#include "radv_meta.h"
#include "radv_private.h"
-#include "nir/nir_builder.h"
#include "sid.h"
-/**
- * Vertex attributes used by all pipelines.
- */
-struct vertex_attrs {
- float position[2]; /**< 3DPRIM_RECTLIST */
-};
-/* passthrough vertex shader */
+
static nir_shader *
-build_nir_vs(void)
+build_dcc_decompress_compute_shader(struct radv_device *dev)
{
- const struct glsl_type *vec4 = glsl_vec4_type();
-
nir_builder b;
- nir_variable *a_position;
- nir_variable *v_position;
+ const struct glsl_type *buf_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
+ false,
+ false,
+ GLSL_TYPE_FLOAT);
+ const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
+ false,
+ false,
+ GLSL_TYPE_FLOAT);
+ nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+ b.shader->info.name = ralloc_strdup(b.shader, "dcc_decompress_compute");
+
+ /* We need at least 16/16/1 to cover an entire DCC block in a single workgroup. */
+ b.shader->info.cs.local_size[0] = 16;
+ b.shader->info.cs.local_size[1] = 16;
+ b.shader->info.cs.local_size[2] = 1;
+ nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
+ buf_type, "s_tex");
+ input_img->data.descriptor_set = 0;
+ input_img->data.binding = 0;
+
+ nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
+ img_type, "out_img");
+ output_img->data.descriptor_set = 0;
+ output_img->data.binding = 1;
+
+ nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
+ nir_ssa_def *wg_id = nir_load_work_group_id(&b);
+ nir_ssa_def *block_size = nir_imm_ivec4(&b,
+ b.shader->info.cs.local_size[0],
+ b.shader->info.cs.local_size[1],
+ b.shader->info.cs.local_size[2], 0);
+
+ nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+ nir_ssa_def *input_img_deref = &nir_build_deref_var(&b, input_img)->dest.ssa;
+
+ nir_tex_instr *tex = nir_tex_instr_create(b.shader, 3);
+ tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
+ tex->op = nir_texop_txf;
+ tex->src[0].src_type = nir_tex_src_coord;
+ tex->src[0].src = nir_src_for_ssa(nir_channels(&b, global_id, 3));
+ tex->src[1].src_type = nir_tex_src_lod;
+ tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
+ tex->src[2].src_type = nir_tex_src_texture_deref;
+ tex->src[2].src = nir_src_for_ssa(input_img_deref);
+ tex->dest_type = nir_type_float;
+ tex->is_array = false;
+ tex->coord_components = 2;
+
+ nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
+ nir_builder_instr_insert(&b, &tex->instr);
+
+ nir_intrinsic_instr *membar = nir_intrinsic_instr_create(b.shader, nir_intrinsic_memory_barrier);
+ nir_builder_instr_insert(&b, &membar->instr);
+
+ nir_intrinsic_instr *bar = nir_intrinsic_instr_create(b.shader, nir_intrinsic_barrier);
+ nir_builder_instr_insert(&b, &bar->instr);
+
+ nir_ssa_def *outval = &tex->dest.ssa;
+ nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_deref_store);
+ store->num_components = 4;
+ store->src[0] = nir_src_for_ssa(&nir_build_deref_var(&b, output_img)->dest.ssa);
+ store->src[1] = nir_src_for_ssa(global_id);
+ store->src[2] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
+ store->src[3] = nir_src_for_ssa(outval);
+
+ nir_builder_instr_insert(&b, &store->instr);
+ return b.shader;
+}
+
+static VkResult
+create_dcc_compress_compute(struct radv_device *device)
+{
+ VkResult result = VK_SUCCESS;
+ struct radv_shader_module cs = { .nir = NULL };
- nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
- b.shader->info->name = ralloc_strdup(b.shader, "meta_fast_clear_vs");
+ cs.nir = build_dcc_decompress_compute_shader(device);
- a_position = nir_variable_create(b.shader, nir_var_shader_in, vec4,
- "a_position");
- a_position->data.location = VERT_ATTRIB_GENERIC0;
+ VkDescriptorSetLayoutCreateInfo ds_create_info = {
+ .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+ .flags = VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR,
+ .bindingCount = 2,
+ .pBindings = (VkDescriptorSetLayoutBinding[]) {
+ {
+ .binding = 0,
+ .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+ .descriptorCount = 1,
+ .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+ .pImmutableSamplers = NULL
+ },
+ {
+ .binding = 1,
+ .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+ .descriptorCount = 1,
+ .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+ .pImmutableSamplers = NULL
+ },
+ }
+ };
- v_position = nir_variable_create(b.shader, nir_var_shader_out, vec4,
- "gl_Position");
- v_position->data.location = VARYING_SLOT_POS;
+ result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
+ &ds_create_info,
+ &device->meta_state.alloc,
+ &device->meta_state.fast_clear_flush.dcc_decompress_compute_ds_layout);
+ if (result != VK_SUCCESS)
+ goto cleanup;
- nir_copy_var(&b, v_position, a_position);
- return b.shader;
-}
+ VkPipelineLayoutCreateInfo pl_create_info = {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+ .setLayoutCount = 1,
+ .pSetLayouts = &device->meta_state.fast_clear_flush.dcc_decompress_compute_ds_layout,
+ .pushConstantRangeCount = 1,
+ .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 8},
+ };
-/* simple passthrough shader */
-static nir_shader *
-build_nir_fs(void)
-{
- nir_builder b;
+ result = radv_CreatePipelineLayout(radv_device_to_handle(device),
+ &pl_create_info,
+ &device->meta_state.alloc,
+ &device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout);
+ if (result != VK_SUCCESS)
+ goto cleanup;
- nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
- b.shader->info->name = ralloc_asprintf(b.shader,
- "meta_fast_clear_noop_fs");
+ /* compute shader */
- return b.shader;
+ VkPipelineShaderStageCreateInfo pipeline_shader_stage = {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+ .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+ .module = radv_shader_module_to_handle(&cs),
+ .pName = "main",
+ .pSpecializationInfo = NULL,
+ };
+
+ VkComputePipelineCreateInfo vk_pipeline_info = {
+ .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+ .stage = pipeline_shader_stage,
+ .flags = 0,
+ .layout = device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout,
+ };
+
+ result = radv_CreateComputePipelines(radv_device_to_handle(device),
+ radv_pipeline_cache_to_handle(&device->meta_state.cache),
+ 1, &vk_pipeline_info, NULL,
+ &device->meta_state.fast_clear_flush.dcc_decompress_compute_pipeline);
+ if (result != VK_SUCCESS)
+ goto cleanup;
+
+cleanup:
+ ralloc_free(cs.nir);
+ return result;
}
static VkResult
attachment.samples = 1;
attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
- attachment.initialLayout = VK_IMAGE_LAYOUT_GENERAL;
- attachment.finalLayout = VK_IMAGE_LAYOUT_GENERAL;
+ attachment.initialLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+ attachment.finalLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
result = radv_CreateRenderPass(device_h,
&(VkRenderPassCreateInfo) {
.pColorAttachments = (VkAttachmentReference[]) {
{
.attachment = 0,
- .layout = VK_IMAGE_LAYOUT_GENERAL,
+ .layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
},
},
.pResolveAttachments = NULL,
return result;
}
+static VkResult
+create_pipeline_layout(struct radv_device *device, VkPipelineLayout *layout)
+{
+ VkPipelineLayoutCreateInfo pl_create_info = {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+ .setLayoutCount = 0,
+ .pSetLayouts = NULL,
+ .pushConstantRangeCount = 0,
+ .pPushConstantRanges = NULL,
+ };
+
+ return radv_CreatePipelineLayout(radv_device_to_handle(device),
+ &pl_create_info,
+ &device->meta_state.alloc,
+ layout);
+}
+
static VkResult
create_pipeline(struct radv_device *device,
- VkShaderModule vs_module_h)
+ VkShaderModule vs_module_h,
+ VkPipelineLayout layout)
{
VkResult result;
VkDevice device_h = radv_device_to_handle(device);
struct radv_shader_module fs_module = {
- .nir = build_nir_fs(),
+ .nir = radv_meta_build_nir_fs_noop(),
};
if (!fs_module.nir) {
const VkPipelineVertexInputStateCreateInfo vi_state = {
.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
- .vertexBindingDescriptionCount = 1,
- .pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
- {
- .binding = 0,
- .stride = sizeof(struct vertex_attrs),
- .inputRate = VK_VERTEX_INPUT_RATE_VERTEX
- },
- },
- .vertexAttributeDescriptionCount = 1,
- .pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
- {
- /* Position */
- .location = 0,
- .binding = 0,
- .format = VK_FORMAT_R32G32_SFLOAT,
- .offset = offsetof(struct vertex_attrs, position),
- },
- }
+ .vertexBindingDescriptionCount = 0,
+ .vertexAttributeDescriptionCount = 0,
};
const VkPipelineInputAssemblyStateCreateInfo ia_state = {
.pViewportState = &(VkPipelineViewportStateCreateInfo) {
.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
- .viewportCount = 0,
- .scissorCount = 0,
+ .viewportCount = 1,
+ .scissorCount = 1,
},
.pRasterizationState = &rs_state,
.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
.alphaToOneEnable = false,
},
.pColorBlendState = &blend_state,
- .pDynamicState = NULL,
+ .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+ .dynamicStateCount = 2,
+ .pDynamicStates = (VkDynamicState[]) {
+ VK_DYNAMIC_STATE_VIEWPORT,
+ VK_DYNAMIC_STATE_SCISSOR,
+ },
+ },
+ .layout = layout,
.renderPass = device->meta_state.fast_clear_flush.pass,
.subpass = 0,
},
.pViewportState = &(VkPipelineViewportStateCreateInfo) {
.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
- .viewportCount = 0,
- .scissorCount = 0,
+ .viewportCount = 1,
+ .scissorCount = 1,
},
.pRasterizationState = &rs_state,
.pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
.alphaToOneEnable = false,
},
.pColorBlendState = &blend_state,
- .pDynamicState = NULL,
+ .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+ .dynamicStateCount = 2,
+ .pDynamicStates = (VkDynamicState[]) {
+ VK_DYNAMIC_STATE_VIEWPORT,
+ VK_DYNAMIC_STATE_SCISSOR,
+ },
+ },
+ .layout = layout,
.renderPass = device->meta_state.fast_clear_flush.pass,
.subpass = 0,
},
&device->meta_state.alloc,
&device->meta_state.fast_clear_flush.fmask_decompress_pipeline);
if (result != VK_SUCCESS)
- goto cleanup_cmask;
+ goto cleanup;
+
+ result = radv_graphics_pipeline_create(device_h,
+ radv_pipeline_cache_to_handle(&device->meta_state.cache),
+ &(VkGraphicsPipelineCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+ .stageCount = 2,
+ .pStages = stages,
+
+ .pVertexInputState = &vi_state,
+ .pInputAssemblyState = &ia_state,
+
+ .pViewportState = &(VkPipelineViewportStateCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+ .viewportCount = 1,
+ .scissorCount = 1,
+ },
+ .pRasterizationState = &rs_state,
+ .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+ .rasterizationSamples = 1,
+ .sampleShadingEnable = false,
+ .pSampleMask = NULL,
+ .alphaToCoverageEnable = false,
+ .alphaToOneEnable = false,
+ },
+ .pColorBlendState = &blend_state,
+ .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+ .dynamicStateCount = 2,
+ .pDynamicStates = (VkDynamicState[]) {
+ VK_DYNAMIC_STATE_VIEWPORT,
+ VK_DYNAMIC_STATE_SCISSOR,
+ },
+ },
+ .layout = layout,
+ .renderPass = device->meta_state.fast_clear_flush.pass,
+ .subpass = 0,
+ },
+ &(struct radv_graphics_pipeline_create_info) {
+ .use_rectlist = true,
+ .custom_blend_mode = V_028808_CB_DCC_DECOMPRESS,
+ },
+ &device->meta_state.alloc,
+ &device->meta_state.fast_clear_flush.dcc_decompress_pipeline);
+ if (result != VK_SUCCESS)
+ goto cleanup;
goto cleanup;
-cleanup_cmask:
- radv_DestroyPipeline(device_h, device->meta_state.fast_clear_flush.cmask_eliminate_pipeline, &device->meta_state.alloc);
+
cleanup:
ralloc_free(fs_module.nir);
return result;
radv_device_finish_meta_fast_clear_flush_state(struct radv_device *device)
{
struct radv_meta_state *state = &device->meta_state;
- VkDevice device_h = radv_device_to_handle(device);
- VkRenderPass pass_h = device->meta_state.fast_clear_flush.pass;
- const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
-
- if (pass_h)
- radv_DestroyRenderPass(device_h, pass_h,
- &device->meta_state.alloc);
- VkPipeline pipeline_h = state->fast_clear_flush.cmask_eliminate_pipeline;
- if (pipeline_h) {
- radv_DestroyPipeline(device_h, pipeline_h, alloc);
- }
-
- pipeline_h = state->fast_clear_flush.fmask_decompress_pipeline;
- if (pipeline_h) {
- radv_DestroyPipeline(device_h, pipeline_h, alloc);
- }
+ radv_DestroyPipeline(radv_device_to_handle(device),
+ state->fast_clear_flush.dcc_decompress_pipeline,
+ &state->alloc);
+ radv_DestroyPipeline(radv_device_to_handle(device),
+ state->fast_clear_flush.fmask_decompress_pipeline,
+ &state->alloc);
+ radv_DestroyPipeline(radv_device_to_handle(device),
+ state->fast_clear_flush.cmask_eliminate_pipeline,
+ &state->alloc);
+ radv_DestroyRenderPass(radv_device_to_handle(device),
+ state->fast_clear_flush.pass, &state->alloc);
+ radv_DestroyPipelineLayout(radv_device_to_handle(device),
+ state->fast_clear_flush.p_layout,
+ &state->alloc);
+
+ radv_DestroyPipeline(radv_device_to_handle(device),
+ state->fast_clear_flush.dcc_decompress_compute_pipeline,
+ &state->alloc);
+ radv_DestroyPipelineLayout(radv_device_to_handle(device),
+ state->fast_clear_flush.dcc_decompress_compute_p_layout,
+ &state->alloc);
+ radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
+ state->fast_clear_flush.dcc_decompress_compute_ds_layout,
+ &state->alloc);
}
-VkResult
-radv_device_init_meta_fast_clear_flush_state(struct radv_device *device)
+static VkResult
+radv_device_init_meta_fast_clear_flush_state_internal(struct radv_device *device)
{
VkResult res = VK_SUCCESS;
- zero(device->meta_state.fast_clear_flush);
+ mtx_lock(&device->meta_state.mtx);
+ if (device->meta_state.fast_clear_flush.cmask_eliminate_pipeline) {
+ mtx_unlock(&device->meta_state.mtx);
+ return VK_SUCCESS;
+ }
- struct radv_shader_module vs_module = { .nir = build_nir_vs() };
+ struct radv_shader_module vs_module = { .nir = radv_meta_build_nir_vs_generate_vertices() };
if (!vs_module.nir) {
/* XXX: Need more accurate error */
res = VK_ERROR_OUT_OF_HOST_MEMORY;
if (res != VK_SUCCESS)
goto fail;
+ res = create_pipeline_layout(device,
+ &device->meta_state.fast_clear_flush.p_layout);
+ if (res != VK_SUCCESS)
+ goto fail;
+
VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
- res = create_pipeline(device, vs_module_h);
+ res = create_pipeline(device, vs_module_h,
+ device->meta_state.fast_clear_flush.p_layout);
+ if (res != VK_SUCCESS)
+ goto fail;
+
+ res = create_dcc_compress_compute(device);
if (res != VK_SUCCESS)
goto fail;
cleanup:
ralloc_free(vs_module.nir);
+ mtx_unlock(&device->meta_state.mtx);
return res;
}
+
+VkResult
+radv_device_init_meta_fast_clear_flush_state(struct radv_device *device, bool on_demand)
+{
+ if (on_demand)
+ return VK_SUCCESS;
+
+ return radv_device_init_meta_fast_clear_flush_state_internal(device);
+}
+
+static void
+radv_emit_set_predication_state_from_image(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ uint64_t pred_offset, bool value)
+{
+ uint64_t va = 0;
+
+ if (value) {
+ va = radv_buffer_get_va(image->bo) + image->offset;
+ va += pred_offset;
+ }
+
+ si_emit_set_predication_state(cmd_buffer, true, va);
+}
+
static void
-emit_fast_clear_flush(struct radv_cmd_buffer *cmd_buffer,
- const VkExtent2D *resolve_extent,
- bool fmask_decompress)
+radv_process_color_image_layer(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *range,
+ int level, int layer)
{
struct radv_device *device = cmd_buffer->device;
- VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
- uint32_t offset;
- const struct vertex_attrs vertex_data[3] = {
- {
- .position = {
- 0,
- 0,
- },
- },
- {
- .position = {
- 0,
- resolve_extent->height,
- },
- },
- {
- .position = {
- resolve_extent->width,
- 0,
- },
- },
- };
+ struct radv_image_view iview;
+ uint32_t width, height;
- cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
- RADV_CMD_FLAG_FLUSH_AND_INV_CB_META);
- radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset);
- struct radv_buffer vertex_buffer = {
- .device = device,
- .size = sizeof(vertex_data),
- .bo = cmd_buffer->upload.upload_bo,
- .offset = offset,
- };
+ width = radv_minify(image->info.width, range->baseMipLevel + level);
+ height = radv_minify(image->info.height, range->baseMipLevel + level);
- VkBuffer vertex_buffer_h = radv_buffer_to_handle(&vertex_buffer);
+ radv_image_view_init(&iview, device,
+ &(VkImageViewCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+ .image = radv_image_to_handle(image),
+ .viewType = radv_meta_get_view_type(image),
+ .format = image->vk_format,
+ .subresourceRange = {
+ .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+ .baseMipLevel = range->baseMipLevel + level,
+ .levelCount = 1,
+ .baseArrayLayer = range->baseArrayLayer + layer,
+ .layerCount = 1,
+ },
+ }, NULL);
- radv_CmdBindVertexBuffers(cmd_buffer_h,
- /*firstBinding*/ 0,
- /*bindingCount*/ 1,
- (VkBuffer[]) { vertex_buffer_h },
- (VkDeviceSize[]) { 0 });
+ VkFramebuffer fb_h;
+ radv_CreateFramebuffer(radv_device_to_handle(device),
+ &(VkFramebufferCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+ .attachmentCount = 1,
+ .pAttachments = (VkImageView[]) {
+ radv_image_view_to_handle(&iview)
+ },
+ .width = width,
+ .height = height,
+ .layers = 1
+ }, &cmd_buffer->pool->alloc, &fb_h);
+
+ radv_CmdBeginRenderPass(radv_cmd_buffer_to_handle(cmd_buffer),
+ &(VkRenderPassBeginInfo) {
+ .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+ .renderPass = device->meta_state.fast_clear_flush.pass,
+ .framebuffer = fb_h,
+ .renderArea = {
+ .offset = {
+ 0,
+ 0,
+ },
+ .extent = {
+ width,
+ height,
+ }
+ },
+ .clearValueCount = 0,
+ .pClearValues = NULL,
+ }, VK_SUBPASS_CONTENTS_INLINE);
+
+ radv_CmdDraw(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
+
+ cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+ RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
+
+ radv_CmdEndRenderPass(radv_cmd_buffer_to_handle(cmd_buffer));
+
+ radv_DestroyFramebuffer(radv_device_to_handle(device), fb_h,
+ &cmd_buffer->pool->alloc);
+}
- VkPipeline pipeline_h;
- if (fmask_decompress)
- pipeline_h = device->meta_state.fast_clear_flush.fmask_decompress_pipeline;
- else
- pipeline_h = device->meta_state.fast_clear_flush.cmask_eliminate_pipeline;
- RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h);
+static void
+radv_process_color_image(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange,
+ bool decompress_dcc)
+{
+ struct radv_meta_saved_state saved_state;
+ VkPipeline *pipeline;
+
+ if (decompress_dcc && radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
+ pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.dcc_decompress_pipeline;
+ } else if (radv_image_has_fmask(image) && !image->tc_compatible_cmask) {
+ pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.fmask_decompress_pipeline;
+ } else {
+ pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.cmask_eliminate_pipeline;
+ }
- if (cmd_buffer->state.pipeline != pipeline) {
- radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
- pipeline_h);
+ if (!*pipeline) {
+ VkResult ret;
+
+ ret = radv_device_init_meta_fast_clear_flush_state_internal(cmd_buffer->device);
+ if (ret != VK_SUCCESS) {
+ cmd_buffer->record_result = ret;
+ return;
+ }
}
- radv_CmdDraw(cmd_buffer_h, 3, 1, 0, 0);
- cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
- RADV_CMD_FLAG_FLUSH_AND_INV_CB_META);
- si_emit_cache_flush(cmd_buffer);
+ radv_meta_save(&saved_state, cmd_buffer,
+ RADV_META_SAVE_GRAPHICS_PIPELINE |
+ RADV_META_SAVE_PASS);
+
+ radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+ VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline);
+
+ for (uint32_t l = 0; l < radv_get_levelCount(image, subresourceRange); ++l) {
+ uint32_t width, height;
+
+ /* Do not decompress levels without DCC. */
+ if (decompress_dcc &&
+ !radv_dcc_enabled(image, subresourceRange->baseMipLevel + l))
+ continue;
+
+ width = radv_minify(image->info.width,
+ subresourceRange->baseMipLevel + l);
+ height = radv_minify(image->info.height,
+ subresourceRange->baseMipLevel + l);
+
+ radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
+ &(VkViewport) {
+ .x = 0,
+ .y = 0,
+ .width = width,
+ .height = height,
+ .minDepth = 0.0f,
+ .maxDepth = 1.0f
+ });
+
+ radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
+ &(VkRect2D) {
+ .offset = { 0, 0 },
+ .extent = { width, height },
+ });
+
+ for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
+ radv_process_color_image_layer(cmd_buffer, image,
+ subresourceRange, l, s);
+ }
+ }
+
+ radv_meta_restore(&saved_state, cmd_buffer);
+}
+
+static void
+radv_emit_color_decompress(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange,
+ bool decompress_dcc)
+{
+ bool old_predicating = false;
+
+ assert(cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL);
+
+ if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
+ uint64_t pred_offset = decompress_dcc ? image->dcc_pred_offset :
+ image->fce_pred_offset;
+ pred_offset += 8 * subresourceRange->baseMipLevel;
+
+ old_predicating = cmd_buffer->state.predicating;
+
+ radv_emit_set_predication_state_from_image(cmd_buffer, image, pred_offset, true);
+ cmd_buffer->state.predicating = true;
+ }
+
+ radv_process_color_image(cmd_buffer, image, subresourceRange,
+ decompress_dcc);
+
+ if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
+ uint64_t pred_offset = decompress_dcc ? image->dcc_pred_offset :
+ image->fce_pred_offset;
+ pred_offset += 8 * subresourceRange->baseMipLevel;
+
+ cmd_buffer->state.predicating = old_predicating;
+
+ radv_emit_set_predication_state_from_image(cmd_buffer, image, pred_offset, false);
+
+ if (cmd_buffer->state.predication_type != -1) {
+ /* Restore previous conditional rendering user state. */
+ si_emit_set_predication_state(cmd_buffer,
+ cmd_buffer->state.predication_type,
+ cmd_buffer->state.predication_va);
+ }
+ }
+
+ if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
+ /* Clear the image's fast-clear eliminate predicate because
+ * FMASK and DCC also imply a fast-clear eliminate.
+ */
+ radv_update_fce_metadata(cmd_buffer, image, subresourceRange, false);
+
+ /* Mark the image as being decompressed. */
+ if (decompress_dcc)
+ radv_update_dcc_metadata(cmd_buffer, image, subresourceRange, false);
+ }
}
-/**
- */
void
radv_fast_clear_flush_image_inplace(struct radv_cmd_buffer *cmd_buffer,
- struct radv_image *image)
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange)
+{
+ radv_emit_color_decompress(cmd_buffer, image, subresourceRange, false);
+}
+
+static void
+radv_decompress_dcc_gfx(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange)
+{
+ radv_emit_color_decompress(cmd_buffer, image, subresourceRange, true);
+}
+
+static void
+radv_decompress_dcc_compute(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange)
{
struct radv_meta_saved_state saved_state;
- struct radv_meta_saved_pass_state saved_pass_state;
- VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
- VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
+ struct radv_image_view load_iview = {0};
+ struct radv_image_view store_iview = {0};
+ struct radv_device *device = cmd_buffer->device;
- radv_meta_save_pass(&saved_pass_state, cmd_buffer);
- radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
+ /* This assumes the image is 2d with 1 layer */
+ struct radv_cmd_state *state = &cmd_buffer->state;
- struct radv_image_view iview;
- radv_image_view_init(&iview, cmd_buffer->device,
- &(VkImageViewCreateInfo) {
- .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
- .image = radv_image_to_handle(image),
- .format = image->vk_format,
- .subresourceRange = {
- .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
- .baseMipLevel = 0,
- .levelCount = 1,
- .baseArrayLayer = 0,
- .layerCount = 1,
- },
- },
- cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
+ state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+ RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
- VkFramebuffer fb_h;
- radv_CreateFramebuffer(device_h,
- &(VkFramebufferCreateInfo) {
- .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
- .attachmentCount = 1,
- .pAttachments = (VkImageView[]) {
- radv_image_view_to_handle(&iview)
- },
- .width = image->extent.width,
- .height = image->extent.height,
- .layers = 1
- },
- &cmd_buffer->pool->alloc,
- &fb_h);
-
- radv_CmdBeginRenderPass(cmd_buffer_h,
- &(VkRenderPassBeginInfo) {
- .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
- .renderPass = cmd_buffer->device->meta_state.fast_clear_flush.pass,
- .framebuffer = fb_h,
- .renderArea = {
- .offset = {
- 0,
- 0,
- },
- .extent = {
- image->extent.width,
- image->extent.height,
- }
- },
- .clearValueCount = 0,
- .pClearValues = NULL,
- },
- VK_SUBPASS_CONTENTS_INLINE);
-
- emit_fast_clear_flush(cmd_buffer,
- &(VkExtent2D) { image->extent.width, image->extent.height },
- image->fmask.size > 0);
- radv_CmdEndRenderPass(cmd_buffer_h);
-
- radv_DestroyFramebuffer(device_h, fb_h,
- &cmd_buffer->pool->alloc);
+ if (!cmd_buffer->device->meta_state.fast_clear_flush.cmask_eliminate_pipeline) {
+ VkResult ret = radv_device_init_meta_fast_clear_flush_state_internal(cmd_buffer->device);
+ if (ret != VK_SUCCESS) {
+ cmd_buffer->record_result = ret;
+ return;
+ }
+ }
+ radv_meta_save(&saved_state, cmd_buffer, RADV_META_SAVE_DESCRIPTORS |
+ RADV_META_SAVE_COMPUTE_PIPELINE);
+
+ radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
+ VK_PIPELINE_BIND_POINT_COMPUTE,
+ device->meta_state.fast_clear_flush.dcc_decompress_compute_pipeline);
+
+ for (uint32_t l = 0; l < radv_get_levelCount(image, subresourceRange); l++) {
+ uint32_t width, height;
+
+ /* Do not decompress levels without DCC. */
+ if (!radv_dcc_enabled(image, subresourceRange->baseMipLevel + l))
+ continue;
+
+ width = radv_minify(image->info.width,
+ subresourceRange->baseMipLevel + l);
+ height = radv_minify(image->info.height,
+ subresourceRange->baseMipLevel + l);
+
+ for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
+ radv_image_view_init(&load_iview, cmd_buffer->device,
+ &(VkImageViewCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+ .image = radv_image_to_handle(image),
+ .viewType = VK_IMAGE_VIEW_TYPE_2D,
+ .format = image->vk_format,
+ .subresourceRange = {
+ .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+ .baseMipLevel = subresourceRange->baseMipLevel + l,
+ .levelCount = 1,
+ .baseArrayLayer = subresourceRange->baseArrayLayer + s,
+ .layerCount = 1
+ },
+ }, NULL);
+ radv_image_view_init(&store_iview, cmd_buffer->device,
+ &(VkImageViewCreateInfo) {
+ .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+ .image = radv_image_to_handle(image),
+ .viewType = VK_IMAGE_VIEW_TYPE_2D,
+ .format = image->vk_format,
+ .subresourceRange = {
+ .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+ .baseMipLevel = subresourceRange->baseMipLevel + l,
+ .levelCount = 1,
+ .baseArrayLayer = subresourceRange->baseArrayLayer + s,
+ .layerCount = 1
+ },
+ }, &(struct radv_image_view_extra_create_info) {
+ .disable_compression = true
+ });
+
+ radv_meta_push_descriptor_set(cmd_buffer,
+ VK_PIPELINE_BIND_POINT_COMPUTE,
+ device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout,
+ 0, /* set */
+ 2, /* descriptorWriteCount */
+ (VkWriteDescriptorSet[]) {
+ {
+ .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+ .dstBinding = 0,
+ .dstArrayElement = 0,
+ .descriptorCount = 1,
+ .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
+ .pImageInfo = (VkDescriptorImageInfo[]) {
+ {
+ .sampler = VK_NULL_HANDLE,
+ .imageView = radv_image_view_to_handle(&load_iview),
+ .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+ },
+ }
+ },
+ {
+ .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+ .dstBinding = 1,
+ .dstArrayElement = 0,
+ .descriptorCount = 1,
+ .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+ .pImageInfo = (VkDescriptorImageInfo[]) {
+ {
+ .sampler = VK_NULL_HANDLE,
+ .imageView = radv_image_view_to_handle(&store_iview),
+ .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
+ },
+ }
+ }
+ });
+
+ radv_unaligned_dispatch(cmd_buffer, width, height, 1);
+ }
+ }
+
+ /* Mark this image as actually being decompressed. */
+ radv_update_dcc_metadata(cmd_buffer, image, subresourceRange, false);
+
+ /* The fill buffer below does its own saving */
radv_meta_restore(&saved_state, cmd_buffer);
- radv_meta_restore_pass(&saved_pass_state, cmd_buffer);
+
+ state->flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+ RADV_CMD_FLAG_INV_VCACHE;
+
+
+ /* Initialize the DCC metadata as "fully expanded". */
+ radv_initialize_dcc(cmd_buffer, image, subresourceRange, 0xffffffff);
+}
+
+void
+radv_decompress_dcc(struct radv_cmd_buffer *cmd_buffer,
+ struct radv_image *image,
+ const VkImageSubresourceRange *subresourceRange)
+{
+ if (cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL)
+ radv_decompress_dcc_gfx(cmd_buffer, image, subresourceRange);
+ else
+ radv_decompress_dcc_compute(cmd_buffer, image, subresourceRange);
}