Home | History | Annotate | Line # | Download | only in vulkan
      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_descriptors.h"
     48 #include "vk_util.h"
     49 
     50 static inline uint8_t *
     51 pool_base(struct tu_descriptor_pool *pool)
     52 {
     53    return pool->host_bo ?: pool->bo.map;
     54 }
     55 
     56 static uint32_t
     57 descriptor_size(VkDescriptorType type)
     58 {
     59    switch (type) {
     60    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
     61    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
     62    case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
     63       /* These are remapped to the special driver-managed descriptor set,
     64        * hence they don't take up any space in the original descriptor set:
     65        * Input attachment doesn't use descriptor sets at all
     66        */
     67       return 0;
     68    case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
     69       /* We make offsets and sizes all 16 dwords, to match how the hardware
     70        * interprets indices passed to sample/load/store instructions in
     71        * multiples of 16 dwords.  This means that "normal" descriptors are all
     72        * of size 16, with padding for smaller descriptors like uniform storage
     73        * descriptors which are less than 16 dwords. However combined images
     74        * and samplers are actually two descriptors, so they have size 2.
     75        */
     76       return A6XX_TEX_CONST_DWORDS * 4 * 2;
     77    default:
     78       return A6XX_TEX_CONST_DWORDS * 4;
     79    }
     80 }
     81 
     82 static uint32_t
     83 mutable_descriptor_size(const VkMutableDescriptorTypeListVALVE *list)
     84 {
     85    uint32_t max_size = 0;
     86 
     87    /* Since we don't support VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER for
     88     * mutable descriptors, max_size should be always A6XX_TEX_CONST_DWORDS * 4.
     89     * But we leave this as-is and add an assert.
     90     */
     91    for (uint32_t i = 0; i < list->descriptorTypeCount; i++) {
     92       uint32_t size = descriptor_size(list->pDescriptorTypes[i]);
     93       max_size = MAX2(max_size, size);
     94    }
     95 
     96    assert(max_size == A6XX_TEX_CONST_DWORDS * 4);
     97 
     98    return max_size;
     99 }
    100 
    101 VKAPI_ATTR VkResult VKAPI_CALL
    102 tu_CreateDescriptorSetLayout(
    103    VkDevice _device,
    104    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
    105    const VkAllocationCallbacks *pAllocator,
    106    VkDescriptorSetLayout *pSetLayout)
    107 {
    108    TU_FROM_HANDLE(tu_device, device, _device);
    109    struct tu_descriptor_set_layout *set_layout;
    110 
    111    assert(pCreateInfo->sType ==
    112           VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO);
    113    const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
    114       vk_find_struct_const(
    115          pCreateInfo->pNext,
    116          DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
    117    const VkMutableDescriptorTypeCreateInfoVALVE *mutable_info =
    118       vk_find_struct_const(
    119          pCreateInfo->pNext,
    120          MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_VALVE);
    121 
    122    uint32_t num_bindings = 0;
    123    uint32_t immutable_sampler_count = 0;
    124    uint32_t ycbcr_sampler_count = 0;
    125    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
    126       num_bindings = MAX2(num_bindings, pCreateInfo->pBindings[j].binding + 1);
    127       if ((pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
    128            pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
    129            pCreateInfo->pBindings[j].pImmutableSamplers) {
    130          immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
    131 
    132          bool has_ycbcr_sampler = false;
    133          for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) {
    134             if (tu_sampler_from_handle(pCreateInfo->pBindings[j].pImmutableSamplers[i])->ycbcr_sampler)
    135                has_ycbcr_sampler = true;
    136          }
    137 
    138          if (has_ycbcr_sampler)
    139             ycbcr_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
    140       }
    141    }
    142 
    143    uint32_t samplers_offset =
    144          offsetof(struct tu_descriptor_set_layout, binding[num_bindings]);
    145 
    146    /* note: only need to store TEX_SAMP_DWORDS for immutable samples,
    147     * but using struct tu_sampler makes things simpler */
    148    uint32_t size = samplers_offset +
    149       immutable_sampler_count * sizeof(struct tu_sampler) +
    150       ycbcr_sampler_count * sizeof(struct tu_sampler_ycbcr_conversion);
    151 
    152    set_layout = vk_object_zalloc(&device->vk, pAllocator, size,
    153                                  VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT);
    154    if (!set_layout)
    155       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
    156 
    157    set_layout->flags = pCreateInfo->flags;
    158 
    159    /* We just allocate all the immutable samplers at the end of the struct */
    160    struct tu_sampler *samplers = (void*) &set_layout->binding[num_bindings];
    161    struct tu_sampler_ycbcr_conversion *ycbcr_samplers =
    162       (void*) &samplers[immutable_sampler_count];
    163 
    164    VkDescriptorSetLayoutBinding *bindings = NULL;
    165    VkResult result = vk_create_sorted_bindings(
    166       pCreateInfo->pBindings, pCreateInfo->bindingCount, &bindings);
    167    if (result != VK_SUCCESS) {
    168       vk_object_free(&device->vk, pAllocator, set_layout);
    169       return vk_error(device, result);
    170    }
    171 
    172    set_layout->binding_count = num_bindings;
    173    set_layout->shader_stages = 0;
    174    set_layout->has_immutable_samplers = false;
    175    set_layout->size = 0;
    176    set_layout->dynamic_ubo = 0;
    177 
    178    uint32_t dynamic_offset_count = 0;
    179 
    180    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
    181       const VkDescriptorSetLayoutBinding *binding = bindings + j;
    182       uint32_t b = binding->binding;
    183 
    184       set_layout->binding[b].type = binding->descriptorType;
    185       set_layout->binding[b].array_size = binding->descriptorCount;
    186       set_layout->binding[b].offset = set_layout->size;
    187       set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
    188       set_layout->binding[b].shader_stages = binding->stageFlags;
    189 
    190       if (binding->descriptorType == VK_DESCRIPTOR_TYPE_MUTABLE_VALVE) {
    191          /* For mutable descriptor types we must allocate a size that fits the
    192           * largest descriptor type that the binding can mutate to.
    193           */
    194          set_layout->binding[b].size =
    195             mutable_descriptor_size(&mutable_info->pMutableDescriptorTypeLists[j]);
    196       } else {
    197          set_layout->binding[b].size = descriptor_size(binding->descriptorType);
    198       }
    199 
    200       if (variable_flags && binding->binding < variable_flags->bindingCount &&
    201           (variable_flags->pBindingFlags[binding->binding] &
    202            VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
    203          assert(!binding->pImmutableSamplers); /* Terribly ill defined  how
    204                                                   many samplers are valid */
    205          assert(binding->binding == num_bindings - 1);
    206 
    207          set_layout->has_variable_descriptors = true;
    208       }
    209 
    210       if ((binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
    211            binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
    212           binding->pImmutableSamplers) {
    213          set_layout->binding[b].immutable_samplers_offset = samplers_offset;
    214          set_layout->has_immutable_samplers = true;
    215 
    216          for (uint32_t i = 0; i < binding->descriptorCount; i++)
    217             samplers[i] = *tu_sampler_from_handle(binding->pImmutableSamplers[i]);
    218 
    219          samplers += binding->descriptorCount;
    220          samplers_offset += sizeof(struct tu_sampler) * binding->descriptorCount;
    221 
    222          bool has_ycbcr_sampler = false;
    223          for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) {
    224             if (tu_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler)
    225                has_ycbcr_sampler = true;
    226          }
    227 
    228          if (has_ycbcr_sampler) {
    229             set_layout->binding[b].ycbcr_samplers_offset =
    230                (const char*)ycbcr_samplers - (const char*)set_layout;
    231             for (uint32_t i = 0; i < binding->descriptorCount; i++) {
    232                struct tu_sampler *sampler = tu_sampler_from_handle(binding->pImmutableSamplers[i]);
    233                if (sampler->ycbcr_sampler)
    234                   ycbcr_samplers[i] = *sampler->ycbcr_sampler;
    235                else
    236                   ycbcr_samplers[i].ycbcr_model = VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY;
    237             }
    238             ycbcr_samplers += binding->descriptorCount;
    239          } else {
    240             set_layout->binding[b].ycbcr_samplers_offset = 0;
    241          }
    242       }
    243 
    244       set_layout->size +=
    245          binding->descriptorCount * set_layout->binding[b].size;
    246       if (binding->descriptorType == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC ||
    247           binding->descriptorType == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) {
    248          if (binding->descriptorType == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) {
    249             STATIC_ASSERT(MAX_DYNAMIC_BUFFERS <= 8 * sizeof(set_layout->dynamic_ubo));
    250             set_layout->dynamic_ubo |=
    251                ((1u << binding->descriptorCount) - 1) << dynamic_offset_count;
    252          }
    253 
    254          dynamic_offset_count += binding->descriptorCount;
    255       }
    256 
    257       set_layout->shader_stages |= binding->stageFlags;
    258    }
    259 
    260    free(bindings);
    261 
    262    set_layout->dynamic_offset_count = dynamic_offset_count;
    263 
    264    *pSetLayout = tu_descriptor_set_layout_to_handle(set_layout);
    265 
    266    return VK_SUCCESS;
    267 }
    268 
    269 VKAPI_ATTR void VKAPI_CALL
    270 tu_DestroyDescriptorSetLayout(VkDevice _device,
    271                               VkDescriptorSetLayout _set_layout,
    272                               const VkAllocationCallbacks *pAllocator)
    273 {
    274    TU_FROM_HANDLE(tu_device, device, _device);
    275    TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout, _set_layout);
    276 
    277    if (!set_layout)
    278       return;
    279 
    280    vk_object_free(&device->vk, pAllocator, set_layout);
    281 }
    282 
    283 VKAPI_ATTR void VKAPI_CALL
    284 tu_GetDescriptorSetLayoutSupport(
    285    VkDevice device,
    286    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
    287    VkDescriptorSetLayoutSupport *pSupport)
    288 {
    289    VkDescriptorSetLayoutBinding *bindings = NULL;
    290    VkResult result = vk_create_sorted_bindings(
    291       pCreateInfo->pBindings, pCreateInfo->bindingCount, &bindings);
    292    if (result != VK_SUCCESS) {
    293       pSupport->supported = false;
    294       return;
    295    }
    296 
    297    const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
    298       vk_find_struct_const(
    299          pCreateInfo->pNext,
    300          DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
    301    VkDescriptorSetVariableDescriptorCountLayoutSupportEXT *variable_count =
    302       vk_find_struct(
    303          (void *) pCreateInfo->pNext,
    304          DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT_EXT);
    305    const VkMutableDescriptorTypeCreateInfoVALVE *mutable_info =
    306       vk_find_struct_const(
    307          pCreateInfo->pNext,
    308          MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_VALVE);
    309 
    310    if (variable_count) {
    311       variable_count->maxVariableDescriptorCount = 0;
    312    }
    313 
    314    bool supported = true;
    315    uint64_t size = 0;
    316    for (uint32_t i = 0; i < pCreateInfo->bindingCount; i++) {
    317       const VkDescriptorSetLayoutBinding *binding = bindings + i;
    318 
    319       uint64_t descriptor_sz;
    320 
    321       if (binding->descriptorType == VK_DESCRIPTOR_TYPE_MUTABLE_VALVE) {
    322          const VkMutableDescriptorTypeListVALVE *list =
    323             &mutable_info->pMutableDescriptorTypeLists[i];
    324 
    325          for (uint32_t j = 0; j < list->descriptorTypeCount; j++) {
    326             /* Don't support the input attachement and combined image sampler type
    327              * for mutable descriptors */
    328             if (list->pDescriptorTypes[j] == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT ||
    329                 list->pDescriptorTypes[j] == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
    330                supported = false;
    331                goto out;
    332             }
    333          }
    334 
    335          descriptor_sz =
    336             mutable_descriptor_size(&mutable_info->pMutableDescriptorTypeLists[i]);
    337       } else {
    338          descriptor_sz = descriptor_size(binding->descriptorType);
    339       }
    340       uint64_t descriptor_alignment = 8;
    341 
    342       if (size && !ALIGN_POT(size, descriptor_alignment)) {
    343          supported = false;
    344       }
    345       size = ALIGN_POT(size, descriptor_alignment);
    346 
    347       uint64_t max_count = UINT64_MAX;
    348       if (descriptor_sz)
    349          max_count = (UINT64_MAX - size) / descriptor_sz;
    350 
    351       if (max_count < binding->descriptorCount) {
    352          supported = false;
    353       }
    354 
    355       if (variable_flags && binding->binding < variable_flags->bindingCount &&
    356           variable_count &&
    357           (variable_flags->pBindingFlags[binding->binding] &
    358            VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
    359          variable_count->maxVariableDescriptorCount =
    360             MIN2(UINT32_MAX, max_count);
    361       }
    362       size += binding->descriptorCount * descriptor_sz;
    363    }
    364 
    365 out:
    366    free(bindings);
    367 
    368    pSupport->supported = supported;
    369 }
    370 
    371 /*
    372  * Pipeline layouts.  These have nothing to do with the pipeline.  They are
    373  * just multiple descriptor set layouts pasted together.
    374  */
    375 
    376 VKAPI_ATTR VkResult VKAPI_CALL
    377 tu_CreatePipelineLayout(VkDevice _device,
    378                         const VkPipelineLayoutCreateInfo *pCreateInfo,
    379                         const VkAllocationCallbacks *pAllocator,
    380                         VkPipelineLayout *pPipelineLayout)
    381 {
    382    TU_FROM_HANDLE(tu_device, device, _device);
    383    struct tu_pipeline_layout *layout;
    384 
    385    assert(pCreateInfo->sType ==
    386           VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
    387 
    388    layout = vk_object_alloc(&device->vk, pAllocator, sizeof(*layout),
    389                             VK_OBJECT_TYPE_PIPELINE_LAYOUT);
    390    if (layout == NULL)
    391       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
    392 
    393    layout->num_sets = pCreateInfo->setLayoutCount;
    394    layout->dynamic_offset_count = 0;
    395 
    396    unsigned dynamic_offset_count = 0;
    397 
    398    for (uint32_t set = 0; set < pCreateInfo->setLayoutCount; set++) {
    399       TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
    400                      pCreateInfo->pSetLayouts[set]);
    401       layout->set[set].layout = set_layout;
    402       layout->set[set].dynamic_offset_start = dynamic_offset_count;
    403       dynamic_offset_count += set_layout->dynamic_offset_count;
    404    }
    405 
    406    layout->dynamic_offset_count = dynamic_offset_count;
    407    layout->push_constant_size = 0;
    408 
    409    for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) {
    410       const VkPushConstantRange *range = pCreateInfo->pPushConstantRanges + i;
    411       layout->push_constant_size =
    412          MAX2(layout->push_constant_size, range->offset + range->size);
    413    }
    414 
    415    layout->push_constant_size = align(layout->push_constant_size, 16);
    416    *pPipelineLayout = tu_pipeline_layout_to_handle(layout);
    417 
    418    return VK_SUCCESS;
    419 }
    420 
    421 VKAPI_ATTR void VKAPI_CALL
    422 tu_DestroyPipelineLayout(VkDevice _device,
    423                          VkPipelineLayout _pipelineLayout,
    424                          const VkAllocationCallbacks *pAllocator)
    425 {
    426    TU_FROM_HANDLE(tu_device, device, _device);
    427    TU_FROM_HANDLE(tu_pipeline_layout, pipeline_layout, _pipelineLayout);
    428 
    429    if (!pipeline_layout)
    430       return;
    431 
    432    vk_object_free(&device->vk, pAllocator, pipeline_layout);
    433 }
    434 
    435 #define EMPTY 1
    436 
    437 static VkResult
    438 tu_descriptor_set_create(struct tu_device *device,
    439             struct tu_descriptor_pool *pool,
    440             const struct tu_descriptor_set_layout *layout,
    441             const uint32_t *variable_count,
    442             struct tu_descriptor_set **out_set)
    443 {
    444    struct tu_descriptor_set *set;
    445    unsigned dynamic_offset = sizeof(struct tu_descriptor_set);
    446    unsigned mem_size = dynamic_offset +
    447       A6XX_TEX_CONST_DWORDS * 4 * layout->dynamic_offset_count;
    448 
    449    if (pool->host_memory_base) {
    450       if (pool->host_memory_end - pool->host_memory_ptr < mem_size)
    451          return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
    452 
    453       set = (struct tu_descriptor_set*)pool->host_memory_ptr;
    454       pool->host_memory_ptr += mem_size;
    455    } else {
    456       set = vk_alloc2(&device->vk.alloc, NULL, mem_size, 8,
    457                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    458 
    459       if (!set)
    460          return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
    461    }
    462 
    463    memset(set, 0, mem_size);
    464    vk_object_base_init(&device->vk, &set->base, VK_OBJECT_TYPE_DESCRIPTOR_SET);
    465 
    466    if (layout->dynamic_offset_count) {
    467       set->dynamic_descriptors = (uint32_t *)((uint8_t*)set + dynamic_offset);
    468    }
    469 
    470    set->layout = layout;
    471    set->pool = pool;
    472    uint32_t layout_size = layout->size;
    473    if (variable_count) {
    474       assert(layout->has_variable_descriptors);
    475       uint32_t stride = layout->binding[layout->binding_count - 1].size;
    476       layout_size = layout->binding[layout->binding_count - 1].offset +
    477                     *variable_count * stride;
    478    }
    479 
    480    if (layout_size) {
    481       set->size = layout_size;
    482 
    483       if (!pool->host_memory_base && pool->entry_count == pool->max_entry_count) {
    484          vk_object_free(&device->vk, NULL, set);
    485          return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
    486       }
    487 
    488       /* try to allocate linearly first, so that we don't spend
    489        * time looking for gaps if the app only allocates &
    490        * resets via the pool. */
    491       if (pool->current_offset + layout_size <= pool->size) {
    492          set->mapped_ptr = (uint32_t*)(pool_base(pool) + pool->current_offset);
    493          set->va = pool->host_bo ? 0 : pool->bo.iova + pool->current_offset;
    494 
    495          if (!pool->host_memory_base) {
    496             pool->entries[pool->entry_count].offset = pool->current_offset;
    497             pool->entries[pool->entry_count].size = layout_size;
    498             pool->entries[pool->entry_count].set = set;
    499             pool->entry_count++;
    500          }
    501          pool->current_offset += layout_size;
    502       } else if (!pool->host_memory_base) {
    503          uint64_t offset = 0;
    504          int index;
    505 
    506          for (index = 0; index < pool->entry_count; ++index) {
    507             if (pool->entries[index].offset - offset >= layout_size)
    508                break;
    509             offset = pool->entries[index].offset + pool->entries[index].size;
    510          }
    511 
    512          if (pool->size - offset < layout_size) {
    513             vk_object_free(&device->vk, NULL, set);
    514             return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
    515          }
    516 
    517          set->mapped_ptr = (uint32_t*)(pool_base(pool) + offset);
    518          set->va = pool->host_bo ? 0 : pool->bo.iova + offset;
    519 
    520          memmove(&pool->entries[index + 1], &pool->entries[index],
    521             sizeof(pool->entries[0]) * (pool->entry_count - index));
    522          pool->entries[index].offset = offset;
    523          pool->entries[index].size = layout_size;
    524          pool->entries[index].set = set;
    525          pool->entry_count++;
    526       } else
    527          return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
    528    }
    529 
    530    if (layout->has_immutable_samplers) {
    531       for (unsigned i = 0; i < layout->binding_count; ++i) {
    532          if (!layout->binding[i].immutable_samplers_offset)
    533             continue;
    534 
    535          unsigned offset = layout->binding[i].offset / 4;
    536          if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
    537             offset += A6XX_TEX_CONST_DWORDS;
    538 
    539          const struct tu_sampler *samplers =
    540             (const struct tu_sampler *)((const char *)layout +
    541                                layout->binding[i].immutable_samplers_offset);
    542          for (unsigned j = 0; j < layout->binding[i].array_size; ++j) {
    543             memcpy(set->mapped_ptr + offset, samplers[j].descriptor,
    544                    sizeof(samplers[j].descriptor));
    545             offset += layout->binding[i].size / 4;
    546          }
    547       }
    548    }
    549 
    550    *out_set = set;
    551    return VK_SUCCESS;
    552 }
    553 
    554 static void
    555 tu_descriptor_set_destroy(struct tu_device *device,
    556              struct tu_descriptor_pool *pool,
    557              struct tu_descriptor_set *set,
    558              bool free_bo)
    559 {
    560    assert(!pool->host_memory_base);
    561 
    562    if (free_bo && set->size && !pool->host_memory_base) {
    563       uint32_t offset = (uint8_t*)set->mapped_ptr - pool_base(pool);
    564 
    565       for (int i = 0; i < pool->entry_count; ++i) {
    566          if (pool->entries[i].offset == offset) {
    567             memmove(&pool->entries[i], &pool->entries[i+1],
    568                sizeof(pool->entries[i]) * (pool->entry_count - i - 1));
    569             --pool->entry_count;
    570             break;
    571          }
    572       }
    573    }
    574 
    575    vk_object_free(&device->vk, NULL, set);
    576 }
    577 
    578 VKAPI_ATTR VkResult VKAPI_CALL
    579 tu_CreateDescriptorPool(VkDevice _device,
    580                         const VkDescriptorPoolCreateInfo *pCreateInfo,
    581                         const VkAllocationCallbacks *pAllocator,
    582                         VkDescriptorPool *pDescriptorPool)
    583 {
    584    TU_FROM_HANDLE(tu_device, device, _device);
    585    struct tu_descriptor_pool *pool;
    586    uint64_t size = sizeof(struct tu_descriptor_pool);
    587    uint64_t bo_size = 0, bo_count = 0, dynamic_count = 0;
    588    VkResult ret;
    589 
    590    const VkMutableDescriptorTypeCreateInfoVALVE *mutable_info =
    591       vk_find_struct_const( pCreateInfo->pNext,
    592          MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_VALVE);
    593 
    594    for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
    595       if (pCreateInfo->pPoolSizes[i].type != VK_DESCRIPTOR_TYPE_SAMPLER)
    596          bo_count += pCreateInfo->pPoolSizes[i].descriptorCount;
    597 
    598       switch(pCreateInfo->pPoolSizes[i].type) {
    599       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
    600       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
    601          dynamic_count += pCreateInfo->pPoolSizes[i].descriptorCount;
    602          break;
    603       case VK_DESCRIPTOR_TYPE_MUTABLE_VALVE:
    604          if (mutable_info && i < mutable_info->mutableDescriptorTypeListCount &&
    605              mutable_info->pMutableDescriptorTypeLists[i].descriptorTypeCount > 0) {
    606             bo_size +=
    607                mutable_descriptor_size(&mutable_info->pMutableDescriptorTypeLists[i]) *
    608                   pCreateInfo->pPoolSizes[i].descriptorCount;
    609          } else {
    610             /* Allocate the maximum size possible.
    611              * Since we don't support VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER for
    612              * mutable descriptors, we can set the default size of descriptor types.
    613              */
    614             bo_size += A6XX_TEX_CONST_DWORDS * 4 *
    615                   pCreateInfo->pPoolSizes[i].descriptorCount;
    616          }
    617          continue;
    618       default:
    619          break;
    620       }
    621 
    622       bo_size += descriptor_size(pCreateInfo->pPoolSizes[i].type) *
    623                            pCreateInfo->pPoolSizes[i].descriptorCount;
    624    }
    625 
    626    if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
    627       uint64_t host_size = pCreateInfo->maxSets * sizeof(struct tu_descriptor_set);
    628       host_size += sizeof(struct tu_bo*) * bo_count;
    629       host_size += A6XX_TEX_CONST_DWORDS * 4 * dynamic_count;
    630       size += host_size;
    631    } else {
    632       size += sizeof(struct tu_descriptor_pool_entry) * pCreateInfo->maxSets;
    633    }
    634 
    635    pool = vk_object_zalloc(&device->vk, pAllocator, size,
    636                           VK_OBJECT_TYPE_DESCRIPTOR_POOL);
    637    if (!pool)
    638       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
    639 
    640    if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
    641       pool->host_memory_base = (uint8_t*)pool + sizeof(struct tu_descriptor_pool);
    642       pool->host_memory_ptr = pool->host_memory_base;
    643       pool->host_memory_end = (uint8_t*)pool + size;
    644    }
    645 
    646    if (bo_size) {
    647       if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_HOST_ONLY_BIT_VALVE)) {
    648          ret = tu_bo_init_new(device, &pool->bo, bo_size, TU_BO_ALLOC_ALLOW_DUMP);
    649          if (ret)
    650             goto fail_alloc;
    651 
    652          ret = tu_bo_map(device, &pool->bo);
    653          if (ret)
    654             goto fail_map;
    655       } else {
    656          pool->host_bo = vk_alloc2(&device->vk.alloc, pAllocator, bo_size, 8,
    657                                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    658          if (!pool->host_bo) {
    659             ret = VK_ERROR_OUT_OF_HOST_MEMORY;
    660             goto fail_alloc;
    661          }
    662       }
    663    }
    664    pool->size = bo_size;
    665    pool->max_entry_count = pCreateInfo->maxSets;
    666 
    667    *pDescriptorPool = tu_descriptor_pool_to_handle(pool);
    668    return VK_SUCCESS;
    669 
    670 fail_map:
    671    tu_bo_finish(device, &pool->bo);
    672 fail_alloc:
    673    vk_object_free(&device->vk, pAllocator, pool);
    674    return ret;
    675 }
    676 
    677 VKAPI_ATTR void VKAPI_CALL
    678 tu_DestroyDescriptorPool(VkDevice _device,
    679                          VkDescriptorPool _pool,
    680                          const VkAllocationCallbacks *pAllocator)
    681 {
    682    TU_FROM_HANDLE(tu_device, device, _device);
    683    TU_FROM_HANDLE(tu_descriptor_pool, pool, _pool);
    684 
    685    if (!pool)
    686       return;
    687 
    688    if (!pool->host_memory_base) {
    689       for(int i = 0; i < pool->entry_count; ++i) {
    690          tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
    691       }
    692    }
    693 
    694    if (pool->size) {
    695       if (pool->host_bo)
    696          vk_free2(&device->vk.alloc, pAllocator, pool->host_bo);
    697       else
    698          tu_bo_finish(device, &pool->bo);
    699    }
    700 
    701    vk_object_free(&device->vk, pAllocator, pool);
    702 }
    703 
    704 VKAPI_ATTR VkResult VKAPI_CALL
    705 tu_ResetDescriptorPool(VkDevice _device,
    706                        VkDescriptorPool descriptorPool,
    707                        VkDescriptorPoolResetFlags flags)
    708 {
    709    TU_FROM_HANDLE(tu_device, device, _device);
    710    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
    711 
    712    if (!pool->host_memory_base) {
    713       for(int i = 0; i < pool->entry_count; ++i) {
    714          tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
    715       }
    716       pool->entry_count = 0;
    717    }
    718 
    719    pool->current_offset = 0;
    720    pool->host_memory_ptr = pool->host_memory_base;
    721 
    722    return VK_SUCCESS;
    723 }
    724 
    725 VKAPI_ATTR VkResult VKAPI_CALL
    726 tu_AllocateDescriptorSets(VkDevice _device,
    727                           const VkDescriptorSetAllocateInfo *pAllocateInfo,
    728                           VkDescriptorSet *pDescriptorSets)
    729 {
    730    TU_FROM_HANDLE(tu_device, device, _device);
    731    TU_FROM_HANDLE(tu_descriptor_pool, pool, pAllocateInfo->descriptorPool);
    732 
    733    VkResult result = VK_SUCCESS;
    734    uint32_t i;
    735    struct tu_descriptor_set *set = NULL;
    736 
    737    const VkDescriptorSetVariableDescriptorCountAllocateInfoEXT *variable_counts =
    738       vk_find_struct_const(pAllocateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_ALLOCATE_INFO_EXT);
    739    const uint32_t zero = 0;
    740 
    741    /* allocate a set of buffers for each shader to contain descriptors */
    742    for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
    743       TU_FROM_HANDLE(tu_descriptor_set_layout, layout,
    744              pAllocateInfo->pSetLayouts[i]);
    745 
    746       const uint32_t *variable_count = NULL;
    747       if (variable_counts) {
    748          if (i < variable_counts->descriptorSetCount)
    749             variable_count = variable_counts->pDescriptorCounts + i;
    750          else
    751             variable_count = &zero;
    752       }
    753 
    754       assert(!(layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
    755 
    756       result = tu_descriptor_set_create(device, pool, layout, variable_count, &set);
    757       if (result != VK_SUCCESS)
    758          break;
    759 
    760       pDescriptorSets[i] = tu_descriptor_set_to_handle(set);
    761    }
    762 
    763    if (result != VK_SUCCESS) {
    764       tu_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool,
    765                i, pDescriptorSets);
    766       for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
    767          pDescriptorSets[i] = VK_NULL_HANDLE;
    768       }
    769    }
    770    return result;
    771 }
    772 
    773 VKAPI_ATTR VkResult VKAPI_CALL
    774 tu_FreeDescriptorSets(VkDevice _device,
    775                       VkDescriptorPool descriptorPool,
    776                       uint32_t count,
    777                       const VkDescriptorSet *pDescriptorSets)
    778 {
    779    TU_FROM_HANDLE(tu_device, device, _device);
    780    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
    781 
    782    for (uint32_t i = 0; i < count; i++) {
    783       TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
    784 
    785       if (set && !pool->host_memory_base)
    786          tu_descriptor_set_destroy(device, pool, set, true);
    787    }
    788    return VK_SUCCESS;
    789 }
    790 
    791 static void
    792 write_texel_buffer_descriptor(uint32_t *dst, const VkBufferView buffer_view)
    793 {
    794    if (buffer_view == VK_NULL_HANDLE) {
    795       memset(dst, 0, A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
    796    } else {
    797       TU_FROM_HANDLE(tu_buffer_view, view, buffer_view);
    798 
    799       memcpy(dst, view->descriptor, sizeof(view->descriptor));
    800    }
    801 }
    802 
    803 static uint32_t get_range(struct tu_buffer *buf, VkDeviceSize offset,
    804                           VkDeviceSize range)
    805 {
    806    if (range == VK_WHOLE_SIZE) {
    807       return buf->size - offset;
    808    } else {
    809       return range;
    810    }
    811 }
    812 
    813 static void
    814 write_buffer_descriptor(const struct tu_device *device,
    815                         uint32_t *dst,
    816                         const VkDescriptorBufferInfo *buffer_info)
    817 {
    818    if (buffer_info->buffer == VK_NULL_HANDLE) {
    819       memset(dst, 0, A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
    820       return;
    821    }
    822 
    823    TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
    824 
    825    assert((buffer_info->offset & 63) == 0); /* minStorageBufferOffsetAlignment */
    826    uint64_t va = tu_buffer_iova(buffer) + buffer_info->offset;
    827    uint32_t range = get_range(buffer, buffer_info->offset, buffer_info->range);
    828 
    829    /* newer a6xx allows using 16-bit descriptor for both 16-bit and 32-bit access */
    830    if (device->physical_device->info->a6xx.storage_16bit) {
    831       dst[0] = A6XX_IBO_0_TILE_MODE(TILE6_LINEAR) | A6XX_IBO_0_FMT(FMT6_16_UINT);
    832       dst[1] = DIV_ROUND_UP(range, 2);
    833    } else {
    834       dst[0] = A6XX_IBO_0_TILE_MODE(TILE6_LINEAR) | A6XX_IBO_0_FMT(FMT6_32_UINT);
    835       dst[1] = DIV_ROUND_UP(range, 4);
    836    }
    837    dst[2] =
    838       A6XX_IBO_2_UNK4 | A6XX_IBO_2_TYPE(A6XX_TEX_1D) | A6XX_IBO_2_UNK31;
    839    dst[3] = 0;
    840    dst[4] = A6XX_IBO_4_BASE_LO(va);
    841    dst[5] = A6XX_IBO_5_BASE_HI(va >> 32);
    842    for (int i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
    843       dst[i] = 0;
    844 }
    845 
    846 static void
    847 write_ubo_descriptor(uint32_t *dst, const VkDescriptorBufferInfo *buffer_info)
    848 {
    849    if (buffer_info->buffer == VK_NULL_HANDLE) {
    850       dst[0] = dst[1] = 0;
    851       return;
    852    }
    853 
    854    TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
    855 
    856    uint32_t range = get_range(buffer, buffer_info->offset, buffer_info->range);
    857    /* The HW range is in vec4 units */
    858    range = ALIGN_POT(range, 16) / 16;
    859    uint64_t va = tu_buffer_iova(buffer) + buffer_info->offset;
    860 
    861    dst[0] = A6XX_UBO_0_BASE_LO(va);
    862    dst[1] = A6XX_UBO_1_BASE_HI(va >> 32) | A6XX_UBO_1_SIZE(range);
    863 }
    864 
    865 static void
    866 write_image_descriptor(uint32_t *dst,
    867                        VkDescriptorType descriptor_type,
    868                        const VkDescriptorImageInfo *image_info)
    869 {
    870    if (image_info->imageView == VK_NULL_HANDLE) {
    871       memset(dst, 0, A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
    872       return;
    873    }
    874 
    875    TU_FROM_HANDLE(tu_image_view, iview, image_info->imageView);
    876 
    877    if (descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) {
    878       memcpy(dst, iview->storage_descriptor, sizeof(iview->storage_descriptor));
    879    } else {
    880       memcpy(dst, iview->descriptor, sizeof(iview->descriptor));
    881    }
    882 }
    883 
    884 static void
    885 write_combined_image_sampler_descriptor(uint32_t *dst,
    886                                         VkDescriptorType descriptor_type,
    887                                         const VkDescriptorImageInfo *image_info,
    888                                         bool has_sampler)
    889 {
    890    write_image_descriptor(dst, descriptor_type, image_info);
    891    /* copy over sampler state */
    892    if (has_sampler) {
    893       TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
    894       memcpy(dst + A6XX_TEX_CONST_DWORDS, sampler->descriptor, sizeof(sampler->descriptor));
    895    }
    896 }
    897 
    898 static void
    899 write_sampler_descriptor(uint32_t *dst, const VkDescriptorImageInfo *image_info)
    900 {
    901    TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
    902 
    903    memcpy(dst, sampler->descriptor, sizeof(sampler->descriptor));
    904 }
    905 
    906 /* note: this is used with immutable samplers in push descriptors */
    907 static void
    908 write_sampler_push(uint32_t *dst, const struct tu_sampler *sampler)
    909 {
    910    memcpy(dst, sampler->descriptor, sizeof(sampler->descriptor));
    911 }
    912 
    913 void
    914 tu_update_descriptor_sets(const struct tu_device *device,
    915                           VkDescriptorSet dstSetOverride,
    916                           uint32_t descriptorWriteCount,
    917                           const VkWriteDescriptorSet *pDescriptorWrites,
    918                           uint32_t descriptorCopyCount,
    919                           const VkCopyDescriptorSet *pDescriptorCopies)
    920 {
    921    uint32_t i, j;
    922    for (i = 0; i < descriptorWriteCount; i++) {
    923       const VkWriteDescriptorSet *writeset = &pDescriptorWrites[i];
    924       TU_FROM_HANDLE(tu_descriptor_set, set, dstSetOverride ?: writeset->dstSet);
    925       const struct tu_descriptor_set_binding_layout *binding_layout =
    926          set->layout->binding + writeset->dstBinding;
    927       uint32_t *ptr = set->mapped_ptr;
    928       /* for immutable samplers with push descriptors: */
    929       const bool copy_immutable_samplers =
    930          dstSetOverride && binding_layout->immutable_samplers_offset;
    931       const struct tu_sampler *samplers =
    932          tu_immutable_samplers(set->layout, binding_layout);
    933 
    934       ptr += binding_layout->offset / 4;
    935 
    936       ptr += (binding_layout->size / 4) * writeset->dstArrayElement;
    937       for (j = 0; j < writeset->descriptorCount; ++j) {
    938          switch(writeset->descriptorType) {
    939          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: {
    940             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
    941             unsigned idx = writeset->dstArrayElement + j;
    942             idx += binding_layout->dynamic_offset_offset;
    943             write_ubo_descriptor(set->dynamic_descriptors + A6XX_TEX_CONST_DWORDS * idx,
    944                                  writeset->pBufferInfo + j);
    945             break;
    946          }
    947          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
    948             write_ubo_descriptor(ptr, writeset->pBufferInfo + j);
    949             break;
    950          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
    951             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
    952             unsigned idx = writeset->dstArrayElement + j;
    953             idx += binding_layout->dynamic_offset_offset;
    954             write_buffer_descriptor(device, set->dynamic_descriptors + A6XX_TEX_CONST_DWORDS * idx,
    955                                     writeset->pBufferInfo + j);
    956             break;
    957          }
    958          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
    959             write_buffer_descriptor(device, ptr, writeset->pBufferInfo + j);
    960             break;
    961          case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
    962          case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
    963             write_texel_buffer_descriptor(ptr, writeset->pTexelBufferView[j]);
    964             break;
    965          case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
    966          case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
    967             write_image_descriptor(ptr, writeset->descriptorType, writeset->pImageInfo + j);
    968             break;
    969          case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
    970             write_combined_image_sampler_descriptor(ptr,
    971                                                     writeset->descriptorType,
    972                                                     writeset->pImageInfo + j,
    973                                                     !binding_layout->immutable_samplers_offset);
    974 
    975             if (copy_immutable_samplers)
    976                write_sampler_push(ptr + A6XX_TEX_CONST_DWORDS, &samplers[writeset->dstArrayElement + j]);
    977             break;
    978          case VK_DESCRIPTOR_TYPE_SAMPLER:
    979             if (!binding_layout->immutable_samplers_offset)
    980                write_sampler_descriptor(ptr, writeset->pImageInfo + j);
    981             else if (copy_immutable_samplers)
    982                write_sampler_push(ptr, &samplers[writeset->dstArrayElement + j]);
    983             break;
    984          case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
    985             /* nothing in descriptor set - framebuffer state is used instead */
    986             break;
    987          default:
    988             unreachable("unimplemented descriptor type");
    989             break;
    990          }
    991          ptr += binding_layout->size / 4;
    992       }
    993    }
    994 
    995    for (i = 0; i < descriptorCopyCount; i++) {
    996       const VkCopyDescriptorSet *copyset = &pDescriptorCopies[i];
    997       TU_FROM_HANDLE(tu_descriptor_set, src_set,
    998                        copyset->srcSet);
    999       TU_FROM_HANDLE(tu_descriptor_set, dst_set,
   1000                        copyset->dstSet);
   1001       const struct tu_descriptor_set_binding_layout *src_binding_layout =
   1002          src_set->layout->binding + copyset->srcBinding;
   1003       const struct tu_descriptor_set_binding_layout *dst_binding_layout =
   1004          dst_set->layout->binding + copyset->dstBinding;
   1005       uint32_t *src_ptr = src_set->mapped_ptr;
   1006       uint32_t *dst_ptr = dst_set->mapped_ptr;
   1007 
   1008       src_ptr += src_binding_layout->offset / 4;
   1009       dst_ptr += dst_binding_layout->offset / 4;
   1010 
   1011       src_ptr += src_binding_layout->size * copyset->srcArrayElement / 4;
   1012       dst_ptr += dst_binding_layout->size * copyset->dstArrayElement / 4;
   1013 
   1014       /* In case of copies between mutable descriptor types
   1015        * and non-mutable descriptor types.
   1016        */
   1017       uint32_t copy_size = MIN2(src_binding_layout->size, dst_binding_layout->size);
   1018 
   1019       for (j = 0; j < copyset->descriptorCount; ++j) {
   1020          switch (src_binding_layout->type) {
   1021          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
   1022          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
   1023             unsigned src_idx = copyset->srcArrayElement + j;
   1024             unsigned dst_idx = copyset->dstArrayElement + j;
   1025             src_idx += src_binding_layout->dynamic_offset_offset;
   1026             dst_idx += dst_binding_layout->dynamic_offset_offset;
   1027 
   1028             uint32_t *src_dynamic, *dst_dynamic;
   1029             src_dynamic = src_set->dynamic_descriptors + src_idx * A6XX_TEX_CONST_DWORDS;
   1030             dst_dynamic = dst_set->dynamic_descriptors + dst_idx * A6XX_TEX_CONST_DWORDS;
   1031             memcpy(dst_dynamic, src_dynamic, A6XX_TEX_CONST_DWORDS * 4);
   1032             break;
   1033          }
   1034          default:
   1035             memcpy(dst_ptr, src_ptr, copy_size);
   1036          }
   1037 
   1038          src_ptr += src_binding_layout->size / 4;
   1039          dst_ptr += dst_binding_layout->size / 4;
   1040       }
   1041    }
   1042 }
   1043 
   1044 VKAPI_ATTR void VKAPI_CALL
   1045 tu_UpdateDescriptorSets(VkDevice _device,
   1046                         uint32_t descriptorWriteCount,
   1047                         const VkWriteDescriptorSet *pDescriptorWrites,
   1048                         uint32_t descriptorCopyCount,
   1049                         const VkCopyDescriptorSet *pDescriptorCopies)
   1050 {
   1051    TU_FROM_HANDLE(tu_device, device, _device);
   1052    tu_update_descriptor_sets(device, VK_NULL_HANDLE,
   1053                              descriptorWriteCount, pDescriptorWrites,
   1054                              descriptorCopyCount, pDescriptorCopies);
   1055 }
   1056 
   1057 VKAPI_ATTR VkResult VKAPI_CALL
   1058 tu_CreateDescriptorUpdateTemplate(
   1059    VkDevice _device,
   1060    const VkDescriptorUpdateTemplateCreateInfo *pCreateInfo,
   1061    const VkAllocationCallbacks *pAllocator,
   1062    VkDescriptorUpdateTemplate *pDescriptorUpdateTemplate)
   1063 {
   1064    TU_FROM_HANDLE(tu_device, device, _device);
   1065    TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
   1066                   pCreateInfo->descriptorSetLayout);
   1067    const uint32_t entry_count = pCreateInfo->descriptorUpdateEntryCount;
   1068    const size_t size =
   1069       sizeof(struct tu_descriptor_update_template) +
   1070       sizeof(struct tu_descriptor_update_template_entry) * entry_count;
   1071    struct tu_descriptor_update_template *templ;
   1072 
   1073    templ = vk_object_alloc(&device->vk, pAllocator, size,
   1074                            VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE);
   1075    if (!templ)
   1076       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
   1077 
   1078    templ->entry_count = entry_count;
   1079 
   1080    if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) {
   1081       TU_FROM_HANDLE(tu_pipeline_layout, pipeline_layout, pCreateInfo->pipelineLayout);
   1082 
   1083       /* descriptorSetLayout should be ignored for push descriptors
   1084        * and instead it refers to pipelineLayout and set.
   1085        */
   1086       assert(pCreateInfo->set < MAX_SETS);
   1087       set_layout = pipeline_layout->set[pCreateInfo->set].layout;
   1088 
   1089       templ->bind_point = pCreateInfo->pipelineBindPoint;
   1090    }
   1091 
   1092    for (uint32_t i = 0; i < entry_count; i++) {
   1093       const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i];
   1094 
   1095       const struct tu_descriptor_set_binding_layout *binding_layout =
   1096          set_layout->binding + entry->dstBinding;
   1097       uint32_t dst_offset, dst_stride;
   1098       const struct tu_sampler *immutable_samplers = NULL;
   1099 
   1100       /* dst_offset is an offset into dynamic_descriptors when the descriptor
   1101        * is dynamic, and an offset into mapped_ptr otherwise.
   1102        */
   1103       switch (entry->descriptorType) {
   1104       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
   1105       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
   1106          dst_offset = (binding_layout->dynamic_offset_offset +
   1107             entry->dstArrayElement) * A6XX_TEX_CONST_DWORDS;
   1108          dst_stride = A6XX_TEX_CONST_DWORDS;
   1109          break;
   1110       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
   1111       case VK_DESCRIPTOR_TYPE_SAMPLER:
   1112          if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR &&
   1113              binding_layout->immutable_samplers_offset) {
   1114             immutable_samplers =
   1115                tu_immutable_samplers(set_layout, binding_layout) + entry->dstArrayElement;
   1116          }
   1117          FALLTHROUGH;
   1118       default:
   1119          dst_offset = binding_layout->offset / 4;
   1120          dst_offset += (binding_layout->size * entry->dstArrayElement) / 4;
   1121          dst_stride = binding_layout->size / 4;
   1122       }
   1123 
   1124       templ->entry[i] = (struct tu_descriptor_update_template_entry) {
   1125          .descriptor_type = entry->descriptorType,
   1126          .descriptor_count = entry->descriptorCount,
   1127          .src_offset = entry->offset,
   1128          .src_stride = entry->stride,
   1129          .dst_offset = dst_offset,
   1130          .dst_stride = dst_stride,
   1131          .has_sampler = !binding_layout->immutable_samplers_offset,
   1132          .immutable_samplers = immutable_samplers,
   1133       };
   1134    }
   1135 
   1136    *pDescriptorUpdateTemplate =
   1137       tu_descriptor_update_template_to_handle(templ);
   1138 
   1139    return VK_SUCCESS;
   1140 }
   1141 
   1142 VKAPI_ATTR void VKAPI_CALL
   1143 tu_DestroyDescriptorUpdateTemplate(
   1144    VkDevice _device,
   1145    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
   1146    const VkAllocationCallbacks *pAllocator)
   1147 {
   1148    TU_FROM_HANDLE(tu_device, device, _device);
   1149    TU_FROM_HANDLE(tu_descriptor_update_template, templ,
   1150                   descriptorUpdateTemplate);
   1151 
   1152    if (!templ)
   1153       return;
   1154 
   1155    vk_object_free(&device->vk, pAllocator, templ);
   1156 }
   1157 
   1158 void
   1159 tu_update_descriptor_set_with_template(
   1160    const struct tu_device *device,
   1161    struct tu_descriptor_set *set,
   1162    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
   1163    const void *pData)
   1164 {
   1165    TU_FROM_HANDLE(tu_descriptor_update_template, templ,
   1166                   descriptorUpdateTemplate);
   1167 
   1168    for (uint32_t i = 0; i < templ->entry_count; i++) {
   1169       uint32_t *ptr = set->mapped_ptr;
   1170       const void *src = ((const char *) pData) + templ->entry[i].src_offset;
   1171       const struct tu_sampler *samplers = templ->entry[i].immutable_samplers;
   1172 
   1173       ptr += templ->entry[i].dst_offset;
   1174       unsigned dst_offset = templ->entry[i].dst_offset;
   1175       for (unsigned j = 0; j < templ->entry[i].descriptor_count; ++j) {
   1176          switch(templ->entry[i].descriptor_type) {
   1177          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: {
   1178             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
   1179             write_ubo_descriptor(set->dynamic_descriptors + dst_offset, src);
   1180             break;
   1181          }
   1182          case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
   1183             write_ubo_descriptor(ptr, src);
   1184             break;
   1185          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
   1186             assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
   1187             write_buffer_descriptor(device, set->dynamic_descriptors + dst_offset, src);
   1188             break;
   1189          }
   1190          case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
   1191             write_buffer_descriptor(device, ptr, src);
   1192             break;
   1193          case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
   1194          case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
   1195             write_texel_buffer_descriptor(ptr, *(VkBufferView *) src);
   1196             break;
   1197          case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
   1198          case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
   1199             write_image_descriptor(ptr, templ->entry[i].descriptor_type,  src);
   1200             break;
   1201          }
   1202          case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
   1203             write_combined_image_sampler_descriptor(ptr,
   1204                                                     templ->entry[i].descriptor_type,
   1205                                                     src,
   1206                                                     templ->entry[i].has_sampler);
   1207             if (samplers)
   1208                write_sampler_push(ptr + A6XX_TEX_CONST_DWORDS, &samplers[j]);
   1209             break;
   1210          case VK_DESCRIPTOR_TYPE_SAMPLER:
   1211             if (templ->entry[i].has_sampler)
   1212                write_sampler_descriptor(ptr, src);
   1213             else if (samplers)
   1214                write_sampler_push(ptr, &samplers[j]);
   1215             break;
   1216          case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
   1217             /* nothing in descriptor set - framebuffer state is used instead */
   1218             break;
   1219          default:
   1220             unreachable("unimplemented descriptor type");
   1221             break;
   1222          }
   1223          src = (char *) src + templ->entry[i].src_stride;
   1224          ptr += templ->entry[i].dst_stride;
   1225          dst_offset += templ->entry[i].dst_stride;
   1226       }
   1227    }
   1228 }
   1229 
   1230 VKAPI_ATTR void VKAPI_CALL
   1231 tu_UpdateDescriptorSetWithTemplate(
   1232    VkDevice _device,
   1233    VkDescriptorSet descriptorSet,
   1234    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
   1235    const void *pData)
   1236 {
   1237    TU_FROM_HANDLE(tu_device, device, _device);
   1238    TU_FROM_HANDLE(tu_descriptor_set, set, descriptorSet);
   1239 
   1240    tu_update_descriptor_set_with_template(device, set, descriptorUpdateTemplate, pData);
   1241 }
   1242 
   1243 VKAPI_ATTR VkResult VKAPI_CALL
   1244 tu_CreateSamplerYcbcrConversion(
   1245    VkDevice _device,
   1246    const VkSamplerYcbcrConversionCreateInfo *pCreateInfo,
   1247    const VkAllocationCallbacks *pAllocator,
   1248    VkSamplerYcbcrConversion *pYcbcrConversion)
   1249 {
   1250    TU_FROM_HANDLE(tu_device, device, _device);
   1251    struct tu_sampler_ycbcr_conversion *conversion;
   1252 
   1253    conversion = vk_object_alloc(&device->vk, pAllocator, sizeof(*conversion),
   1254                                 VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION);
   1255    if (!conversion)
   1256       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
   1257 
   1258    conversion->format = pCreateInfo->format;
   1259    conversion->ycbcr_model = pCreateInfo->ycbcrModel;
   1260    conversion->ycbcr_range = pCreateInfo->ycbcrRange;
   1261    conversion->components = pCreateInfo->components;
   1262    conversion->chroma_offsets[0] = pCreateInfo->xChromaOffset;
   1263    conversion->chroma_offsets[1] = pCreateInfo->yChromaOffset;
   1264    conversion->chroma_filter = pCreateInfo->chromaFilter;
   1265 
   1266    *pYcbcrConversion = tu_sampler_ycbcr_conversion_to_handle(conversion);
   1267    return VK_SUCCESS;
   1268 }
   1269 
   1270 VKAPI_ATTR void VKAPI_CALL
   1271 tu_DestroySamplerYcbcrConversion(VkDevice _device,
   1272                                  VkSamplerYcbcrConversion ycbcrConversion,
   1273                                  const VkAllocationCallbacks *pAllocator)
   1274 {
   1275    TU_FROM_HANDLE(tu_device, device, _device);
   1276    TU_FROM_HANDLE(tu_sampler_ycbcr_conversion, ycbcr_conversion, ycbcrConversion);
   1277 
   1278    if (!ycbcr_conversion)
   1279       return;
   1280 
   1281    vk_object_free(&device->vk, pAllocator, ycbcr_conversion);
   1282 }
   1283