src/amd/vulkan/radv_meta_fast_clear.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdbool.h>
  26
  27 #include "radv_meta.h"
  28 #include "radv_private.h"
  29 #include "sid.h"
  30
  31
  32 static nir_shader *
  33 build_dcc_decompress_compute_shader(struct radv_device *dev)
  34 {
  35         nir_builder b;
  36         const struct glsl_type *buf_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
  37                                                              false,
  38                                                              false,
  39                                                              GLSL_TYPE_FLOAT);
  40         const struct glsl_type *img_type = glsl_image_type(GLSL_SAMPLER_DIM_2D,
  41                                                            false,
  42                                                            GLSL_TYPE_FLOAT);
  43         nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
  44         b.shader->info.name = ralloc_strdup(b.shader, "dcc_decompress_compute");
  45
  46         /* We need at least 16/16/1 to cover an entire DCC block in a single workgroup. */
  47         b.shader->info.cs.local_size[0] = 16;
  48         b.shader->info.cs.local_size[1] = 16;
  49         b.shader->info.cs.local_size[2] = 1;
  50         nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
  51                                                       buf_type, "s_tex");
  52         input_img->data.descriptor_set = 0;
  53         input_img->data.binding = 0;
  54
  55         nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
  56                                                        img_type, "out_img");
  57         output_img->data.descriptor_set = 0;
  58         output_img->data.binding = 1;
  59
  60         nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
  61         nir_ssa_def *wg_id = nir_load_work_group_id(&b);
  62         nir_ssa_def *block_size = nir_imm_ivec4(&b,
  63                                                 b.shader->info.cs.local_size[0],
  64                                                 b.shader->info.cs.local_size[1],
  65                                                 b.shader->info.cs.local_size[2], 0);
  66
  67         nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
  68         nir_ssa_def *input_img_deref = &nir_build_deref_var(&b, input_img)->dest.ssa;
  69
  70         nir_tex_instr *tex = nir_tex_instr_create(b.shader, 3);
  71         tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
  72         tex->op = nir_texop_txf;
  73         tex->src[0].src_type = nir_tex_src_coord;
  74         tex->src[0].src = nir_src_for_ssa(nir_channels(&b, global_id, 3));
  75         tex->src[1].src_type = nir_tex_src_lod;
  76         tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
  77         tex->src[2].src_type = nir_tex_src_texture_deref;
  78         tex->src[2].src = nir_src_for_ssa(input_img_deref);
  79         tex->dest_type = nir_type_float;
  80         tex->is_array = false;
  81         tex->coord_components = 2;
  82
  83         nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
  84         nir_builder_instr_insert(&b, &tex->instr);
  85
  86         nir_intrinsic_instr *membar = nir_intrinsic_instr_create(b.shader, nir_intrinsic_memory_barrier);
  87         nir_builder_instr_insert(&b, &membar->instr);
  88
  89         nir_intrinsic_instr *bar = nir_intrinsic_instr_create(b.shader, nir_intrinsic_control_barrier);
  90         nir_builder_instr_insert(&b, &bar->instr);
  91
  92         nir_ssa_def *outval = &tex->dest.ssa;
  93         nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_deref_store);
  94         store->num_components = 4;
  95         store->src[0] = nir_src_for_ssa(&nir_build_deref_var(&b, output_img)->dest.ssa);
  96         store->src[1] = nir_src_for_ssa(global_id);
  97         store->src[2] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
  98         store->src[3] = nir_src_for_ssa(outval);
  99         store->src[4] = nir_src_for_ssa(nir_imm_int(&b, 0));
 100
 101         nir_builder_instr_insert(&b, &store->instr);
 102         return b.shader;
 103 }
 104
 105 static VkResult
 106 create_dcc_compress_compute(struct radv_device *device)
 107 {
 108         VkResult result = VK_SUCCESS;
 109         struct radv_shader_module cs = { .nir = NULL };
 110
 111         cs.nir = build_dcc_decompress_compute_shader(device);
 112
 113         VkDescriptorSetLayoutCreateInfo ds_create_info = {
 114                 .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
 115                 .flags = VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR,
 116                 .bindingCount = 2,
 117                 .pBindings = (VkDescriptorSetLayoutBinding[]) {
 118                         {
 119                                 .binding = 0,
 120                                 .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
 121                                 .descriptorCount = 1,
 122                                 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
 123                                 .pImmutableSamplers = NULL
 124                         },
 125                         {
 126                                 .binding = 1,
 127                                 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
 128                                 .descriptorCount = 1,
 129                                 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
 130                                 .pImmutableSamplers = NULL
 131                         },
 132                 }
 133         };
 134
 135         result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
 136                                                 &ds_create_info,
 137                                                 &device->meta_state.alloc,
 138                                                 &device->meta_state.fast_clear_flush.dcc_decompress_compute_ds_layout);
 139         if (result != VK_SUCCESS)
 140                 goto cleanup;
 141
 142
 143         VkPipelineLayoutCreateInfo pl_create_info = {
 144                 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
 145                 .setLayoutCount = 1,
 146                 .pSetLayouts = &device->meta_state.fast_clear_flush.dcc_decompress_compute_ds_layout,
 147                 .pushConstantRangeCount = 1,
 148                 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 8},
 149         };
 150
 151         result = radv_CreatePipelineLayout(radv_device_to_handle(device),
 152                                           &pl_create_info,
 153                                           &device->meta_state.alloc,
 154                                           &device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout);
 155         if (result != VK_SUCCESS)
 156                 goto cleanup;
 157
 158         /* compute shader */
 159
 160         VkPipelineShaderStageCreateInfo pipeline_shader_stage = {
 161                 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
 162                 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
 163                 .module = radv_shader_module_to_handle(&cs),
 164                 .pName = "main",
 165                 .pSpecializationInfo = NULL,
 166         };
 167
 168         VkComputePipelineCreateInfo vk_pipeline_info = {
 169                 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
 170                 .stage = pipeline_shader_stage,
 171                 .flags = 0,
 172                 .layout = device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout,
 173         };
 174
 175         result = radv_CreateComputePipelines(radv_device_to_handle(device),
 176                                              radv_pipeline_cache_to_handle(&device->meta_state.cache),
 177                                              1, &vk_pipeline_info, NULL,
 178                                              &device->meta_state.fast_clear_flush.dcc_decompress_compute_pipeline);
 179         if (result != VK_SUCCESS)
 180                 goto cleanup;
 181
 182 cleanup:
 183         ralloc_free(cs.nir);
 184         return result;
 185 }
 186
 187 static VkResult
 188 create_pass(struct radv_device *device)
 189 {
 190         VkResult result;
 191         VkDevice device_h = radv_device_to_handle(device);
 192         const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
 193         VkAttachmentDescription attachment;
 194
 195         attachment.format = VK_FORMAT_UNDEFINED;
 196         attachment.samples = 1;
 197         attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
 198         attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 199         attachment.initialLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
 200         attachment.finalLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
 201
 202         result = radv_CreateRenderPass(device_h,
 203                                        &(VkRenderPassCreateInfo) {
 204                                                .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
 205                                                        .attachmentCount = 1,
 206                                                        .pAttachments = &attachment,
 207                                                        .subpassCount = 1,
 208                                                        .pSubpasses = &(VkSubpassDescription) {
 209                                                        .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
 210                                                        .inputAttachmentCount = 0,
 211                                                        .colorAttachmentCount = 1,
 212                                                        .pColorAttachments = (VkAttachmentReference[]) {
 213                                                                {
 214                                                                        .attachment = 0,
 215                                                                        .layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
 216                                                                },
 217                                                        },
 218                                                        .pResolveAttachments = NULL,
 219                                                        .pDepthStencilAttachment = &(VkAttachmentReference) {
 220                                                                .attachment = VK_ATTACHMENT_UNUSED,
 221                                                        },
 222                                                        .preserveAttachmentCount = 0,
 223                                                        .pPreserveAttachments = NULL,
 224                                                },
 225                                                         .dependencyCount = 2,
 226                                                         .pDependencies = (VkSubpassDependency[]) {
 227                                                                 {
 228                                                                         .srcSubpass = VK_SUBPASS_EXTERNAL,
 229                                                                         .dstSubpass = 0,
 230                                                                         .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
 231                                                                         .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
 232                                                                         .srcAccessMask = 0,
 233                                                                         .dstAccessMask = 0,
 234                                                                         .dependencyFlags = 0
 235                                                                 },
 236                                                                 {
 237                                                                         .srcSubpass = 0,
 238                                                                         .dstSubpass = VK_SUBPASS_EXTERNAL,
 239                                                                         .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
 240                                                                         .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
 241                                                                         .srcAccessMask = 0,
 242                                                                         .dstAccessMask = 0,
 243                                                                         .dependencyFlags = 0
 244                                                                 }
 245                                                         },
 246                                        },
 247                                        alloc,
 248                                        &device->meta_state.fast_clear_flush.pass);
 249
 250         return result;
 251 }
 252
 253 static VkResult
 254 create_pipeline_layout(struct radv_device *device, VkPipelineLayout *layout)
 255 {
 256         VkPipelineLayoutCreateInfo pl_create_info = {
 257                 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
 258                 .setLayoutCount = 0,
 259                 .pSetLayouts = NULL,
 260                 .pushConstantRangeCount = 0,
 261                 .pPushConstantRanges = NULL,
 262         };
 263
 264         return radv_CreatePipelineLayout(radv_device_to_handle(device),
 265                                          &pl_create_info,
 266                                          &device->meta_state.alloc,
 267                                          layout);
 268 }
 269
 270 static VkResult
 271 create_pipeline(struct radv_device *device,
 272                 VkShaderModule vs_module_h,
 273                 VkPipelineLayout layout)
 274 {
 275         VkResult result;
 276         VkDevice device_h = radv_device_to_handle(device);
 277
 278         struct radv_shader_module fs_module = {
 279                 .nir = radv_meta_build_nir_fs_noop(),
 280         };
 281
 282         if (!fs_module.nir) {
 283                 /* XXX: Need more accurate error */
 284                 result = VK_ERROR_OUT_OF_HOST_MEMORY;
 285                 goto cleanup;
 286         }
 287
 288         const VkPipelineShaderStageCreateInfo stages[2] = {
 289                 {
 290                         .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
 291                         .stage = VK_SHADER_STAGE_VERTEX_BIT,
 292                         .module = vs_module_h,
 293                         .pName = "main",
 294                 },
 295                 {
 296                         .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
 297                         .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
 298                         .module = radv_shader_module_to_handle(&fs_module),
 299                         .pName = "main",
 300                 },
 301         };
 302
 303         const VkPipelineVertexInputStateCreateInfo vi_state = {
 304                 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
 305                 .vertexBindingDescriptionCount = 0,
 306                 .vertexAttributeDescriptionCount = 0,
 307         };
 308
 309         const VkPipelineInputAssemblyStateCreateInfo ia_state = {
 310                 .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
 311                 .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
 312                 .primitiveRestartEnable = false,
 313         };
 314
 315         const VkPipelineColorBlendStateCreateInfo blend_state = {
 316                 .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
 317                 .logicOpEnable = false,
 318                 .attachmentCount = 1,
 319                 .pAttachments = (VkPipelineColorBlendAttachmentState []) {
 320                         {
 321                                 .colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
 322                                 VK_COLOR_COMPONENT_G_BIT |
 323                                 VK_COLOR_COMPONENT_B_BIT |
 324                                 VK_COLOR_COMPONENT_A_BIT,
 325                         },
 326                 }
 327         };
 328         const VkPipelineRasterizationStateCreateInfo rs_state = {
 329                 .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
 330                 .depthClampEnable = false,
 331                 .rasterizerDiscardEnable = false,
 332                 .polygonMode = VK_POLYGON_MODE_FILL,
 333                 .cullMode = VK_CULL_MODE_NONE,
 334                 .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
 335         };
 336
 337         result = radv_graphics_pipeline_create(device_h,
 338                                                radv_pipeline_cache_to_handle(&device->meta_state.cache),
 339                                                &(VkGraphicsPipelineCreateInfo) {
 340                                                        .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
 341                                                        .stageCount = 2,
 342                                                        .pStages = stages,
 343
 344                                                        .pVertexInputState = &vi_state,
 345                                                        .pInputAssemblyState = &ia_state,
 346
 347                                                .pViewportState = &(VkPipelineViewportStateCreateInfo) {
 348                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
 349                                                        .viewportCount = 1,
 350                                                        .scissorCount = 1,
 351                                                },
 352                                                        .pRasterizationState = &rs_state,
 353                                                .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
 354                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
 355                                                        .rasterizationSamples = 1,
 356                                                        .sampleShadingEnable = false,
 357                                                        .pSampleMask = NULL,
 358                                                        .alphaToCoverageEnable = false,
 359                                                        .alphaToOneEnable = false,
 360                                                },
 361                                                 .pColorBlendState = &blend_state,
 362                                                 .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
 363                                                         .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
 364                                                         .dynamicStateCount = 2,
 365                                                         .pDynamicStates = (VkDynamicState[]) {
 366                                                                 VK_DYNAMIC_STATE_VIEWPORT,
 367                                                                 VK_DYNAMIC_STATE_SCISSOR,
 368                                                         },
 369                                                 },
 370                                                 .layout = layout,
 371                                                 .renderPass = device->meta_state.fast_clear_flush.pass,
 372                                                 .subpass = 0,
 373                                                },
 374                                                &(struct radv_graphics_pipeline_create_info) {
 375                                                        .use_rectlist = true,
 376                                                        .custom_blend_mode = V_028808_CB_ELIMINATE_FAST_CLEAR,
 377                                                },
 378                                                &device->meta_state.alloc,
 379                                                &device->meta_state.fast_clear_flush.cmask_eliminate_pipeline);
 380         if (result != VK_SUCCESS)
 381                 goto cleanup;
 382
 383         result = radv_graphics_pipeline_create(device_h,
 384                                                radv_pipeline_cache_to_handle(&device->meta_state.cache),
 385                                                &(VkGraphicsPipelineCreateInfo) {
 386                                                        .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
 387                                                        .stageCount = 2,
 388                                                        .pStages = stages,
 389
 390                                                        .pVertexInputState = &vi_state,
 391                                                        .pInputAssemblyState = &ia_state,
 392
 393                                                .pViewportState = &(VkPipelineViewportStateCreateInfo) {
 394                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
 395                                                        .viewportCount = 1,
 396                                                        .scissorCount = 1,
 397                                                },
 398                                                        .pRasterizationState = &rs_state,
 399                                                .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
 400                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
 401                                                        .rasterizationSamples = 1,
 402                                                        .sampleShadingEnable = false,
 403                                                        .pSampleMask = NULL,
 404                                                        .alphaToCoverageEnable = false,
 405                                                        .alphaToOneEnable = false,
 406                                                },
 407                                                 .pColorBlendState = &blend_state,
 408                                                 .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
 409                                                         .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
 410                                                         .dynamicStateCount = 2,
 411                                                         .pDynamicStates = (VkDynamicState[]) {
 412                                                                 VK_DYNAMIC_STATE_VIEWPORT,
 413                                                                 VK_DYNAMIC_STATE_SCISSOR,
 414                                                         },
 415                                                 },
 416                                                 .layout = layout,
 417                                                 .renderPass = device->meta_state.fast_clear_flush.pass,
 418                                                 .subpass = 0,
 419                                                },
 420                                                &(struct radv_graphics_pipeline_create_info) {
 421                                                        .use_rectlist = true,
 422                                                        .custom_blend_mode = V_028808_CB_FMASK_DECOMPRESS,
 423                                                },
 424                                                &device->meta_state.alloc,
 425                                                &device->meta_state.fast_clear_flush.fmask_decompress_pipeline);
 426         if (result != VK_SUCCESS)
 427                 goto cleanup;
 428
 429         result = radv_graphics_pipeline_create(device_h,
 430                                                radv_pipeline_cache_to_handle(&device->meta_state.cache),
 431                                                &(VkGraphicsPipelineCreateInfo) {
 432                                                        .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
 433                                                        .stageCount = 2,
 434                                                        .pStages = stages,
 435
 436                                                        .pVertexInputState = &vi_state,
 437                                                        .pInputAssemblyState = &ia_state,
 438
 439                                                .pViewportState = &(VkPipelineViewportStateCreateInfo) {
 440                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
 441                                                        .viewportCount = 1,
 442                                                        .scissorCount = 1,
 443                                                },
 444                                                        .pRasterizationState = &rs_state,
 445                                                .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
 446                                                        .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
 447                                                        .rasterizationSamples = 1,
 448                                                        .sampleShadingEnable = false,
 449                                                        .pSampleMask = NULL,
 450                                                        .alphaToCoverageEnable = false,
 451                                                        .alphaToOneEnable = false,
 452                                                },
 453                                                 .pColorBlendState = &blend_state,
 454                                                 .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
 455                                                         .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
 456                                                         .dynamicStateCount = 2,
 457                                                         .pDynamicStates = (VkDynamicState[]) {
 458                                                                 VK_DYNAMIC_STATE_VIEWPORT,
 459                                                                 VK_DYNAMIC_STATE_SCISSOR,
 460                                                         },
 461                                                 },
 462                                                 .layout = layout,
 463                                                 .renderPass = device->meta_state.fast_clear_flush.pass,
 464                                                 .subpass = 0,
 465                                                },
 466                                                &(struct radv_graphics_pipeline_create_info) {
 467                                                        .use_rectlist = true,
 468                                                        .custom_blend_mode = V_028808_CB_DCC_DECOMPRESS,
 469                                                },
 470                                                &device->meta_state.alloc,
 471                                                &device->meta_state.fast_clear_flush.dcc_decompress_pipeline);
 472         if (result != VK_SUCCESS)
 473                 goto cleanup;
 474
 475         goto cleanup;
 476
 477 cleanup:
 478         ralloc_free(fs_module.nir);
 479         return result;
 480 }
 481
 482 void
 483 radv_device_finish_meta_fast_clear_flush_state(struct radv_device *device)
 484 {
 485         struct radv_meta_state *state = &device->meta_state;
 486
 487         radv_DestroyPipeline(radv_device_to_handle(device),
 488                              state->fast_clear_flush.dcc_decompress_pipeline,
 489                              &state->alloc);
 490         radv_DestroyPipeline(radv_device_to_handle(device),
 491                              state->fast_clear_flush.fmask_decompress_pipeline,
 492                              &state->alloc);
 493         radv_DestroyPipeline(radv_device_to_handle(device),
 494                              state->fast_clear_flush.cmask_eliminate_pipeline,
 495                              &state->alloc);
 496         radv_DestroyRenderPass(radv_device_to_handle(device),
 497                                state->fast_clear_flush.pass, &state->alloc);
 498         radv_DestroyPipelineLayout(radv_device_to_handle(device),
 499                                    state->fast_clear_flush.p_layout,
 500                                    &state->alloc);
 501
 502         radv_DestroyPipeline(radv_device_to_handle(device),
 503                              state->fast_clear_flush.dcc_decompress_compute_pipeline,
 504                              &state->alloc);
 505         radv_DestroyPipelineLayout(radv_device_to_handle(device),
 506                                    state->fast_clear_flush.dcc_decompress_compute_p_layout,
 507                                    &state->alloc);
 508         radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
 509                                         state->fast_clear_flush.dcc_decompress_compute_ds_layout,
 510                                         &state->alloc);
 511 }
 512
 513 static VkResult
 514 radv_device_init_meta_fast_clear_flush_state_internal(struct radv_device *device)
 515 {
 516         VkResult res = VK_SUCCESS;
 517
 518         mtx_lock(&device->meta_state.mtx);
 519         if (device->meta_state.fast_clear_flush.cmask_eliminate_pipeline) {
 520                 mtx_unlock(&device->meta_state.mtx);
 521                 return VK_SUCCESS;
 522         }
 523
 524         struct radv_shader_module vs_module = { .nir = radv_meta_build_nir_vs_generate_vertices() };
 525         if (!vs_module.nir) {
 526                 /* XXX: Need more accurate error */
 527                 res = VK_ERROR_OUT_OF_HOST_MEMORY;
 528                 goto fail;
 529         }
 530
 531         res = create_pass(device);
 532         if (res != VK_SUCCESS)
 533                 goto fail;
 534
 535         res = create_pipeline_layout(device,
 536                                      &device->meta_state.fast_clear_flush.p_layout);
 537         if (res != VK_SUCCESS)
 538                 goto fail;
 539
 540         VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
 541         res = create_pipeline(device, vs_module_h,
 542                               device->meta_state.fast_clear_flush.p_layout);
 543         if (res != VK_SUCCESS)
 544                 goto fail;
 545
 546         res = create_dcc_compress_compute(device);
 547         if (res != VK_SUCCESS)
 548                 goto fail;
 549
 550         goto cleanup;
 551
 552 fail:
 553         radv_device_finish_meta_fast_clear_flush_state(device);
 554
 555 cleanup:
 556         ralloc_free(vs_module.nir);
 557         mtx_unlock(&device->meta_state.mtx);
 558
 559         return res;
 560 }
 561
 562
 563 VkResult
 564 radv_device_init_meta_fast_clear_flush_state(struct radv_device *device, bool on_demand)
 565 {
 566         if (on_demand)
 567                 return VK_SUCCESS;
 568
 569         return radv_device_init_meta_fast_clear_flush_state_internal(device);
 570 }
 571
 572 static void
 573 radv_emit_set_predication_state_from_image(struct radv_cmd_buffer *cmd_buffer,
 574                                       struct radv_image *image,
 575                                       uint64_t pred_offset, bool value)
 576 {
 577         uint64_t va = 0;
 578
 579         if (value) {
 580                 va = radv_buffer_get_va(image->bo) + image->offset;
 581                 va += pred_offset;
 582         }
 583
 584         si_emit_set_predication_state(cmd_buffer, true, va);
 585 }
 586
 587 static void
 588 radv_process_color_image_layer(struct radv_cmd_buffer *cmd_buffer,
 589                                struct radv_image *image,
 590                                const VkImageSubresourceRange *range,
 591                                int level, int layer)
 592 {
 593         struct radv_device *device = cmd_buffer->device;
 594         struct radv_image_view iview;
 595         uint32_t width, height;
 596
 597         width = radv_minify(image->info.width, range->baseMipLevel + level);
 598         height = radv_minify(image->info.height, range->baseMipLevel + level);
 599
 600         radv_image_view_init(&iview, device,
 601                              &(VkImageViewCreateInfo) {
 602                                 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 603                                 .image = radv_image_to_handle(image),
 604                                 .viewType = radv_meta_get_view_type(image),
 605                                 .format = image->vk_format,
 606                                 .subresourceRange = {
 607                                         .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 608                                         .baseMipLevel = range->baseMipLevel + level,
 609                                         .levelCount = 1,
 610                                         .baseArrayLayer = range->baseArrayLayer + layer,
 611                                         .layerCount = 1,
 612                                  },
 613                               }, NULL);
 614
 615         VkFramebuffer fb_h;
 616         radv_CreateFramebuffer(radv_device_to_handle(device),
 617                                &(VkFramebufferCreateInfo) {
 618                                         .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
 619                                         .attachmentCount = 1,
 620                                         .pAttachments = (VkImageView[]) {
 621                                                 radv_image_view_to_handle(&iview)
 622                                         },
 623                                         .width = width,
 624                                         .height = height,
 625                                         .layers = 1
 626                                 }, &cmd_buffer->pool->alloc, &fb_h);
 627
 628         radv_cmd_buffer_begin_render_pass(cmd_buffer,
 629                                           &(VkRenderPassBeginInfo) {
 630                                                 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
 631                                                 .renderPass = device->meta_state.fast_clear_flush.pass,
 632                                                 .framebuffer = fb_h,
 633                                                 .renderArea = {
 634                                                         .offset = { 0, 0, },
 635                                                         .extent = { width, height, }
 636                                                 },
 637                                                 .clearValueCount = 0,
 638                                                 .pClearValues = NULL,
 639                                         });
 640
 641         radv_cmd_buffer_set_subpass(cmd_buffer,
 642                                     &cmd_buffer->state.pass->subpasses[0]);
 643
 644         radv_CmdDraw(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
 645
 646         cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
 647                                         RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
 648
 649         radv_cmd_buffer_end_render_pass(cmd_buffer);
 650
 651         radv_DestroyFramebuffer(radv_device_to_handle(device), fb_h,
 652                                 &cmd_buffer->pool->alloc);
 653 }
 654
 655 static void
 656 radv_process_color_image(struct radv_cmd_buffer *cmd_buffer,
 657                          struct radv_image *image,
 658                          const VkImageSubresourceRange *subresourceRange,
 659                          bool decompress_dcc)
 660 {
 661         struct radv_meta_saved_state saved_state;
 662         VkPipeline *pipeline;
 663
 664         if (decompress_dcc && radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
 665                 pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.dcc_decompress_pipeline;
 666         } else if (radv_image_has_fmask(image) && !image->tc_compatible_cmask) {
 667                 pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.fmask_decompress_pipeline;
 668         } else {
 669                 pipeline = &cmd_buffer->device->meta_state.fast_clear_flush.cmask_eliminate_pipeline;
 670         }
 671
 672         if (!*pipeline) {
 673                 VkResult ret;
 674
 675                 ret = radv_device_init_meta_fast_clear_flush_state_internal(cmd_buffer->device);
 676                 if (ret != VK_SUCCESS) {
 677                         cmd_buffer->record_result = ret;
 678                         return;
 679                 }
 680         }
 681
 682         radv_meta_save(&saved_state, cmd_buffer,
 683                        RADV_META_SAVE_GRAPHICS_PIPELINE |
 684                        RADV_META_SAVE_PASS);
 685
 686         radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
 687                              VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline);
 688
 689         for (uint32_t l = 0; l < radv_get_levelCount(image, subresourceRange); ++l) {
 690                 uint32_t width, height;
 691
 692                 /* Do not decompress levels without DCC. */
 693                 if (decompress_dcc &&
 694                     !radv_dcc_enabled(image, subresourceRange->baseMipLevel + l))
 695                         continue;
 696
 697                 width = radv_minify(image->info.width,
 698                                     subresourceRange->baseMipLevel + l);
 699                 height = radv_minify(image->info.height,
 700                                     subresourceRange->baseMipLevel + l);
 701
 702                 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
 703                                     &(VkViewport) {
 704                                         .x = 0,
 705                                         .y = 0,
 706                                         .width = width,
 707                                         .height = height,
 708                                         .minDepth = 0.0f,
 709                                         .maxDepth = 1.0f
 710                                     });
 711
 712                 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
 713                                    &(VkRect2D) {
 714                                         .offset = { 0, 0 },
 715                                         .extent = { width, height },
 716                                    });
 717
 718                 for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
 719                         radv_process_color_image_layer(cmd_buffer, image,
 720                                                        subresourceRange, l, s);
 721                 }
 722         }
 723
 724         radv_meta_restore(&saved_state, cmd_buffer);
 725 }
 726
 727 static void
 728 radv_emit_color_decompress(struct radv_cmd_buffer *cmd_buffer,
 729                            struct radv_image *image,
 730                            const VkImageSubresourceRange *subresourceRange,
 731                            bool decompress_dcc)
 732 {
 733         bool old_predicating = false;
 734
 735         assert(cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL);
 736
 737         if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
 738                 uint64_t pred_offset = decompress_dcc ? image->dcc_pred_offset :
 739                                                         image->fce_pred_offset;
 740                 pred_offset += 8 * subresourceRange->baseMipLevel;
 741
 742                 old_predicating = cmd_buffer->state.predicating;
 743
 744                 radv_emit_set_predication_state_from_image(cmd_buffer, image, pred_offset, true);
 745                 cmd_buffer->state.predicating = true;
 746         }
 747
 748         radv_process_color_image(cmd_buffer, image, subresourceRange,
 749                                  decompress_dcc);
 750
 751         if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
 752                 uint64_t pred_offset = decompress_dcc ? image->dcc_pred_offset :
 753                                                         image->fce_pred_offset;
 754                 pred_offset += 8 * subresourceRange->baseMipLevel;
 755
 756                 cmd_buffer->state.predicating = old_predicating;
 757
 758                 radv_emit_set_predication_state_from_image(cmd_buffer, image, pred_offset, false);
 759
 760                 if (cmd_buffer->state.predication_type != -1) {
 761                         /* Restore previous conditional rendering user state. */
 762                         si_emit_set_predication_state(cmd_buffer,
 763                                                       cmd_buffer->state.predication_type,
 764                                                       cmd_buffer->state.predication_va);
 765                 }
 766         }
 767
 768         if (radv_dcc_enabled(image, subresourceRange->baseMipLevel)) {
 769                 /* Clear the image's fast-clear eliminate predicate because
 770                  * FMASK and DCC also imply a fast-clear eliminate.
 771                  */
 772                 radv_update_fce_metadata(cmd_buffer, image, subresourceRange, false);
 773
 774                 /* Mark the image as being decompressed. */
 775                 if (decompress_dcc)
 776                         radv_update_dcc_metadata(cmd_buffer, image, subresourceRange, false);
 777         }
 778 }
 779
 780 void
 781 radv_fast_clear_flush_image_inplace(struct radv_cmd_buffer *cmd_buffer,
 782                                     struct radv_image *image,
 783                                     const VkImageSubresourceRange *subresourceRange)
 784 {
 785         struct radv_barrier_data barrier = {};
 786
 787         if (radv_image_has_fmask(image)) {
 788                 barrier.layout_transitions.fmask_decompress = 1;
 789         } else {
 790                 barrier.layout_transitions.fast_clear_eliminate = 1;
 791         }
 792         radv_describe_layout_transition(cmd_buffer, &barrier);
 793
 794         radv_emit_color_decompress(cmd_buffer, image, subresourceRange, false);
 795 }
 796
 797 static void
 798 radv_decompress_dcc_gfx(struct radv_cmd_buffer *cmd_buffer,
 799                         struct radv_image *image,
 800                         const VkImageSubresourceRange *subresourceRange)
 801 {
 802         radv_emit_color_decompress(cmd_buffer, image, subresourceRange, true);
 803 }
 804
 805 static void
 806 radv_decompress_dcc_compute(struct radv_cmd_buffer *cmd_buffer,
 807                             struct radv_image *image,
 808                             const VkImageSubresourceRange *subresourceRange)
 809 {
 810         struct radv_meta_saved_state saved_state;
 811         struct radv_image_view load_iview = {0};
 812         struct radv_image_view store_iview = {0};
 813         struct radv_device *device = cmd_buffer->device;
 814
 815         /* This assumes the image is 2d with 1 layer */
 816         struct radv_cmd_state *state = &cmd_buffer->state;
 817
 818         state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
 819                              RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
 820
 821         if (!cmd_buffer->device->meta_state.fast_clear_flush.cmask_eliminate_pipeline) {
 822                 VkResult ret = radv_device_init_meta_fast_clear_flush_state_internal(cmd_buffer->device);
 823                 if (ret != VK_SUCCESS) {
 824                         cmd_buffer->record_result = ret;
 825                         return;
 826                 }
 827         }
 828
 829         radv_meta_save(&saved_state, cmd_buffer, RADV_META_SAVE_DESCRIPTORS |
 830                                                  RADV_META_SAVE_COMPUTE_PIPELINE);
 831
 832         radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
 833                              VK_PIPELINE_BIND_POINT_COMPUTE,
 834                              device->meta_state.fast_clear_flush.dcc_decompress_compute_pipeline);
 835
 836         for (uint32_t l = 0; l < radv_get_levelCount(image, subresourceRange); l++) {
 837                 uint32_t width, height;
 838
 839                 /* Do not decompress levels without DCC. */
 840                 if (!radv_dcc_enabled(image, subresourceRange->baseMipLevel + l))
 841                         continue;
 842
 843                 width = radv_minify(image->info.width,
 844                                     subresourceRange->baseMipLevel + l);
 845                 height = radv_minify(image->info.height,
 846                                     subresourceRange->baseMipLevel + l);
 847
 848                 for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
 849                         radv_image_view_init(&load_iview, cmd_buffer->device,
 850                                              &(VkImageViewCreateInfo) {
 851                                                      .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 852                                                              .image = radv_image_to_handle(image),
 853                                                              .viewType = VK_IMAGE_VIEW_TYPE_2D,
 854                                                              .format = image->vk_format,
 855                                                              .subresourceRange = {
 856                                                                 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 857                                                                 .baseMipLevel = subresourceRange->baseMipLevel + l,
 858                                                                 .levelCount = 1,
 859                                                                 .baseArrayLayer = subresourceRange->baseArrayLayer + s,
 860                                                                 .layerCount = 1
 861                                                              },
 862                                              }, NULL);
 863                         radv_image_view_init(&store_iview, cmd_buffer->device,
 864                                              &(VkImageViewCreateInfo) {
 865                                                      .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 866                                                              .image = radv_image_to_handle(image),
 867                                                              .viewType = VK_IMAGE_VIEW_TYPE_2D,
 868                                                              .format = image->vk_format,
 869                                                              .subresourceRange = {
 870                                                                 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 871                                                                 .baseMipLevel = subresourceRange->baseMipLevel + l,
 872                                                                 .levelCount = 1,
 873                                                                 .baseArrayLayer = subresourceRange->baseArrayLayer + s,
 874                                                                 .layerCount = 1
 875                                                              },
 876                                              }, &(struct radv_image_view_extra_create_info) {
 877                                                      .disable_compression = true
 878                                              });
 879
 880                         radv_meta_push_descriptor_set(cmd_buffer,
 881                                                       VK_PIPELINE_BIND_POINT_COMPUTE,
 882                                                       device->meta_state.fast_clear_flush.dcc_decompress_compute_p_layout,
 883                                                       0, /* set */
 884                                                       2, /* descriptorWriteCount */
 885                                                       (VkWriteDescriptorSet[]) {
 886                                                       {
 887                                                                        .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
 888                                                                        .dstBinding = 0,
 889                                                                        .dstArrayElement = 0,
 890                                                                        .descriptorCount = 1,
 891                                                                        .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
 892                                                                        .pImageInfo = (VkDescriptorImageInfo[]) {
 893                                                                                {
 894                                                                                        .sampler = VK_NULL_HANDLE,
 895                                                                                        .imageView = radv_image_view_to_handle(&load_iview),
 896                                                                                        .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
 897                                                                                },
 898                                                                        }
 899                                                               },
 900                                                               {
 901                                                                        .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
 902                                                                        .dstBinding = 1,
 903                                                                        .dstArrayElement = 0,
 904                                                                        .descriptorCount = 1,
 905                                                                        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
 906                                                                        .pImageInfo = (VkDescriptorImageInfo[]) {
 907                                                                                {
 908                                                                                        .sampler = VK_NULL_HANDLE,
 909                                                                                        .imageView = radv_image_view_to_handle(&store_iview),
 910                                                                                        .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
 911                                                                                },
 912                                                                        }
 913                                                               }
 914                                                       });
 915
 916                         radv_unaligned_dispatch(cmd_buffer, width, height, 1);
 917                 }
 918         }
 919
 920         /* Mark this image as actually being decompressed. */
 921         radv_update_dcc_metadata(cmd_buffer, image, subresourceRange, false);
 922
 923         /* The fill buffer below does its own saving */
 924         radv_meta_restore(&saved_state, cmd_buffer);
 925
 926         state->flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
 927                              RADV_CMD_FLAG_INV_VCACHE;
 928
 929
 930         /* Initialize the DCC metadata as "fully expanded". */
 931         radv_initialize_dcc(cmd_buffer, image, subresourceRange, 0xffffffff);
 932 }
 933
 934 void
 935 radv_decompress_dcc(struct radv_cmd_buffer *cmd_buffer,
 936                     struct radv_image *image,
 937                     const VkImageSubresourceRange *subresourceRange)
 938 {
 939         struct radv_barrier_data barrier = {};
 940
 941         barrier.layout_transitions.dcc_decompress = 1;
 942         radv_describe_layout_transition(cmd_buffer, &barrier);
 943
 944         if (cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL)
 945                 radv_decompress_dcc_gfx(cmd_buffer, image, subresourceRange);
 946         else
 947                 radv_decompress_dcc_compute(cmd_buffer, image, subresourceRange);
 948 }