src/amd/vulkan/radv_meta_bufimage.c

   1 /*
   2  * Copyright © 2016 Red Hat.
   3  * Copyright © 2016 Bas Nieuwenhuizen
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24 #include "radv_meta.h"
  25 #include "nir/nir_builder.h"
  26
  27 /*
  28  * Compute shader implementation of image->buffer copy.
  29  */
  30
  31 static nir_shader *
  32 build_nir_itob_compute_shader(struct radv_device *dev)
  33 {
  34         nir_builder b;
  35         const struct glsl_type *sampler_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
  36                                                                  false,
  37                                                                  false,
  38                                                                  GLSL_TYPE_FLOAT);
  39         const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_BUF,
  40                                                              false,
  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, "meta_itob_cs");
  45         b.shader->info.cs.local_size[0] = 16;
  46         b.shader->info.cs.local_size[1] = 16;
  47         b.shader->info.cs.local_size[2] = 1;
  48         nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
  49                                                       sampler_type, "s_tex");
  50         input_img->data.descriptor_set = 0;
  51         input_img->data.binding = 0;
  52
  53         nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
  54                                                        img_type, "out_img");
  55         output_img->data.descriptor_set = 0;
  56         output_img->data.binding = 1;
  57
  58         nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
  59         nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
  60         nir_ssa_def *block_size = nir_imm_ivec4(&b,
  61                                                 b.shader->info.cs.local_size[0],
  62                                                 b.shader->info.cs.local_size[1],
  63                                                 b.shader->info.cs.local_size[2], 0);
  64
  65         nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
  66
  67
  68
  69         nir_intrinsic_instr *offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
  70         offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
  71         offset->num_components = 2;
  72         nir_ssa_dest_init(&offset->instr, &offset->dest, 2, 32, "offset");
  73         nir_builder_instr_insert(&b, &offset->instr);
  74
  75         nir_intrinsic_instr *stride = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
  76         stride->src[0] = nir_src_for_ssa(nir_imm_int(&b, 8));
  77         stride->num_components = 1;
  78         nir_ssa_dest_init(&stride->instr, &stride->dest, 1, 32, "stride");
  79         nir_builder_instr_insert(&b, &stride->instr);
  80
  81         nir_ssa_def *img_coord = nir_iadd(&b, global_id, &offset->dest.ssa);
  82
  83         nir_tex_instr *tex = nir_tex_instr_create(b.shader, 2);
  84         tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
  85         tex->op = nir_texop_txf;
  86         tex->src[0].src_type = nir_tex_src_coord;
  87         tex->src[0].src = nir_src_for_ssa(img_coord);
  88         tex->src[1].src_type = nir_tex_src_lod;
  89         tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
  90         tex->dest_type = nir_type_float;
  91         tex->is_array = false;
  92         tex->coord_components = 2;
  93         tex->texture = nir_deref_var_create(tex, input_img);
  94         tex->sampler = NULL;
  95
  96         nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
  97         nir_builder_instr_insert(&b, &tex->instr);
  98
  99         nir_ssa_def *pos_x = nir_channel(&b, global_id, 0);
 100         nir_ssa_def *pos_y = nir_channel(&b, global_id, 1);
 101
 102         nir_ssa_def *tmp = nir_imul(&b, pos_y, &stride->dest.ssa);
 103         tmp = nir_iadd(&b, tmp, pos_x);
 104
 105         nir_ssa_def *coord = nir_vec4(&b, tmp, tmp, tmp, tmp);
 106
 107         nir_ssa_def *outval = &tex->dest.ssa;
 108         nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_store);
 109         store->src[0] = nir_src_for_ssa(coord);
 110         store->src[1] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
 111         store->src[2] = nir_src_for_ssa(outval);
 112         store->variables[0] = nir_deref_var_create(store, output_img);
 113
 114         nir_builder_instr_insert(&b, &store->instr);
 115         return b.shader;
 116 }
 117
 118 /* Image to buffer - don't write use image accessors */
 119 static VkResult
 120 radv_device_init_meta_itob_state(struct radv_device *device)
 121 {
 122         VkResult result;
 123         struct radv_shader_module cs = { .nir = NULL };
 124
 125         zero(device->meta_state.itob);
 126
 127         cs.nir = build_nir_itob_compute_shader(device);
 128
 129         /*
 130          * two descriptors one for the image being sampled
 131          * one for the buffer being written.
 132          */
 133         VkDescriptorSetLayoutCreateInfo ds_create_info = {
 134                 .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
 135                 .bindingCount = 2,
 136                 .pBindings = (VkDescriptorSetLayoutBinding[]) {
 137                         {
 138                                 .binding = 0,
 139                                 .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
 140                                 .descriptorCount = 1,
 141                                 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
 142                                 .pImmutableSamplers = NULL
 143                         },
 144                         {
 145                                 .binding = 1,
 146                                 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER,
 147                                 .descriptorCount = 1,
 148                                 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
 149                                 .pImmutableSamplers = NULL
 150                         },
 151                 }
 152         };
 153
 154         result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
 155                                                 &ds_create_info,
 156                                                 &device->meta_state.alloc,
 157                                                 &device->meta_state.itob.img_ds_layout);
 158         if (result != VK_SUCCESS)
 159                 goto fail;
 160
 161
 162         VkPipelineLayoutCreateInfo pl_create_info = {
 163                 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
 164                 .setLayoutCount = 1,
 165                 .pSetLayouts = &device->meta_state.itob.img_ds_layout,
 166                 .pushConstantRangeCount = 1,
 167                 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 12},
 168         };
 169
 170         result = radv_CreatePipelineLayout(radv_device_to_handle(device),
 171                                           &pl_create_info,
 172                                           &device->meta_state.alloc,
 173                                           &device->meta_state.itob.img_p_layout);
 174         if (result != VK_SUCCESS)
 175                 goto fail;
 176
 177         /* compute shader */
 178
 179         VkPipelineShaderStageCreateInfo pipeline_shader_stage = {
 180                 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
 181                 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
 182                 .module = radv_shader_module_to_handle(&cs),
 183                 .pName = "main",
 184                 .pSpecializationInfo = NULL,
 185         };
 186
 187         VkComputePipelineCreateInfo vk_pipeline_info = {
 188                 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
 189                 .stage = pipeline_shader_stage,
 190                 .flags = 0,
 191                 .layout = device->meta_state.itob.img_p_layout,
 192         };
 193
 194         result = radv_CreateComputePipelines(radv_device_to_handle(device),
 195                                              radv_pipeline_cache_to_handle(&device->meta_state.cache),
 196                                              1, &vk_pipeline_info, NULL,
 197                                              &device->meta_state.itob.pipeline);
 198         if (result != VK_SUCCESS)
 199                 goto fail;
 200
 201         ralloc_free(cs.nir);
 202         return VK_SUCCESS;
 203 fail:
 204         ralloc_free(cs.nir);
 205         return result;
 206 }
 207
 208 static void
 209 radv_device_finish_meta_itob_state(struct radv_device *device)
 210 {
 211         if (device->meta_state.itob.img_p_layout) {
 212                 radv_DestroyPipelineLayout(radv_device_to_handle(device),
 213                                            device->meta_state.itob.img_p_layout,
 214                                            &device->meta_state.alloc);
 215         }
 216         if (device->meta_state.itob.img_ds_layout) {
 217                 radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
 218                                                 device->meta_state.itob.img_ds_layout,
 219                                                 &device->meta_state.alloc);
 220         }
 221         if (device->meta_state.itob.pipeline) {
 222                 radv_DestroyPipeline(radv_device_to_handle(device),
 223                                      device->meta_state.itob.pipeline,
 224                                      &device->meta_state.alloc);
 225         }
 226 }
 227
 228 void
 229 radv_device_finish_meta_bufimage_state(struct radv_device *device)
 230 {
 231         radv_device_finish_meta_itob_state(device);
 232 }
 233
 234 VkResult
 235 radv_device_init_meta_bufimage_state(struct radv_device *device)
 236 {
 237         VkResult result;
 238
 239         result = radv_device_init_meta_itob_state(device);
 240         if (result != VK_SUCCESS)
 241                 return result;
 242         return VK_SUCCESS;
 243 }
 244
 245 void
 246 radv_meta_begin_bufimage(struct radv_cmd_buffer *cmd_buffer,
 247                          struct radv_meta_saved_compute_state *save)
 248 {
 249         radv_meta_save_compute(save, cmd_buffer, 12);
 250 }
 251
 252 void
 253 radv_meta_end_bufimage(struct radv_cmd_buffer *cmd_buffer,
 254                        struct radv_meta_saved_compute_state *save)
 255 {
 256         radv_meta_restore_compute(save, cmd_buffer, 12);
 257 }
 258
 259 static void
 260 create_iview(struct radv_cmd_buffer *cmd_buffer,
 261              struct radv_meta_blit2d_surf *surf,
 262              VkImageUsageFlags usage,
 263              struct radv_image_view *iview)
 264 {
 265
 266         radv_image_view_init(iview, cmd_buffer->device,
 267                              &(VkImageViewCreateInfo) {
 268                                      .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 269                                              .image = radv_image_to_handle(surf->image),
 270                                              .viewType = VK_IMAGE_VIEW_TYPE_2D,
 271                                              .format = surf->format,
 272                                              .subresourceRange = {
 273                                              .aspectMask = surf->aspect_mask,
 274                                              .baseMipLevel = surf->level,
 275                                              .levelCount = 1,
 276                                              .baseArrayLayer = surf->layer,
 277                                              .layerCount = 1
 278                                      },
 279                                              }, cmd_buffer, usage);
 280 }
 281
 282 static void
 283 create_bview(struct radv_cmd_buffer *cmd_buffer,
 284              struct radv_buffer *buffer,
 285              unsigned offset,
 286              VkFormat format,
 287              struct radv_buffer_view *bview)
 288 {
 289         radv_buffer_view_init(bview, cmd_buffer->device,
 290                               &(VkBufferViewCreateInfo) {
 291                                       .sType = VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO,
 292                                       .flags = 0,
 293                                       .buffer = radv_buffer_to_handle(buffer),
 294                                       .format = format,
 295                                       .offset = offset,
 296                                       .range = VK_WHOLE_SIZE,
 297                               }, cmd_buffer);
 298
 299 }
 300
 301 struct itob_temps {
 302         struct radv_image_view src_iview;
 303
 304         struct radv_buffer_view dst_bview;
 305         VkDescriptorSet set;
 306 };
 307
 308 static void
 309 itob_bind_src_image(struct radv_cmd_buffer *cmd_buffer,
 310                    struct radv_meta_blit2d_surf *src,
 311                    struct radv_meta_blit2d_rect *rect,
 312                    struct itob_temps *tmp)
 313 {
 314         create_iview(cmd_buffer, src, VK_IMAGE_USAGE_SAMPLED_BIT, &tmp->src_iview);
 315 }
 316
 317 static void
 318 itob_bind_dst_buffer(struct radv_cmd_buffer *cmd_buffer,
 319                      struct radv_meta_blit2d_buffer *dst,
 320                      struct radv_meta_blit2d_rect *rect,
 321                      struct itob_temps *tmp)
 322 {
 323         create_bview(cmd_buffer, dst->buffer, dst->offset, dst->format, &tmp->dst_bview);
 324 }
 325
 326 static void
 327 itob_bind_descriptors(struct radv_cmd_buffer *cmd_buffer,
 328                       struct itob_temps *tmp)
 329 {
 330         struct radv_device *device = cmd_buffer->device;
 331         VkDevice vk_device = radv_device_to_handle(cmd_buffer->device);
 332
 333         radv_temp_descriptor_set_create(device, cmd_buffer,
 334                                         device->meta_state.itob.img_ds_layout,
 335                                         &tmp->set);
 336
 337         radv_UpdateDescriptorSets(vk_device,
 338                                   2, /* writeCount */
 339                                   (VkWriteDescriptorSet[]) {
 340                                           {
 341                                                   .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
 342                                                   .dstSet = tmp->set,
 343                                                   .dstBinding = 0,
 344                                                   .dstArrayElement = 0,
 345                                                   .descriptorCount = 1,
 346                                                   .descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE,
 347                                                   .pImageInfo = (VkDescriptorImageInfo[]) {
 348                                                           {
 349                                                                   .sampler = NULL,
 350                                                                   .imageView = radv_image_view_to_handle(&tmp->src_iview),
 351                                                                   .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
 352                                                           },
 353                                                   }
 354                                           },
 355                                           {
 356                                                   .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
 357                                                   .dstSet = tmp->set,
 358                                                   .dstBinding = 1,
 359                                                   .dstArrayElement = 0,
 360                                                   .descriptorCount = 1,
 361                                                   .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER,
 362                                                   .pTexelBufferView = (VkBufferView[])  { radv_buffer_view_to_handle(&tmp->dst_bview) },
 363                                           }
 364                                   }, 0, NULL);
 365
 366         radv_CmdBindDescriptorSets(radv_cmd_buffer_to_handle(cmd_buffer),
 367                                    VK_PIPELINE_BIND_POINT_COMPUTE,
 368                                    device->meta_state.itob.img_p_layout, 0, 1,
 369                                    &tmp->set, 0, NULL);
 370 }
 371
 372 static void
 373 itob_unbind_src_image(struct radv_cmd_buffer *cmd_buffer,
 374                       struct itob_temps *temps)
 375 {
 376 }
 377
 378 static void
 379 bind_pipeline(struct radv_cmd_buffer *cmd_buffer)
 380 {
 381         VkPipeline pipeline =
 382                 cmd_buffer->device->meta_state.itob.pipeline;
 383
 384         if (cmd_buffer->state.compute_pipeline != radv_pipeline_from_handle(pipeline)) {
 385                 radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
 386                                      VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
 387         }
 388 }
 389
 390 void
 391 radv_meta_image_to_buffer(struct radv_cmd_buffer *cmd_buffer,
 392                           struct radv_meta_blit2d_surf *src,
 393                           struct radv_meta_blit2d_buffer *dst,
 394                           unsigned num_rects,
 395                           struct radv_meta_blit2d_rect *rects)
 396 {
 397         struct radv_device *device = cmd_buffer->device;
 398
 399         for (unsigned r = 0; r < num_rects; ++r) {
 400                 struct itob_temps temps;
 401
 402                 itob_bind_src_image(cmd_buffer, src, &rects[r], &temps);
 403                 itob_bind_dst_buffer(cmd_buffer, dst, &rects[r], &temps);
 404                 itob_bind_descriptors(cmd_buffer, &temps);
 405
 406                 bind_pipeline(cmd_buffer);
 407
 408                 unsigned push_constants[3] = {
 409                         rects[r].src_x,
 410                         rects[r].src_y,
 411                         dst->pitch
 412                 };
 413                 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
 414                                       device->meta_state.itob.img_p_layout,
 415                                       VK_SHADER_STAGE_COMPUTE_BIT, 0, 12,
 416                                       push_constants);
 417
 418                 radv_unaligned_dispatch(cmd_buffer, rects[r].width, rects[r].height, 1);
 419                 radv_temp_descriptor_set_destroy(cmd_buffer->device, temps.set);
 420                 itob_unbind_src_image(cmd_buffer, &temps);
 421         }
 422
 423 }