src/freedreno/vulkan/tu_descriptor_set.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
  22  * DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 /**
  26  * @file
  27  *
  28  * We use the bindless descriptor model, which maps fairly closely to how
  29  * Vulkan descriptor sets work. The two exceptions are input attachments and
  30  * dynamic descriptors, which have to be patched when recording command
  31  * buffers. We reserve an extra descriptor set for these. This descriptor set
  32  * contains all the input attachments in the pipeline, in order, and then all
  33  * the dynamic descriptors. The dynamic descriptors are stored in the CPU-side
  34  * datastructure for each tu_descriptor_set, and then combined into one big
  35  * descriptor set at CmdBindDescriptors time/draw time.
  36  */
  37
  38 #include "tu_private.h"
  39
  40 #include <assert.h>
  41 #include <fcntl.h>
  42 #include <stdbool.h>
  43 #include <string.h>
  44 #include <unistd.h>
  45
  46 #include "util/mesa-sha1.h"
  47 #include "vk_util.h"
  48
  49 static int
  50 binding_compare(const void *av, const void *bv)
  51 {
  52    const VkDescriptorSetLayoutBinding *a =
  53       (const VkDescriptorSetLayoutBinding *) av;
  54    const VkDescriptorSetLayoutBinding *b =
  55       (const VkDescriptorSetLayoutBinding *) bv;
  56
  57    return (a->binding < b->binding) ? -1 : (a->binding > b->binding) ? 1 : 0;
  58 }
  59
  60 static VkDescriptorSetLayoutBinding *
  61 create_sorted_bindings(const VkDescriptorSetLayoutBinding *bindings,
  62                        unsigned count)
  63 {
  64    VkDescriptorSetLayoutBinding *sorted_bindings =
  65       malloc(count * sizeof(VkDescriptorSetLayoutBinding));
  66    if (!sorted_bindings)
  67       return NULL;
  68
  69    memcpy(sorted_bindings, bindings,
  70           count * sizeof(VkDescriptorSetLayoutBinding));
  71
  72    qsort(sorted_bindings, count, sizeof(VkDescriptorSetLayoutBinding),
  73          binding_compare);
  74
  75    return sorted_bindings;
  76 }
  77
  78 static uint32_t
  79 descriptor_size(VkDescriptorType type)
  80 {
  81    switch (type) {
  82    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
  83    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
  84    case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
  85       /* These are remapped to the special driver-managed descriptor set,
  86        * hence they don't take up any space in the original descriptor set:
  87        */
  88       return 0;
  89    case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
  90       /* We make offsets and sizes all 16 dwords, to match how the hardware
  91        * interprets indices passed to sample/load/store instructions in
  92        * multiples of 16 dwords.  This means that "normal" descriptors are all
  93        * of size 16, with padding for smaller descriptors like uniform storage
  94        * descriptors which are less than 16 dwords. However combined images
  95        * and samplers are actually two descriptors, so they have size 2.
  96        */
  97       return A6XX_TEX_CONST_DWORDS * 4 * 2;
  98    default:
  99       return A6XX_TEX_CONST_DWORDS * 4;
 100    }
 101 }
 102
 103 VkResult
 104 tu_CreateDescriptorSetLayout(
 105    VkDevice _device,
 106    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
 107    const VkAllocationCallbacks *pAllocator,
 108    VkDescriptorSetLayout *pSetLayout)
 109 {
 110    TU_FROM_HANDLE(tu_device, device, _device);
 111    struct tu_descriptor_set_layout *set_layout;
 112
 113    assert(pCreateInfo->sType ==
 114           VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO);
 115    const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
 116       vk_find_struct_const(
 117          pCreateInfo->pNext,
 118          DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
 119
 120    uint32_t max_binding = 0;
 121    uint32_t immutable_sampler_count = 0;
 122    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
 123       max_binding = MAX2(max_binding, pCreateInfo->pBindings[j].binding);
 124       if ((pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
 125            pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
 126            pCreateInfo->pBindings[j].pImmutableSamplers) {
 127          immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
 128       }
 129    }
 130
 131    uint32_t samplers_offset = sizeof(struct tu_descriptor_set_layout) +
 132       (max_binding + 1) * sizeof(set_layout->binding[0]);
 133    uint32_t size = samplers_offset + immutable_sampler_count * A6XX_TEX_SAMP_DWORDS * 4;
 134
 135    set_layout = vk_alloc2(&device->alloc, pAllocator, size, 8,
 136                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 137    if (!set_layout)
 138       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 139
 140    set_layout->flags = pCreateInfo->flags;
 141
 142    /* We just allocate all the immutable samplers at the end of the struct */
 143    struct tu_sampler *samplers = (void*) &set_layout->binding[max_binding + 1];
 144
 145    VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(
 146       pCreateInfo->pBindings, pCreateInfo->bindingCount);
 147    if (!bindings) {
 148       vk_free2(&device->alloc, pAllocator, set_layout);
 149       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 150    }
 151
 152    set_layout->binding_count = max_binding + 1;
 153    set_layout->shader_stages = 0;
 154    set_layout->has_immutable_samplers = false;
 155    set_layout->size = 0;
 156    set_layout->dynamic_ubo = 0;
 157
 158    memset(set_layout->binding, 0,
 159           size - sizeof(struct tu_descriptor_set_layout));
 160
 161    uint32_t dynamic_offset_count = 0;
 162    uint32_t input_attachment_count = 0;
 163    uint32_t buffer_count = 0;
 164
 165    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
 166       const VkDescriptorSetLayoutBinding *binding = bindings + j;
 167       uint32_t b = binding->binding;
 168
 169       set_layout->binding[b].type = binding->descriptorType;
 170       set_layout->binding[b].array_size = binding->descriptorCount;
 171       set_layout->binding[b].offset = set_layout->size;
 172       set_layout->binding[b].buffer_offset = buffer_count;
 173       set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
 174       set_layout->binding[b].input_attachment_offset = input_attachment_count;
 175       set_layout->binding[b].size = descriptor_size(binding->descriptorType);
 176       set_layout->binding[b].shader_stages = binding->stageFlags;
 177
 178       if (variable_flags && binding->binding < variable_flags->bindingCount &&
 179           (variable_flags->pBindingFlags[binding->binding] &
 180            VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
 181          assert(!binding->pImmutableSamplers); /* Terribly ill defined  how
 182                                                   many samplers are valid */
 183          assert(binding->binding == max_binding);
 184
 185          set_layout->has_variable_descriptors = true;
 186       }
 187
 188       if ((binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
 189            binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
 190           binding->pImmutableSamplers) {
 191          set_layout->binding[b].immutable_samplers_offset = samplers_offset;
 192          set_layout->has_immutable_samplers = true;
 193
 194          for (uint32_t i = 0; i < binding->descriptorCount; i++)
 195             samplers[i] = *tu_sampler_from_handle(binding->pImmutableSamplers[i]);
 196
 197          samplers += binding->descriptorCount;
 198          samplers_offset += sizeof(struct tu_sampler) * binding->descriptorCount;
 199       }
 200
 201       set_layout->size +=
 202          binding->descriptorCount * set_layout->binding[b].size;
 203       if (binding->descriptorType != VK_DESCRIPTOR_TYPE_SAMPLER &&
 204           binding->descriptorType != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
 205          buffer_count += binding->descriptorCount;
 206       if (binding->descriptorType == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC ||
 207           binding->descriptorType == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) {
 208          if (binding->descriptorType == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) {
 209             STATIC_ASSERT(MAX_DYNAMIC_BUFFERS <= 8 * sizeof(set_layout->dynamic_ubo));
 210             set_layout->dynamic_ubo |=
 211                ((1u << binding->descriptorCount) - 1) << dynamic_offset_count;
 212          }
 213
 214          dynamic_offset_count += binding->descriptorCount;
 215       }
 216       if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT)
 217          input_attachment_count += binding->descriptorCount;
 218       set_layout->shader_stages |= binding->stageFlags;
 219    }
 220
 221    free(bindings);
 222
 223    set_layout->dynamic_offset_count = dynamic_offset_count;
 224    set_layout->input_attachment_count = input_attachment_count;
 225    set_layout->buffer_count = buffer_count;
 226
 227    *pSetLayout = tu_descriptor_set_layout_to_handle(set_layout);
 228
 229    return VK_SUCCESS;
 230 }
 231
 232 void
 233 tu_DestroyDescriptorSetLayout(VkDevice _device,
 234                               VkDescriptorSetLayout _set_layout,
 235                               const VkAllocationCallbacks *pAllocator)
 236 {
 237    TU_FROM_HANDLE(tu_device, device, _device);
 238    TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout, _set_layout);
 239
 240    if (!set_layout)
 241       return;
 242
 243    vk_free2(&device->alloc, pAllocator, set_layout);
 244 }
 245
 246 void
 247 tu_GetDescriptorSetLayoutSupport(
 248    VkDevice device,
 249    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
 250    VkDescriptorSetLayoutSupport *pSupport)
 251 {
 252    VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(
 253       pCreateInfo->pBindings, pCreateInfo->bindingCount);
 254    if (!bindings) {
 255       pSupport->supported = false;
 256       return;
 257    }
 258
 259    const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
 260       vk_find_struct_const(
 261          pCreateInfo->pNext,
 262          DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
 263    VkDescriptorSetVariableDescriptorCountLayoutSupportEXT *variable_count =
 264       vk_find_struct(
 265          (void *) pCreateInfo->pNext,
 266          DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT_EXT);
 267    if (variable_count) {
 268       variable_count->maxVariableDescriptorCount = 0;
 269    }
 270
 271    bool supported = true;
 272    uint64_t size = 0;
 273    for (uint32_t i = 0; i < pCreateInfo->bindingCount; i++) {
 274       const VkDescriptorSetLayoutBinding *binding = bindings + i;
 275
 276       uint64_t descriptor_sz = descriptor_size(binding->descriptorType);
 277       uint64_t descriptor_alignment = 8;
 278
 279       if (size && !align_u64(size, descriptor_alignment)) {
 280          supported = false;
 281       }
 282       size = align_u64(size, descriptor_alignment);
 283
 284       uint64_t max_count = UINT64_MAX;
 285       if (descriptor_sz)
 286          max_count = (UINT64_MAX - size) / descriptor_sz;
 287
 288       if (max_count < binding->descriptorCount) {
 289          supported = false;
 290       }
 291       if (variable_flags && binding->binding < variable_flags->bindingCount &&
 292           variable_count &&
 293           (variable_flags->pBindingFlags[binding->binding] &
 294            VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
 295          variable_count->maxVariableDescriptorCount =
 296             MIN2(UINT32_MAX, max_count);
 297       }
 298       size += binding->descriptorCount * descriptor_sz;
 299    }
 300
 301    free(bindings);
 302
 303    pSupport->supported = supported;
 304 }
 305
 306 /*
 307  * Pipeline layouts.  These have nothing to do with the pipeline.  They are
 308  * just multiple descriptor set layouts pasted together.
 309  */
 310
 311 VkResult
 312 tu_CreatePipelineLayout(VkDevice _device,
 313                         const VkPipelineLayoutCreateInfo *pCreateInfo,
 314                         const VkAllocationCallbacks *pAllocator,
 315                         VkPipelineLayout *pPipelineLayout)
 316 {
 317    TU_FROM_HANDLE(tu_device, device, _device);
 318    struct tu_pipeline_layout *layout;
 319    struct mesa_sha1 ctx;
 320
 321    assert(pCreateInfo->sType ==
 322           VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
 323
 324    layout = vk_alloc2(&device->alloc, pAllocator, sizeof(*layout), 8,
 325                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 326    if (layout == NULL)
 327       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 328
 329    layout->num_sets = pCreateInfo->setLayoutCount;
 330    layout->input_attachment_count = 0;
 331    layout->dynamic_offset_count = 0;
 332
 333    unsigned dynamic_offset_count = 0, input_attachment_count = 0;
 334
 335    _mesa_sha1_init(&ctx);
 336    for (uint32_t set = 0; set < pCreateInfo->setLayoutCount; set++) {
 337       TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
 338                      pCreateInfo->pSetLayouts[set]);
 339       layout->set[set].layout = set_layout;
 340       layout->set[set].dynamic_offset_start = dynamic_offset_count;
 341       layout->set[set].input_attachment_start = input_attachment_count;
 342       dynamic_offset_count += set_layout->dynamic_offset_count;
 343       input_attachment_count += set_layout->input_attachment_count;
 344
 345       for (uint32_t b = 0; b < set_layout->binding_count; b++) {
 346          if (set_layout->binding[b].immutable_samplers_offset)
 347             _mesa_sha1_update(
 348                &ctx,
 349                tu_immutable_samplers(set_layout, set_layout->binding + b),
 350                set_layout->binding[b].array_size * 4 * sizeof(uint32_t));
 351       }
 352       _mesa_sha1_update(
 353          &ctx, set_layout->binding,
 354          sizeof(set_layout->binding[0]) * set_layout->binding_count);
 355    }
 356
 357    layout->dynamic_offset_count = dynamic_offset_count;
 358    layout->input_attachment_count = input_attachment_count;
 359    layout->push_constant_size = 0;
 360
 361    for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) {
 362       const VkPushConstantRange *range = pCreateInfo->pPushConstantRanges + i;
 363       layout->push_constant_size =
 364          MAX2(layout->push_constant_size, range->offset + range->size);
 365    }
 366
 367    layout->push_constant_size = align(layout->push_constant_size, 16);
 368    _mesa_sha1_update(&ctx, &layout->push_constant_size,
 369                      sizeof(layout->push_constant_size));
 370    _mesa_sha1_final(&ctx, layout->sha1);
 371    *pPipelineLayout = tu_pipeline_layout_to_handle(layout);
 372
 373    return VK_SUCCESS;
 374 }
 375
 376 void
 377 tu_DestroyPipelineLayout(VkDevice _device,
 378                          VkPipelineLayout _pipelineLayout,
 379                          const VkAllocationCallbacks *pAllocator)
 380 {
 381    TU_FROM_HANDLE(tu_device, device, _device);
 382    TU_FROM_HANDLE(tu_pipeline_layout, pipeline_layout, _pipelineLayout);
 383
 384    if (!pipeline_layout)
 385       return;
 386    vk_free2(&device->alloc, pAllocator, pipeline_layout);
 387 }
 388
 389 #define EMPTY 1
 390
 391 static VkResult
 392 tu_descriptor_set_create(struct tu_device *device,
 393             struct tu_descriptor_pool *pool,
 394             const struct tu_descriptor_set_layout *layout,
 395             const uint32_t *variable_count,
 396             struct tu_descriptor_set **out_set)
 397 {
 398    struct tu_descriptor_set *set;
 399    uint32_t buffer_count = layout->buffer_count;
 400    if (variable_count) {
 401       unsigned stride = 1;
 402       if (layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_SAMPLER ||
 403           layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
 404          stride = 0;
 405       buffer_count = layout->binding[layout->binding_count - 1].buffer_offset +
 406                      *variable_count * stride;
 407    }
 408    unsigned dynamic_offset = sizeof(struct tu_descriptor_set) +
 409       sizeof(struct tu_bo *) * buffer_count;
 410    unsigned mem_size = dynamic_offset +
 411       A6XX_TEX_CONST_DWORDS * 4 * (layout->dynamic_offset_count +
 412                                    layout->input_attachment_count);;
 413
 414    if (pool->host_memory_base) {
 415       if (pool->host_memory_end - pool->host_memory_ptr < mem_size)
 416          return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
 417
 418       set = (struct tu_descriptor_set*)pool->host_memory_ptr;
 419       pool->host_memory_ptr += mem_size;
 420    } else {
 421       set = vk_alloc2(&device->alloc, NULL, mem_size, 8,
 422                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 423
 424       if (!set)
 425          return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 426    }
 427
 428    memset(set, 0, mem_size);
 429
 430    if (layout->dynamic_offset_count + layout->input_attachment_count > 0) {
 431       set->dynamic_descriptors = (uint32_t *)((uint8_t*)set + dynamic_offset);
 432    }
 433
 434    set->layout = layout;
 435    set->pool = pool;
 436    uint32_t layout_size = layout->size;
 437    if (variable_count) {
 438       assert(layout->has_variable_descriptors);
 439       uint32_t stride = layout->binding[layout->binding_count - 1].size;
 440       layout_size = layout->binding[layout->binding_count - 1].offset +
 441                     *variable_count * stride;
 442    }
 443
 444    if (layout_size) {
 445       set->size = layout_size;
 446
 447       if (!pool->host_memory_base && pool->entry_count == pool->max_entry_count) {
 448          vk_free2(&device->alloc, NULL, set);
 449          return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
 450       }
 451
 452       /* try to allocate linearly first, so that we don't spend
 453        * time looking for gaps if the app only allocates &
 454        * resets via the pool. */
 455       if (pool->current_offset + layout_size <= pool->size) {
 456          set->mapped_ptr = (uint32_t*)(pool->bo.map + pool->current_offset);
 457          set->va = pool->bo.iova + pool->current_offset;
 458          if (!pool->host_memory_base) {
 459             pool->entries[pool->entry_count].offset = pool->current_offset;
 460             pool->entries[pool->entry_count].size = layout_size;
 461             pool->entries[pool->entry_count].set = set;
 462             pool->entry_count++;
 463          }
 464          pool->current_offset += layout_size;
 465       } else if (!pool->host_memory_base) {
 466          uint64_t offset = 0;
 467          int index;
 468
 469          for (index = 0; index < pool->entry_count; ++index) {
 470             if (pool->entries[index].offset - offset >= layout_size)
 471                break;
 472             offset = pool->entries[index].offset + pool->entries[index].size;
 473          }
 474
 475          if (pool->size - offset < layout_size) {
 476             vk_free2(&device->alloc, NULL, set);
 477             return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
 478          }
 479
 480          set->mapped_ptr = (uint32_t*)(pool->bo.map + offset);
 481          set->va = pool->bo.iova + offset;
 482          memmove(&pool->entries[index + 1], &pool->entries[index],
 483             sizeof(pool->entries[0]) * (pool->entry_count - index));
 484          pool->entries[index].offset = offset;
 485          pool->entries[index].size = layout_size;
 486          pool->entries[index].set = set;
 487          pool->entry_count++;
 488       } else
 489          return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
 490    }
 491
 492    if (layout->has_immutable_samplers) {
 493       for (unsigned i = 0; i < layout->binding_count; ++i) {
 494          if (!layout->binding[i].immutable_samplers_offset)
 495             continue;
 496
 497          unsigned offset = layout->binding[i].offset / 4;
 498          if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
 499             offset += A6XX_TEX_CONST_DWORDS;
 500
 501          const struct tu_sampler *samplers =
 502             (const struct tu_sampler *)((const char *)layout +
 503                                layout->binding[i].immutable_samplers_offset);
 504          for (unsigned j = 0; j < layout->binding[i].array_size; ++j) {
 505             memcpy(set->mapped_ptr + offset, samplers + j,
 506                    sizeof(struct tu_sampler));
 507             offset += layout->binding[i].size / 4;
 508          }
 509       }
 510    }
 511
 512    *out_set = set;
 513    return VK_SUCCESS;
 514 }
 515
 516 static void
 517 tu_descriptor_set_destroy(struct tu_device *device,
 518              struct tu_descriptor_pool *pool,
 519              struct tu_descriptor_set *set,
 520              bool free_bo)
 521 {
 522    assert(!pool->host_memory_base);
 523
 524    if (free_bo && set->size && !pool->host_memory_base) {
 525       uint32_t offset = (uint8_t*)set->mapped_ptr - (uint8_t*)pool->bo.map;
 526       for (int i = 0; i < pool->entry_count; ++i) {
 527          if (pool->entries[i].offset == offset) {
 528             memmove(&pool->entries[i], &pool->entries[i+1],
 529                sizeof(pool->entries[i]) * (pool->entry_count - i - 1));
 530             --pool->entry_count;
 531             break;
 532          }
 533       }
 534    }
 535    vk_free2(&device->alloc, NULL, set);
 536 }
 537
 538 VkResult
 539 tu_CreateDescriptorPool(VkDevice _device,
 540                         const VkDescriptorPoolCreateInfo *pCreateInfo,
 541                         const VkAllocationCallbacks *pAllocator,
 542                         VkDescriptorPool *pDescriptorPool)
 543 {
 544    TU_FROM_HANDLE(tu_device, device, _device);
 545    struct tu_descriptor_pool *pool;
 546    uint64_t size = sizeof(struct tu_descriptor_pool);
 547    uint64_t bo_size = 0, bo_count = 0, dynamic_count = 0;
 548
 549    for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
 550       if (pCreateInfo->pPoolSizes[i].type != VK_DESCRIPTOR_TYPE_SAMPLER)
 551          bo_count += pCreateInfo->pPoolSizes[i].descriptorCount;
 552
 553       switch(pCreateInfo->pPoolSizes[i].type) {
 554       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 555       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 556       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 557          dynamic_count += pCreateInfo->pPoolSizes[i].descriptorCount;
 558       default:
 559          break;
 560       }
 561
 562       bo_size += descriptor_size(pCreateInfo->pPoolSizes[i].type) *
 563                            pCreateInfo->pPoolSizes[i].descriptorCount;
 564    }
 565
 566    if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
 567       uint64_t host_size = pCreateInfo->maxSets * sizeof(struct tu_descriptor_set);
 568       host_size += sizeof(struct tu_bo*) * bo_count;
 569       host_size += A6XX_TEX_CONST_DWORDS * 4 * dynamic_count;
 570       size += host_size;
 571    } else {
 572       size += sizeof(struct tu_descriptor_pool_entry) * pCreateInfo->maxSets;
 573    }
 574
 575    pool = vk_alloc2(&device->alloc, pAllocator, size, 8,
 576                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 577    if (!pool)
 578       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 579
 580    memset(pool, 0, sizeof(*pool));
 581
 582    if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
 583       pool->host_memory_base = (uint8_t*)pool + sizeof(struct tu_descriptor_pool);
 584       pool->host_memory_ptr = pool->host_memory_base;
 585       pool->host_memory_end = (uint8_t*)pool + size;
 586    }
 587
 588    if (bo_size) {
 589       VkResult ret;
 590
 591       ret = tu_bo_init_new(device, &pool->bo, bo_size);
 592       assert(ret == VK_SUCCESS);
 593
 594       ret = tu_bo_map(device, &pool->bo);
 595       assert(ret == VK_SUCCESS);
 596    }
 597    pool->size = bo_size;
 598    pool->max_entry_count = pCreateInfo->maxSets;
 599
 600    *pDescriptorPool = tu_descriptor_pool_to_handle(pool);
 601    return VK_SUCCESS;
 602 }
 603
 604 void
 605 tu_DestroyDescriptorPool(VkDevice _device,
 606                          VkDescriptorPool _pool,
 607                          const VkAllocationCallbacks *pAllocator)
 608 {
 609    TU_FROM_HANDLE(tu_device, device, _device);
 610    TU_FROM_HANDLE(tu_descriptor_pool, pool, _pool);
 611
 612    if (!pool)
 613       return;
 614
 615    if (!pool->host_memory_base) {
 616       for(int i = 0; i < pool->entry_count; ++i) {
 617          tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
 618       }
 619    }
 620
 621    if (pool->size)
 622       tu_bo_finish(device, &pool->bo);
 623    vk_free2(&device->alloc, pAllocator, pool);
 624 }
 625
 626 VkResult
 627 tu_ResetDescriptorPool(VkDevice _device,
 628                        VkDescriptorPool descriptorPool,
 629                        VkDescriptorPoolResetFlags flags)
 630 {
 631    TU_FROM_HANDLE(tu_device, device, _device);
 632    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
 633
 634    if (!pool->host_memory_base) {
 635       for(int i = 0; i < pool->entry_count; ++i) {
 636          tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
 637       }
 638       pool->entry_count = 0;
 639    }
 640
 641    pool->current_offset = 0;
 642    pool->host_memory_ptr = pool->host_memory_base;
 643
 644    return VK_SUCCESS;
 645 }
 646
 647 VkResult
 648 tu_AllocateDescriptorSets(VkDevice _device,
 649                           const VkDescriptorSetAllocateInfo *pAllocateInfo,
 650                           VkDescriptorSet *pDescriptorSets)
 651 {
 652    TU_FROM_HANDLE(tu_device, device, _device);
 653    TU_FROM_HANDLE(tu_descriptor_pool, pool, pAllocateInfo->descriptorPool);
 654
 655    VkResult result = VK_SUCCESS;
 656    uint32_t i;
 657    struct tu_descriptor_set *set = NULL;
 658
 659    const VkDescriptorSetVariableDescriptorCountAllocateInfoEXT *variable_counts =
 660       vk_find_struct_const(pAllocateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_ALLOCATE_INFO_EXT);
 661    const uint32_t zero = 0;
 662
 663    /* allocate a set of buffers for each shader to contain descriptors */
 664    for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
 665       TU_FROM_HANDLE(tu_descriptor_set_layout, layout,
 666              pAllocateInfo->pSetLayouts[i]);
 667
 668       const uint32_t *variable_count = NULL;
 669       if (variable_counts) {
 670          if (i < variable_counts->descriptorSetCount)
 671             variable_count = variable_counts->pDescriptorCounts + i;
 672          else
 673             variable_count = &zero;
 674       }
 675
 676       assert(!(layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
 677
 678       result = tu_descriptor_set_create(device, pool, layout, variable_count, &set);
 679       if (result != VK_SUCCESS)
 680          break;
 681
 682       pDescriptorSets[i] = tu_descriptor_set_to_handle(set);
 683    }
 684
 685    if (result != VK_SUCCESS) {
 686       tu_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool,
 687                i, pDescriptorSets);
 688       for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
 689          pDescriptorSets[i] = VK_NULL_HANDLE;
 690       }
 691    }
 692    return result;
 693 }
 694
 695 VkResult
 696 tu_FreeDescriptorSets(VkDevice _device,
 697                       VkDescriptorPool descriptorPool,
 698                       uint32_t count,
 699                       const VkDescriptorSet *pDescriptorSets)
 700 {
 701    TU_FROM_HANDLE(tu_device, device, _device);
 702    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
 703
 704    for (uint32_t i = 0; i < count; i++) {
 705       TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
 706
 707       if (set && !pool->host_memory_base)
 708          tu_descriptor_set_destroy(device, pool, set, true);
 709    }
 710    return VK_SUCCESS;
 711 }
 712
 713 static void write_texel_buffer_descriptor(struct tu_device *device,
 714                                           struct tu_cmd_buffer *cmd_buffer,
 715                                           unsigned *dst,
 716                                           struct tu_bo **buffer_list,
 717                                           const VkBufferView buffer_view)
 718 {
 719    TU_FROM_HANDLE(tu_buffer_view, view, buffer_view);
 720
 721    memcpy(dst, view->descriptor, sizeof(view->descriptor));
 722
 723    if (cmd_buffer)
 724       tu_bo_list_add(&cmd_buffer->bo_list, view->buffer->bo, MSM_SUBMIT_BO_READ);
 725    else
 726       *buffer_list = view->buffer->bo;
 727 }
 728
 729 static uint32_t get_range(struct tu_buffer *buf, VkDeviceSize offset,
 730                           VkDeviceSize range)
 731 {
 732    if (range == VK_WHOLE_SIZE) {
 733       return buf->size - offset;
 734    } else {
 735       return range;
 736    }
 737 }
 738
 739 static void write_buffer_descriptor(struct tu_device *device,
 740                                     struct tu_cmd_buffer *cmd_buffer,
 741                                     unsigned *dst,
 742                                     struct tu_bo **buffer_list,
 743                                     const VkDescriptorBufferInfo *buffer_info)
 744 {
 745    TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
 746
 747    uint64_t va = tu_buffer_iova(buffer) + buffer_info->offset;
 748    uint32_t range = get_range(buffer, buffer_info->offset, buffer_info->range);
 749    range = ALIGN_POT(range, 4) / 4;
 750    dst[0] =
 751       A6XX_IBO_0_TILE_MODE(TILE6_LINEAR) | A6XX_IBO_0_FMT(FMT6_32_UINT);
 752    dst[1] = range;
 753    dst[2] =
 754       A6XX_IBO_2_UNK4 | A6XX_IBO_2_TYPE(A6XX_TEX_1D) | A6XX_IBO_2_UNK31;
 755    dst[3] = 0;
 756    dst[4] = A6XX_IBO_4_BASE_LO(va);
 757    dst[5] = A6XX_IBO_5_BASE_HI(va >> 32);
 758    for (int i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
 759       dst[i] = 0;
 760
 761    if (cmd_buffer)
 762       tu_bo_list_add(&cmd_buffer->bo_list, buffer->bo, MSM_SUBMIT_BO_READ);
 763    else
 764       *buffer_list = buffer->bo;
 765 }
 766
 767 static void write_ubo_descriptor(struct tu_device *device,
 768                                  struct tu_cmd_buffer *cmd_buffer,
 769                                  unsigned *dst,
 770                                  struct tu_bo **buffer_list,
 771                                  const VkDescriptorBufferInfo *buffer_info)
 772 {
 773    TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
 774
 775    uint32_t range = get_range(buffer, buffer_info->offset, buffer_info->range);
 776    /* The HW range is in vec4 units */
 777    range = ALIGN_POT(range, 16) / 16;
 778    uint64_t va = tu_buffer_iova(buffer) + buffer_info->offset;
 779    dst[0] = A6XX_UBO_0_BASE_LO(va);
 780    dst[1] = A6XX_UBO_1_BASE_HI(va >> 32) | A6XX_UBO_1_SIZE(range);
 781
 782    if (cmd_buffer)
 783       tu_bo_list_add(&cmd_buffer->bo_list, buffer->bo, MSM_SUBMIT_BO_READ);
 784    else
 785       *buffer_list = buffer->bo;
 786 }
 787
 788 static void
 789 write_image_descriptor(struct tu_device *device,
 790              struct tu_cmd_buffer *cmd_buffer,
 791              unsigned *dst,
 792              struct tu_bo **buffer_list,
 793              VkDescriptorType descriptor_type,
 794              const VkDescriptorImageInfo *image_info)
 795 {
 796    TU_FROM_HANDLE(tu_image_view, iview, image_info->imageView);
 797
 798    if (descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) {
 799       memcpy(dst, iview->storage_descriptor, sizeof(iview->storage_descriptor));
 800    } else {
 801       memcpy(dst, iview->descriptor, sizeof(iview->descriptor));
 802    }
 803
 804    if (cmd_buffer)
 805       tu_bo_list_add(&cmd_buffer->bo_list, iview->image->bo, MSM_SUBMIT_BO_READ);
 806    else
 807       *buffer_list = iview->image->bo;
 808 }
 809
 810 static void
 811 write_combined_image_sampler_descriptor(struct tu_device *device,
 812                struct tu_cmd_buffer *cmd_buffer,
 813                unsigned sampler_offset,
 814                unsigned *dst,
 815                struct tu_bo **buffer_list,
 816                VkDescriptorType descriptor_type,
 817                const VkDescriptorImageInfo *image_info,
 818                bool has_sampler)
 819 {
 820    TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
 821
 822    write_image_descriptor(device, cmd_buffer, dst, buffer_list,
 823                           descriptor_type, image_info);
 824    /* copy over sampler state */
 825    if (has_sampler) {
 826       memcpy(dst + A6XX_TEX_CONST_DWORDS, sampler, sizeof(*sampler));
 827    }
 828 }
 829
 830 static void
 831 write_sampler_descriptor(struct tu_device *device,
 832                          unsigned *dst,
 833                          const VkDescriptorImageInfo *image_info)
 834 {
 835    TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
 836
 837    memcpy(dst, sampler, sizeof(*sampler));
 838 }
 839
 840 void
 841 tu_update_descriptor_sets(struct tu_device *device,
 842                           struct tu_cmd_buffer *cmd_buffer,
 843                           VkDescriptorSet dstSetOverride,
 844                           uint32_t descriptorWriteCount,
 845                           const VkWriteDescriptorSet *pDescriptorWrites,
 846                           uint32_t descriptorCopyCount,
 847                           const VkCopyDescriptorSet *pDescriptorCopies)
 848 {
 849    uint32_t i, j;
 850    for (i = 0; i < descriptorWriteCount; i++) {
 851       const VkWriteDescriptorSet *writeset = &pDescriptorWrites[i];
 852       TU_FROM_HANDLE(tu_descriptor_set, set,
 853                        dstSetOverride ? dstSetOverride : writeset->dstSet);
 854       const struct tu_descriptor_set_binding_layout *binding_layout =
 855          set->layout->binding + writeset->dstBinding;
 856       uint32_t *ptr = set->mapped_ptr;
 857       struct tu_bo **buffer_list = set->buffers;
 858
 859       ptr += binding_layout->offset / 4;
 860
 861       ptr += (binding_layout->size / 4) * writeset->dstArrayElement;
 862       buffer_list += binding_layout->buffer_offset;
 863       buffer_list += writeset->dstArrayElement;
 864       for (j = 0; j < writeset->descriptorCount; ++j) {
 865          switch(writeset->descriptorType) {
 866          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: {
 867             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
 868             unsigned idx = writeset->dstArrayElement + j;
 869             idx += set->layout->input_attachment_count + binding_layout->dynamic_offset_offset;
 870             write_ubo_descriptor(device, cmd_buffer,
 871                                  set->dynamic_descriptors + A6XX_TEX_CONST_DWORDS * idx,
 872                                  buffer_list, writeset->pBufferInfo + j);
 873             break;
 874          }
 875          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 876             write_ubo_descriptor(device, cmd_buffer, ptr, buffer_list,
 877                      writeset->pBufferInfo + j);
 878             break;
 879          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
 880             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
 881             unsigned idx = writeset->dstArrayElement + j;
 882             idx += set->layout->input_attachment_count + binding_layout->dynamic_offset_offset;
 883             write_buffer_descriptor(device, cmd_buffer,
 884                                     set->dynamic_descriptors + A6XX_TEX_CONST_DWORDS * idx,
 885                                     buffer_list, writeset->pBufferInfo + j);
 886             break;
 887          }
 888          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 889             write_buffer_descriptor(device, cmd_buffer, ptr, buffer_list,
 890                      writeset->pBufferInfo + j);
 891             break;
 892          case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 893          case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 894             write_texel_buffer_descriptor(device, cmd_buffer, ptr, buffer_list,
 895                      writeset->pTexelBufferView[j]);
 896             break;
 897          case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 898          case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
 899             write_image_descriptor(device, cmd_buffer, ptr, buffer_list,
 900                                    writeset->descriptorType,
 901                                    writeset->pImageInfo + j);
 902             break;
 903          case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: {
 904             unsigned idx = writeset->dstArrayElement + j;
 905             idx += binding_layout->input_attachment_offset;
 906             write_image_descriptor(device, cmd_buffer,
 907                                     set->dynamic_descriptors + A6XX_TEX_CONST_DWORDS * idx,
 908                                     buffer_list, writeset->descriptorType,
 909                                     writeset->pImageInfo + j);
 910             break;
 911          }
 912          case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 913             write_combined_image_sampler_descriptor(device, cmd_buffer,
 914                                                     A6XX_TEX_CONST_DWORDS * 4,
 915                                                     ptr, buffer_list,
 916                                                     writeset->descriptorType,
 917                                                     writeset->pImageInfo + j,
 918                                                     !binding_layout->immutable_samplers_offset);
 919             break;
 920          case VK_DESCRIPTOR_TYPE_SAMPLER:
 921             write_sampler_descriptor(device, ptr, writeset->pImageInfo + j);
 922             break;
 923          default:
 924             unreachable("unimplemented descriptor type");
 925             break;
 926          }
 927          ptr += binding_layout->size / 4;
 928          ++buffer_list;
 929       }
 930    }
 931
 932    for (i = 0; i < descriptorCopyCount; i++) {
 933       const VkCopyDescriptorSet *copyset = &pDescriptorCopies[i];
 934       TU_FROM_HANDLE(tu_descriptor_set, src_set,
 935                        copyset->srcSet);
 936       TU_FROM_HANDLE(tu_descriptor_set, dst_set,
 937                        copyset->dstSet);
 938       const struct tu_descriptor_set_binding_layout *src_binding_layout =
 939          src_set->layout->binding + copyset->srcBinding;
 940       const struct tu_descriptor_set_binding_layout *dst_binding_layout =
 941          dst_set->layout->binding + copyset->dstBinding;
 942       uint32_t *src_ptr = src_set->mapped_ptr;
 943       uint32_t *dst_ptr = dst_set->mapped_ptr;
 944       struct tu_bo **src_buffer_list = src_set->buffers;
 945       struct tu_bo **dst_buffer_list = dst_set->buffers;
 946
 947       src_ptr += src_binding_layout->offset / 4;
 948       dst_ptr += dst_binding_layout->offset / 4;
 949
 950       src_ptr += src_binding_layout->size * copyset->srcArrayElement / 4;
 951       dst_ptr += dst_binding_layout->size * copyset->dstArrayElement / 4;
 952
 953       src_buffer_list += src_binding_layout->buffer_offset;
 954       src_buffer_list += copyset->srcArrayElement;
 955
 956       dst_buffer_list += dst_binding_layout->buffer_offset;
 957       dst_buffer_list += copyset->dstArrayElement;
 958
 959       for (j = 0; j < copyset->descriptorCount; ++j) {
 960          switch (src_binding_layout->type) {
 961          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 962          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
 963             unsigned src_idx = copyset->srcArrayElement + j;
 964             unsigned dst_idx = copyset->dstArrayElement + j;
 965             src_idx += src_set->layout->input_attachment_count;
 966             dst_idx += dst_set->layout->input_attachment_count;
 967             src_idx += src_binding_layout->dynamic_offset_offset;
 968             dst_idx += dst_binding_layout->dynamic_offset_offset;
 969
 970             uint32_t *src_dynamic, *dst_dynamic;
 971             src_dynamic = src_set->dynamic_descriptors + src_idx * A6XX_TEX_CONST_DWORDS;
 972             dst_dynamic = dst_set->dynamic_descriptors + dst_idx * A6XX_TEX_CONST_DWORDS;
 973             memcpy(dst_dynamic, src_dynamic, A6XX_TEX_CONST_DWORDS * 4);
 974             break;
 975          }
 976          case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: {
 977             unsigned src_idx = copyset->srcArrayElement + j;
 978             unsigned dst_idx = copyset->dstArrayElement + j;
 979             src_idx += src_binding_layout->input_attachment_offset;
 980             dst_idx += dst_binding_layout->input_attachment_offset;
 981
 982             uint32_t *src_dynamic, *dst_dynamic;
 983             src_dynamic = src_set->dynamic_descriptors + src_idx * A6XX_TEX_CONST_DWORDS;
 984             dst_dynamic = dst_set->dynamic_descriptors + dst_idx * A6XX_TEX_CONST_DWORDS;
 985             memcpy(dst_dynamic, src_dynamic, A6XX_TEX_CONST_DWORDS * 4);
 986             break;
 987          }
 988          default:
 989             memcpy(dst_ptr, src_ptr, src_binding_layout->size);
 990          }
 991
 992          src_ptr += src_binding_layout->size / 4;
 993          dst_ptr += dst_binding_layout->size / 4;
 994
 995          if (src_binding_layout->type != VK_DESCRIPTOR_TYPE_SAMPLER) {
 996             /* Sampler descriptors don't have a buffer list. */
 997             dst_buffer_list[j] = src_buffer_list[j];
 998          }
 999       }
1000    }
1001 }
1002
1003 void
1004 tu_UpdateDescriptorSets(VkDevice _device,
1005                         uint32_t descriptorWriteCount,
1006                         const VkWriteDescriptorSet *pDescriptorWrites,
1007                         uint32_t descriptorCopyCount,
1008                         const VkCopyDescriptorSet *pDescriptorCopies)
1009 {
1010    TU_FROM_HANDLE(tu_device, device, _device);
1011
1012    tu_update_descriptor_sets(device, NULL, VK_NULL_HANDLE,
1013                              descriptorWriteCount, pDescriptorWrites,
1014                              descriptorCopyCount, pDescriptorCopies);
1015 }
1016
1017 VkResult
1018 tu_CreateDescriptorUpdateTemplate(
1019    VkDevice _device,
1020    const VkDescriptorUpdateTemplateCreateInfo *pCreateInfo,
1021    const VkAllocationCallbacks *pAllocator,
1022    VkDescriptorUpdateTemplate *pDescriptorUpdateTemplate)
1023 {
1024    TU_FROM_HANDLE(tu_device, device, _device);
1025    TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
1026                   pCreateInfo->descriptorSetLayout);
1027    const uint32_t entry_count = pCreateInfo->descriptorUpdateEntryCount;
1028    const size_t size =
1029       sizeof(struct tu_descriptor_update_template) +
1030       sizeof(struct tu_descriptor_update_template_entry) * entry_count;
1031    struct tu_descriptor_update_template *templ;
1032
1033    templ = vk_alloc2(&device->alloc, pAllocator, size, 8,
1034                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1035    if (!templ)
1036       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1037
1038    templ->entry_count = entry_count;
1039
1040    if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) {
1041       TU_FROM_HANDLE(tu_pipeline_layout, pipeline_layout, pCreateInfo->pipelineLayout);
1042
1043       /* descriptorSetLayout should be ignored for push descriptors
1044        * and instead it refers to pipelineLayout and set.
1045        */
1046       assert(pCreateInfo->set < MAX_SETS);
1047       set_layout = pipeline_layout->set[pCreateInfo->set].layout;
1048    }
1049
1050    for (uint32_t i = 0; i < entry_count; i++) {
1051       const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i];
1052
1053       const struct tu_descriptor_set_binding_layout *binding_layout =
1054          set_layout->binding + entry->dstBinding;
1055       const uint32_t buffer_offset = binding_layout->buffer_offset +
1056          entry->dstArrayElement;
1057       uint32_t dst_offset, dst_stride;
1058
1059       /* dst_offset is an offset into dynamic_descriptors when the descriptor
1060        * is dynamic, and an offset into mapped_ptr otherwise.
1061        */
1062       switch (entry->descriptorType) {
1063       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1064       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
1065          dst_offset = (set_layout->input_attachment_count +
1066             binding_layout->dynamic_offset_offset +
1067             entry->dstArrayElement) * A6XX_TEX_CONST_DWORDS;
1068          dst_stride = A6XX_TEX_CONST_DWORDS;
1069          break;
1070       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
1071          dst_offset = (binding_layout->input_attachment_offset +
1072             entry->dstArrayElement) * A6XX_TEX_CONST_DWORDS;
1073          dst_stride = A6XX_TEX_CONST_DWORDS;
1074          break;
1075       default:
1076          dst_offset = binding_layout->offset / 4;
1077          dst_offset += (binding_layout->size * entry->dstArrayElement) / 4;
1078          dst_stride = binding_layout->size / 4;
1079       }
1080
1081       templ->entry[i] = (struct tu_descriptor_update_template_entry) {
1082          .descriptor_type = entry->descriptorType,
1083          .descriptor_count = entry->descriptorCount,
1084          .src_offset = entry->offset,
1085          .src_stride = entry->stride,
1086          .dst_offset = dst_offset,
1087          .dst_stride = dst_stride,
1088          .buffer_offset = buffer_offset,
1089          .has_sampler = !binding_layout->immutable_samplers_offset,
1090       };
1091    }
1092
1093    *pDescriptorUpdateTemplate =
1094       tu_descriptor_update_template_to_handle(templ);
1095
1096    return VK_SUCCESS;
1097 }
1098
1099 void
1100 tu_DestroyDescriptorUpdateTemplate(
1101    VkDevice _device,
1102    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1103    const VkAllocationCallbacks *pAllocator)
1104 {
1105    TU_FROM_HANDLE(tu_device, device, _device);
1106    TU_FROM_HANDLE(tu_descriptor_update_template, templ,
1107                   descriptorUpdateTemplate);
1108
1109    if (!templ)
1110       return;
1111
1112    vk_free2(&device->alloc, pAllocator, templ);
1113 }
1114
1115 void
1116 tu_update_descriptor_set_with_template(
1117    struct tu_device *device,
1118    struct tu_cmd_buffer *cmd_buffer,
1119    struct tu_descriptor_set *set,
1120    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1121    const void *pData)
1122 {
1123    TU_FROM_HANDLE(tu_descriptor_update_template, templ,
1124                   descriptorUpdateTemplate);
1125
1126    for (uint32_t i = 0; i < templ->entry_count; i++) {
1127       uint32_t *ptr = set->mapped_ptr;
1128       const void *src = ((const char *) pData) + templ->entry[i].src_offset;
1129       struct tu_bo **buffer_list = set->buffers;
1130
1131       ptr += templ->entry[i].dst_offset;
1132       buffer_list += templ->entry[i].buffer_offset;
1133       unsigned dst_offset = templ->entry[i].dst_offset;
1134       for (unsigned j = 0; j < templ->entry[i].descriptor_count; ++j) {
1135          switch(templ->entry[i].descriptor_type) {
1136          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: {
1137             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
1138             write_ubo_descriptor(device, cmd_buffer,
1139                                  set->dynamic_descriptors + dst_offset,
1140                                  buffer_list, src);
1141             break;
1142          }
1143          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
1144             write_ubo_descriptor(device, cmd_buffer, ptr, buffer_list, src);
1145             break;
1146          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
1147             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
1148             write_buffer_descriptor(device, cmd_buffer,
1149                                     set->dynamic_descriptors + dst_offset,
1150                                     buffer_list, src);
1151             break;
1152          }
1153          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
1154             write_buffer_descriptor(device, cmd_buffer, ptr, buffer_list, src);
1155             break;
1156          case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
1157          case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
1158             write_texel_buffer_descriptor(device, cmd_buffer, ptr,
1159                                           buffer_list, *(VkBufferView *) src);
1160             break;
1161          case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
1162          case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
1163             write_image_descriptor(device, cmd_buffer, ptr, buffer_list,
1164                                    templ->entry[i].descriptor_type,
1165                                    src);
1166             break;
1167          case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: {
1168             write_image_descriptor(device, cmd_buffer,
1169                                     set->dynamic_descriptors + dst_offset,
1170                                     buffer_list, templ->entry[i].descriptor_type,
1171                                     src);
1172             break;
1173          }
1174          case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
1175             write_combined_image_sampler_descriptor(device, cmd_buffer,
1176                                                     A6XX_TEX_CONST_DWORDS * 4,
1177                                                     ptr, buffer_list,
1178                                                     templ->entry[i].descriptor_type,
1179                                                     src,
1180                                                     templ->entry[i].has_sampler);
1181             break;
1182          case VK_DESCRIPTOR_TYPE_SAMPLER:
1183             write_sampler_descriptor(device, ptr, src);
1184             break;
1185          default:
1186             unreachable("unimplemented descriptor type");
1187             break;
1188          }
1189          src = (char *) src + templ->entry[i].src_stride;
1190          ptr += templ->entry[i].dst_stride;
1191          dst_offset += templ->entry[i].dst_stride;
1192          ++buffer_list;
1193       }
1194    }
1195 }
1196
1197 void
1198 tu_UpdateDescriptorSetWithTemplate(
1199    VkDevice _device,
1200    VkDescriptorSet descriptorSet,
1201    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1202    const void *pData)
1203 {
1204    TU_FROM_HANDLE(tu_device, device, _device);
1205    TU_FROM_HANDLE(tu_descriptor_set, set, descriptorSet);
1206
1207    tu_update_descriptor_set_with_template(device, NULL, set,
1208                                           descriptorUpdateTemplate, pData);
1209 }
1210
1211 VkResult
1212 tu_CreateSamplerYcbcrConversion(
1213    VkDevice device,
1214    const VkSamplerYcbcrConversionCreateInfo *pCreateInfo,
1215    const VkAllocationCallbacks *pAllocator,
1216    VkSamplerYcbcrConversion *pYcbcrConversion)
1217 {
1218    *pYcbcrConversion = VK_NULL_HANDLE;
1219    return VK_SUCCESS;
1220 }
1221
1222 void
1223 tu_DestroySamplerYcbcrConversion(VkDevice device,
1224                                  VkSamplerYcbcrConversion ycbcrConversion,
1225                                  const VkAllocationCallbacks *pAllocator)
1226 {
1227    /* Do nothing. */
1228 }