->attachments[i].aux_usage =
+ fb_attachment_get_aux_usage(cmd_buffer->device, framebuffer, i);
+
+ struct isl_view view = iview->isl;
+ view.usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT;
+ isl_surf_fill_state(isl_dev,
+ state->attachments[i].color_rt_state.map,
+ .surf = &iview->image->color_surface.isl,
+ .view = &view,
+ .aux_surf = &iview->image->aux_surface.isl,
+ .aux_usage = state->attachments[i].aux_usage,
+ .mocs = cmd_buffer->device->default_mocs);
+
+ add_image_view_relocs(cmd_buffer, iview,
+ state->attachments[i].aux_usage,
+ state->attachments[i].color_rt_state);
+ } else {
+ state->attachments[i].aux_usage = ISL_AUX_USAGE_NONE;
+ }
+ }
+
+ if (!cmd_buffer->device->info.has_llc)
+ anv_state_clflush(state->render_pass_states);
+ }
+}
+
+VkResult
+genX(BeginCommandBuffer)(
+ VkCommandBuffer commandBuffer,
+ const VkCommandBufferBeginInfo* pBeginInfo)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ /* If this is the first vkBeginCommandBuffer, we must *initialize* the
+ * command buffer's state. Otherwise, we must *reset* its state. In both
+ * cases we reset it.
+ *
+ * From the Vulkan 1.0 spec:
+ *
+ * If a command buffer is in the executable state and the command buffer
+ * was allocated from a command pool with the
+ * VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
+ * vkBeginCommandBuffer implicitly resets the command buffer, behaving
+ * as if vkResetCommandBuffer had been called with
+ * VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
+ * the command buffer in the recording state.
+ */
+ anv_cmd_buffer_reset(cmd_buffer);
+
+ cmd_buffer->usage_flags = pBeginInfo->flags;
+
+ assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY ||
+ !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT));
+
+ genX(cmd_buffer_emit_state_base_address)(cmd_buffer);
+
+ if (cmd_buffer->usage_flags &
+ VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
+ cmd_buffer->state.pass =
+ anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
+ cmd_buffer->state.subpass =
+ &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ cmd_buffer->state.framebuffer = NULL;
+
+ genX(cmd_buffer_setup_attachments)(cmd_buffer, cmd_buffer->state.pass,
+ NULL, NULL);
+
+ cmd_buffer->state.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
+ }
+
+ return VK_SUCCESS;
+}
+
+VkResult
+genX(EndCommandBuffer)(
+ VkCommandBuffer commandBuffer)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ anv_cmd_buffer_end_batch_buffer(cmd_buffer);
+
+ return VK_SUCCESS;
+}
+
+void
+genX(CmdExecuteCommands)(
+ VkCommandBuffer commandBuffer,
+ uint32_t commandBufferCount,
+ const VkCommandBuffer* pCmdBuffers)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);
+
+ assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
+
+ for (uint32_t i = 0; i < commandBufferCount; i++) {
+ ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
+
+ assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
+
+ if (secondary->usage_flags &
+ VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
+ /* If we're continuing a render pass from the primary, we need to
+ * copy the surface states for the current subpass into the storage
+ * we allocated for them in BeginCommandBuffer.
+ */
+ struct anv_bo *ss_bo = &primary->device->surface_state_block_pool.bo;
+ struct anv_state src_state = primary->state.render_pass_states;
+ struct anv_state dst_state = secondary->state.render_pass_states;
+ assert(src_state.alloc_size == dst_state.alloc_size);
+
+ genX(cmd_buffer_gpu_memcpy)(primary, ss_bo, dst_state.offset,
+ ss_bo, src_state.offset,
+ src_state.alloc_size);
+ }
+
+ anv_cmd_buffer_add_secondary(primary, secondary);
+ }
+
+ /* Each of the secondary command buffers will use its own state base
+ * address. We need to re-emit state base address for the primary after
+ * all of the secondaries are done.
+ *
+ * TODO: Maybe we want to make this a dirty bit to avoid extra state base
+ * address calls?
+ */
+ genX(cmd_buffer_emit_state_base_address)(primary);
+}
+
#define IVB_L3SQCREG1_SQGHPCI_DEFAULT 0x00730000
#define VLV_L3SQCREG1_SQGHPCI_DEFAULT 0x00d30000
#define HSW_L3SQCREG1_SQGHPCI_DEFAULT 0x00610000
pipe_bits |= ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT;
break;
case VK_ACCESS_SHADER_READ_BIT:
- case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT:
+ case VK_ACCESS_INPUT_ATTACHMENT_READ_BIT:
case VK_ACCESS_TRANSFER_READ_BIT:
pipe_bits |= ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT;
break;
cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_ALL_GRAPHICS;
}
+static VkResult
+emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
+ gl_shader_stage stage,
+ struct anv_state *bt_state)
+{
+ struct anv_subpass *subpass = cmd_buffer->state.subpass;
+ struct anv_pipeline *pipeline;
+ uint32_t bias, state_offset;
+
+ switch (stage) {
+ case MESA_SHADER_COMPUTE:
+ pipeline = cmd_buffer->state.compute_pipeline;
+ bias = 1;
+ break;
+ default:
+ pipeline = cmd_buffer->state.pipeline;
+ bias = 0;
+ break;
+ }
+
+ 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;
+ if (bias + map->surface_count == 0) {
+ *bt_state = (struct anv_state) { 0, };
+ return VK_SUCCESS;
+ }
+
+ *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
+ bias + map->surface_count,
+ &state_offset);
+ uint32_t *bt_map = bt_state->map;
+
+ if (bt_state->map == NULL)
+ return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+
+ if (stage == MESA_SHADER_COMPUTE &&
+ get_cs_prog_data(cmd_buffer->state.compute_pipeline)->uses_num_work_groups) {
+ struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo;
+ uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset;
+
+ struct anv_state surface_state;
+ surface_state =
+ anv_cmd_buffer_alloc_surface_state(cmd_buffer);
+
+ const enum isl_format format =
+ anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
+ anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
+ format, bo_offset, 12, 1);
+
+ bt_map[0] = surface_state.offset + state_offset;
+ add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
+ }
+
+ if (map->surface_count == 0)
+ goto out;
+
+ if (map->image_count > 0) {
+ VkResult result =
+ anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
+ if (result != VK_SUCCESS)
+ return result;
+
+ cmd_buffer->state.push_constants_dirty |= 1 << stage;
+ }
+
+ uint32_t image = 0;
+ for (uint32_t s = 0; s < map->surface_count; s++) {
+ struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
+
+ struct anv_state surface_state;
+
+ if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
+ /* Color attachment binding */
+ assert(stage == MESA_SHADER_FRAGMENT);
+ assert(binding->binding == 0);
+ if (binding->index < subpass->color_count) {
+ const unsigned att = subpass->color_attachments[binding->index];
+ surface_state = cmd_buffer->state.attachments[att].color_rt_state;
+ } else {
+ surface_state = cmd_buffer->state.null_surface_state;
+ }
+
+ bt_map[bias + s] = surface_state.offset + state_offset;
+ continue;
+ }
+
+ struct anv_descriptor_set *set =
+ cmd_buffer->state.descriptors[binding->set];
+ uint32_t offset = set->layout->binding[binding->binding].descriptor_index;
+ struct anv_descriptor *desc = &set->descriptors[offset + binding->index];
+
+ switch (desc->type) {
+ case VK_DESCRIPTOR_TYPE_SAMPLER:
+ /* Nothing for us to do here */
+ continue;
+
+ case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+ case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+ case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+ surface_state = desc->image_view->sampler_surface_state;
+ assert(surface_state.alloc_size);
+ add_image_view_relocs(cmd_buffer, desc->image_view,
+ desc->image_view->image->aux_usage,
+ surface_state);
+ break;
+
+ case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
+ surface_state = desc->image_view->storage_surface_state;
+ assert(surface_state.alloc_size);
+ add_image_view_relocs(cmd_buffer, desc->image_view,
+ desc->image_view->image->aux_usage,
+ surface_state);
+
+ struct brw_image_param *image_param =
+ &cmd_buffer->state.push_constants[stage]->images[image++];
+
+ *image_param = desc->image_view->storage_image_param;
+ image_param->surface_idx = bias + s;
+ break;
+ }
+
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+ case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+ surface_state = desc->buffer_view->surface_state;
+ assert(surface_state.alloc_size);
+ add_surface_state_reloc(cmd_buffer, surface_state,
+ desc->buffer_view->bo,
+ desc->buffer_view->offset);
+ break;
+
+ case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+ surface_state = desc->buffer_view->storage_surface_state;
+ assert(surface_state.alloc_size);
+ add_surface_state_reloc(cmd_buffer, surface_state,
+ desc->buffer_view->bo,
+ desc->buffer_view->offset);
+
+ struct brw_image_param *image_param =
+ &cmd_buffer->state.push_constants[stage]->images[image++];
+
+ *image_param = desc->buffer_view->storage_image_param;
+ image_param->surface_idx = bias + s;
+ break;
+
+ default:
+ assert(!"Invalid descriptor type");
+ continue;
+ }
+
+ bt_map[bias + s] = surface_state.offset + state_offset;
+ }
+ assert(image == map->image_count);
+
+ out:
+ if (!cmd_buffer->device->info.has_llc)
+ anv_state_clflush(*bt_state);
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+emit_samplers(struct anv_cmd_buffer *cmd_buffer,
+ gl_shader_stage stage,
+ struct anv_state *state)
+{
+ struct anv_pipeline *pipeline;
+
+ if (stage == MESA_SHADER_COMPUTE)
+ pipeline = cmd_buffer->state.compute_pipeline;
+ else
+ pipeline = cmd_buffer->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;
+ if (map->sampler_count == 0) {
+ *state = (struct anv_state) { 0, };
+ return VK_SUCCESS;
+ }
+
+ uint32_t size = map->sampler_count * 16;
+ *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
+
+ if (state->map == NULL)
+ return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+
+ for (uint32_t s = 0; s < map->sampler_count; s++) {
+ struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
+ struct anv_descriptor_set *set =
+ cmd_buffer->state.descriptors[binding->set];
+ uint32_t offset = set->layout->binding[binding->binding].descriptor_index;
+ struct anv_descriptor *desc = &set->descriptors[offset + binding->index];
+
+ if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
+ desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
+ continue;
+
+ struct anv_sampler *sampler = desc->sampler;
+
+ /* This can happen if we have an unfilled slot since TYPE_SAMPLER
+ * happens to be zero.
+ */
+ if (sampler == NULL)
+ continue;
+
+ memcpy(state->map + (s * 16),
+ sampler->state, sizeof(sampler->state));
+ }
+
+ if (!cmd_buffer->device->info.has_llc)
+ anv_state_clflush(*state);
+
+ return VK_SUCCESS;
+}
+
+static uint32_t
+flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
+{
+ VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty &
+ cmd_buffer->state.pipeline->active_stages;
+
+ VkResult result = VK_SUCCESS;
+ anv_foreach_stage(s, dirty) {
+ result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
+ if (result != VK_SUCCESS)
+ break;
+ result = emit_binding_table(cmd_buffer, s,
+ &cmd_buffer->state.binding_tables[s]);
+ if (result != VK_SUCCESS)
+ break;
+ }
+
+ if (result != VK_SUCCESS) {
+ assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);
+
+ result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
+ assert(result == VK_SUCCESS);
+
+ /* 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);
+
+ /* Re-emit all active binding tables */
+ dirty |= cmd_buffer->state.pipeline->active_stages;
+ anv_foreach_stage(s, dirty) {
+ result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
+ if (result != VK_SUCCESS)
+ return result;
+ result = emit_binding_table(cmd_buffer, s,
+ &cmd_buffer->state.binding_tables[s]);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ }
+
+ cmd_buffer->state.descriptors_dirty &= ~dirty;
+
+ return dirty;
+}
+
static void
cmd_buffer_emit_descriptor_pointers(struct anv_cmd_buffer *cmd_buffer,
uint32_t stages)
*/
uint32_t dirty = 0;
if (cmd_buffer->state.descriptors_dirty)
- dirty = anv_cmd_buffer_flush_descriptor_sets(cmd_buffer);
+ dirty = flush_descriptor_sets(cmd_buffer);
if (cmd_buffer->state.push_constants_dirty) {
#if GEN_GEN >= 9
}
}
+static VkResult
+flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
+{
+ struct anv_pipeline *pipeline = cmd_buffer->state.compute_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);
+ assert(result == VK_SUCCESS);
+
+ /* 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);
+ assert(result == VK_SUCCESS);
+ }
+
+ result = emit_samplers(cmd_buffer, MESA_SHADER_COMPUTE, &samplers);
+ assert(result == VK_SUCCESS);
+
+ 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_pipeline;
+ MAYBE_UNUSED VkResult result;
+
+ assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
+
+ genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);
+
+ genX(flush_pipeline_select_gpgpu)(cmd_buffer);
+
+ if (cmd_buffer->state.compute_dirty & ANV_CMD_DIRTY_PIPELINE) {
+ /* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
+ *
+ * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
+ * the only bits that are changed are scoreboard related: Scoreboard
+ * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard * Delta. For
+ * these scoreboard related states, a MEDIA_STATE_FLUSH is
+ * sufficient."
+ */
+ 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);
+ }
+
+ if ((cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_COMPUTE_BIT) ||
+ (cmd_buffer->state.compute_dirty & ANV_CMD_DIRTY_PIPELINE)) {
+ /* FIXME: figure out descriptors for gen7 */
+ result = flush_compute_descriptor_set(cmd_buffer);
+ assert(result == VK_SUCCESS);
+ cmd_buffer->state.descriptors_dirty &= ~VK_SHADER_STAGE_COMPUTE_BIT;
+ }
+
+ if (cmd_buffer->state.push_constants_dirty & VK_SHADER_STAGE_COMPUTE_BIT) {
+ struct anv_state push_state =
+ anv_cmd_buffer_cs_push_constants(cmd_buffer);
+
+ if (push_state.alloc_size) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_CURBE_LOAD), curbe) {
+ curbe.CURBETotalDataLength = push_state.alloc_size;
+ curbe.CURBEDataStartAddress = push_state.offset;
+ }
+ }
+ }
+
+ cmd_buffer->state.compute_dirty = 0;
+
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+}
+
#if GEN_GEN == 7
static bool
}
}
-struct anv_state
-genX(cmd_buffer_alloc_null_surface_state)(struct anv_cmd_buffer *cmd_buffer,
- struct anv_framebuffer *fb)
+void
+genX(cmd_buffer_emit_gen7_depth_flush)(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_state state =
- anv_state_stream_alloc(&cmd_buffer->surface_state_stream, 64, 64);
-
- struct GENX(RENDER_SURFACE_STATE) null_ss = {
- .SurfaceType = SURFTYPE_NULL,
- .SurfaceArray = fb->layers > 0,
- .SurfaceFormat = ISL_FORMAT_R8G8B8A8_UNORM,
-#if GEN_GEN >= 8
- .TileMode = YMAJOR,
-#else
- .TiledSurface = true,
-#endif
- .Width = fb->width - 1,
- .Height = fb->height - 1,
- .Depth = fb->layers - 1,
- .RenderTargetViewExtent = fb->layers - 1,
- };
-
- GENX(RENDER_SURFACE_STATE_pack)(NULL, state.map, &null_ss);
-
- if (!cmd_buffer->device->info.has_llc)
- anv_state_clflush(state);
+ if (GEN_GEN >= 8)
+ return;
- return state;
+ /* From the Haswell PRM, documentation for 3DSTATE_DEPTH_BUFFER:
+ *
+ * "Restriction: Prior to changing Depth/Stencil Buffer state (i.e., any
+ * combination of 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS,
+ * 3DSTATE_STENCIL_BUFFER, 3DSTATE_HIER_DEPTH_BUFFER) SW must first
+ * issue a pipelined depth stall (PIPE_CONTROL with Depth Stall bit
+ * set), followed by a pipelined depth cache flush (PIPE_CONTROL with
+ * Depth Flush Bit set, followed by another pipelined depth stall
+ * (PIPE_CONTROL with Depth Stall Bit set), unless SW can otherwise
+ * guarantee that the pipeline from WM onwards is already flushed (e.g.,
+ * via a preceding MI_FLUSH)."
+ */
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
+ pipe.DepthStallEnable = true;
+ }
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
+ pipe.DepthCacheFlushEnable = true;
+ }
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
+ pipe.DepthStallEnable = true;
+ }
}
static void
anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
const struct anv_image *image = iview ? iview->image : NULL;
const bool has_depth = image && (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT);
+ const bool has_hiz = image != NULL && anv_image_has_hiz(image);
const bool has_stencil =
image && (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT);
/* FIXME: Implement the PMA stall W/A */
/* FIXME: Width and Height are wrong */
+ genX(cmd_buffer_emit_gen7_depth_flush)(cmd_buffer);
+
/* Emit 3DSTATE_DEPTH_BUFFER */
if (has_depth) {
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_DEPTH_BUFFER), db) {
db.SurfaceType = SURFTYPE_2D;
db.DepthWriteEnable = true;
db.StencilWriteEnable = has_stencil;
- db.HierarchicalDepthBufferEnable = false;
+
+ if (cmd_buffer->state.pass->subpass_count == 1) {
+ db.HierarchicalDepthBufferEnable = has_hiz;
+ } else {
+ anv_finishme("Multiple-subpass HiZ not implemented");
+ }
db.SurfaceFormat = isl_surf_get_depth_format(&device->isl_dev,
&image->depth_surface.isl);
.bo = image->bo,
.offset = image->offset + image->depth_surface.offset,
};
- db.DepthBufferObjectControlState = GENX(MOCS),
+ db.DepthBufferObjectControlState = GENX(MOCS);
db.SurfacePitch = image->depth_surface.isl.row_pitch - 1;
db.Height = image->extent.height - 1;
db.Width = image->extent.width - 1;
- db.LOD = iview->base_mip;
+ db.LOD = iview->isl.base_level;
db.Depth = image->array_size - 1; /* FIXME: 3-D */
- db.MinimumArrayElement = iview->base_layer;
+ db.MinimumArrayElement = iview->isl.base_array_layer;
#if GEN_GEN >= 8
db.SurfaceQPitch =
- isl_surf_get_array_pitch_el_rows(&image->depth_surface.isl) >> 2,
+ isl_surf_get_array_pitch_el_rows(&image->depth_surface.isl) >> 2;
#endif
db.RenderTargetViewExtent = 1 - 1;
}
}
}
+ if (has_hiz) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_HIER_DEPTH_BUFFER), hdb) {
+ hdb.HierarchicalDepthBufferObjectControlState = GENX(MOCS);
+ hdb.SurfacePitch = image->aux_surface.isl.row_pitch - 1;
+ hdb.SurfaceBaseAddress = (struct anv_address) {
+ .bo = image->bo,
+ .offset = image->offset + image->aux_surface.offset,
+ };
+#if GEN_GEN >= 8
+ /* From the SKL PRM Vol2a:
+ *
+ * The interpretation of this field is dependent on Surface Type
+ * as follows:
+ * - SURFTYPE_1D: distance in pixels between array slices
+ * - SURFTYPE_2D/CUBE: distance in rows between array slices
+ * - SURFTYPE_3D: distance in rows between R - slices
+ */
+ hdb.SurfaceQPitch =
+ image->aux_surface.isl.dim == ISL_SURF_DIM_1D ?
+ isl_surf_get_array_pitch_el(&image->aux_surface.isl) >> 2 :
+ isl_surf_get_array_pitch_el_rows(&image->aux_surface.isl) >> 2;
+#endif
+ }
+ } else {
+ anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_HIER_DEPTH_BUFFER), hdb);
+ }
+
/* Emit 3DSTATE_STENCIL_BUFFER */
if (has_stencil) {
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_STENCIL_BUFFER), sb) {
#if GEN_GEN >= 8 || GEN_IS_HASWELL
- sb.StencilBufferEnable = true,
+ sb.StencilBufferEnable = true;
#endif
- sb.StencilBufferObjectControlState = GENX(MOCS),
+ sb.StencilBufferObjectControlState = GENX(MOCS);
- sb.SurfacePitch = image->stencil_surface.isl.row_pitch - 1,
+ sb.SurfacePitch = image->stencil_surface.isl.row_pitch - 1;
#if GEN_GEN >= 8
- sb.SurfaceQPitch = isl_surf_get_array_pitch_el_rows(&image->stencil_surface.isl) >> 2,
+ sb.SurfaceQPitch = isl_surf_get_array_pitch_el_rows(&image->stencil_surface.isl) >> 2;
#endif
sb.SurfaceBaseAddress = (struct anv_address) {
.bo = image->bo,
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_STENCIL_BUFFER), sb);
}
- /* Disable hierarchial depth buffers. */
- anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_HIER_DEPTH_BUFFER), hz);
-
- /* Clear the clear params. */
- anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CLEAR_PARAMS), cp);
+ /* From the IVB PRM Vol2P1, 11.5.5.4 3DSTATE_CLEAR_PARAMS:
+ *
+ * 3DSTATE_CLEAR_PARAMS must always be programmed in the along with
+ * the other Depth/Stencil state commands(i.e. 3DSTATE_DEPTH_BUFFER,
+ * 3DSTATE_STENCIL_BUFFER, or 3DSTATE_HIER_DEPTH_BUFFER)
+ *
+ * Testing also shows that some variant of this restriction may exist HSW+.
+ * On BDW+, it is not possible to emit 2 of these packets consecutively when
+ * both have DepthClearValueValid set. An analysis of such state programming
+ * on SKL showed that the GPU doesn't register the latter packet's clear
+ * value.
+ */
+ anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CLEAR_PARAMS), cp) {
+ if (has_hiz) {
+ cp.DepthClearValueValid = true;
+ const uint32_t ds =
+ cmd_buffer->state.subpass->depth_stencil_attachment;
+ cp.DepthClearValue =
+ cmd_buffer->state.attachments[ds].clear_value.depthStencil.depth;
+ }
+ }
}
-/**
- * @see anv_cmd_buffer_set_subpass()
- */
-void
+static void
genX(cmd_buffer_set_subpass)(struct anv_cmd_buffer *cmd_buffer,
struct anv_subpass *subpass)
{
cmd_buffer->state.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
cmd_buffer_emit_depth_stencil(cmd_buffer);
+ genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_HIZ_RESOLVE);
+ genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_DEPTH_CLEAR);
+
+ anv_cmd_buffer_clear_subpass(cmd_buffer);
}
void genX(CmdBeginRenderPass)(
cmd_buffer->state.framebuffer = framebuffer;
cmd_buffer->state.pass = pass;
cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
- anv_cmd_state_setup_attachments(cmd_buffer, pRenderPassBegin);
+ genX(cmd_buffer_setup_attachments)(cmd_buffer, pass, framebuffer,
+ pRenderPassBegin->pClearValues);
genX(flush_pipeline_select_3d)(cmd_buffer);
genX(cmd_buffer_set_subpass)(cmd_buffer, pass->subpasses);
- anv_cmd_buffer_clear_subpass(cmd_buffer);
}
void genX(CmdNextSubpass)(
anv_cmd_buffer_resolve_subpass(cmd_buffer);
genX(cmd_buffer_set_subpass)(cmd_buffer, cmd_buffer->state.subpass + 1);
- anv_cmd_buffer_clear_subpass(cmd_buffer);
}
void genX(CmdEndRenderPass)(
{
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_DEPTH_RESOLVE);
anv_cmd_buffer_resolve_subpass(cmd_buffer);
#ifndef NDEBUG
}
static void
-emit_ps_depth_count(struct anv_batch *batch,
+emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
struct anv_bo *bo, uint32_t offset)
{
- anv_batch_emit(batch, GENX(PIPE_CONTROL), pc) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
pc.DestinationAddressType = DAT_PPGTT;
pc.PostSyncOperation = WritePSDepthCount;
pc.DepthStallEnable = true;
pc.Address = (struct anv_address) { bo, offset };
+
+ if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
+ pc.CommandStreamerStallEnable = true;
}
}
static void
-emit_query_availability(struct anv_batch *batch,
+emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
struct anv_bo *bo, uint32_t offset)
{
- anv_batch_emit(batch, GENX(PIPE_CONTROL), pc) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
pc.DestinationAddressType = DAT_PPGTT;
pc.PostSyncOperation = WriteImmediateData;
pc.Address = (struct anv_address) { bo, offset };
switch (pool->type) {
case VK_QUERY_TYPE_OCCLUSION:
- emit_ps_depth_count(&cmd_buffer->batch, &pool->bo,
+ emit_ps_depth_count(cmd_buffer, &pool->bo,
query * sizeof(struct anv_query_pool_slot));
break;
switch (pool->type) {
case VK_QUERY_TYPE_OCCLUSION:
- emit_ps_depth_count(&cmd_buffer->batch, &pool->bo,
+ emit_ps_depth_count(cmd_buffer, &pool->bo,
query * sizeof(struct anv_query_pool_slot) + 8);
- emit_query_availability(&cmd_buffer->batch, &pool->bo,
+ emit_query_availability(cmd_buffer, &pool->bo,
query * sizeof(struct anv_query_pool_slot) + 16);
break;
default:
/* Everything else is bottom-of-pipe */
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
- pc.DestinationAddressType = DAT_PPGTT,
- pc.PostSyncOperation = WriteTimestamp,
+ pc.DestinationAddressType = DAT_PPGTT;
+ pc.PostSyncOperation = WriteTimestamp;
pc.Address = (struct anv_address) { &pool->bo, offset };
+
+ if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
+ pc.CommandStreamerStallEnable = true;
}
break;
}
- emit_query_availability(&cmd_buffer->batch, &pool->bo, query + 16);
+ emit_query_availability(cmd_buffer, &pool->bo, query + 16);
}
#if GEN_GEN > 7 || GEN_IS_HASWELL