X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Famd%2Fvulkan%2Fradv_descriptor_set.c;h=8d1343d2d8bb6ec81638d0913f312d451cb9186a;hb=178adfa6a8e5694a077c3e11fdcc5bc0802839ab;hp=b608063d0706afe533cacb493738dc5322a220fe;hpb=74d92e547c0c5118e15df71b81a46a588e6a0a01;p=mesa.git diff --git a/src/amd/vulkan/radv_descriptor_set.c b/src/amd/vulkan/radv_descriptor_set.c index b608063d070..8d1343d2d8b 100644 --- a/src/amd/vulkan/radv_descriptor_set.c +++ b/src/amd/vulkan/radv_descriptor_set.c @@ -30,6 +30,43 @@ #include "util/mesa-sha1.h" #include "radv_private.h" #include "sid.h" +#include "vk_format.h" +#include "vk_util.h" + + +static bool has_equal_immutable_samplers(const VkSampler *samplers, uint32_t count) +{ + if (!samplers) + return false; + for(uint32_t i = 1; i < count; ++i) { + if (memcmp(radv_sampler_from_handle(samplers[0])->state, + radv_sampler_from_handle(samplers[i])->state, 16)) { + return false; + } + } + return true; +} + +static int binding_compare(const void* av, const void *bv) +{ + const VkDescriptorSetLayoutBinding *a = (const VkDescriptorSetLayoutBinding*)av; + const VkDescriptorSetLayoutBinding *b = (const VkDescriptorSetLayoutBinding*)bv; + + return (a->binding < b->binding) ? -1 : (a->binding > b->binding) ? 1 : 0; +} + +static VkDescriptorSetLayoutBinding * +create_sorted_bindings(const VkDescriptorSetLayoutBinding *bindings, unsigned count) { + VkDescriptorSetLayoutBinding *sorted_bindings = malloc(count * sizeof(VkDescriptorSetLayoutBinding)); + if (!sorted_bindings) + return NULL; + + memcpy(sorted_bindings, bindings, count * sizeof(VkDescriptorSetLayoutBinding)); + + qsort(sorted_bindings, count, sizeof(VkDescriptorSetLayoutBinding), binding_compare); + + return sorted_bindings; +} VkResult radv_CreateDescriptorSetLayout( VkDevice _device, @@ -41,31 +78,72 @@ VkResult radv_CreateDescriptorSetLayout( struct radv_descriptor_set_layout *set_layout; assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO); + const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags = + vk_find_struct_const(pCreateInfo->pNext, DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO); uint32_t max_binding = 0; uint32_t immutable_sampler_count = 0; + uint32_t ycbcr_sampler_count = 0; for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) { max_binding = MAX2(max_binding, pCreateInfo->pBindings[j].binding); - if (pCreateInfo->pBindings[j].pImmutableSamplers) + if ((pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER || + pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) && + pCreateInfo->pBindings[j].pImmutableSamplers) { immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount; + + bool has_ycbcr_sampler = false; + for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) { + if (radv_sampler_from_handle(pCreateInfo->pBindings[j].pImmutableSamplers[i])->ycbcr_sampler) + has_ycbcr_sampler = true; + } + + if (has_ycbcr_sampler) + ycbcr_sampler_count += pCreateInfo->pBindings[j].descriptorCount; + } } uint32_t samplers_offset = sizeof(struct radv_descriptor_set_layout) + (max_binding + 1) * sizeof(set_layout->binding[0]); size_t size = samplers_offset + immutable_sampler_count * 4 * sizeof(uint32_t); + if (ycbcr_sampler_count > 0) { + size += ycbcr_sampler_count * sizeof(struct radv_sampler_ycbcr_conversion) + (max_binding + 1) * sizeof(uint32_t); + } - set_layout = vk_alloc2(&device->alloc, pAllocator, size, 8, - VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + set_layout = vk_zalloc2(&device->vk.alloc, pAllocator, size, 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!set_layout) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); + + vk_object_base_init(&device->vk, &set_layout->base, + VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT); set_layout->flags = pCreateInfo->flags; + set_layout->layout_size = size; /* We just allocate all the samplers at the end of the struct */ uint32_t *samplers = (uint32_t*)&set_layout->binding[max_binding + 1]; + struct radv_sampler_ycbcr_conversion *ycbcr_samplers = NULL; + uint32_t *ycbcr_sampler_offsets = NULL; + + if (ycbcr_sampler_count > 0) { + ycbcr_sampler_offsets = samplers + 4 * immutable_sampler_count; + set_layout->ycbcr_sampler_offsets_offset = (char*)ycbcr_sampler_offsets - (char*)set_layout; + ycbcr_samplers = (struct radv_sampler_ycbcr_conversion *)(ycbcr_sampler_offsets + max_binding + 1); + } else + set_layout->ycbcr_sampler_offsets_offset = 0; + + VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(pCreateInfo->pBindings, + pCreateInfo->bindingCount); + if (!bindings) { + vk_object_base_finish(&set_layout->base); + vk_free2(&device->vk.alloc, pAllocator, set_layout); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); + } set_layout->binding_count = max_binding + 1; set_layout->shader_stages = 0; + set_layout->dynamic_shader_stages = 0; + set_layout->has_immutable_samplers = false; set_layout->size = 0; memset(set_layout->binding, 0, size - sizeof(struct radv_descriptor_set_layout)); @@ -74,10 +152,29 @@ VkResult radv_CreateDescriptorSetLayout( uint32_t dynamic_offset_count = 0; for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) { - const VkDescriptorSetLayoutBinding *binding = &pCreateInfo->pBindings[j]; + const VkDescriptorSetLayoutBinding *binding = bindings + j; uint32_t b = binding->binding; - uint32_t alignment; + uint32_t alignment = 0; unsigned binding_buffer_count = 0; + uint32_t descriptor_count = binding->descriptorCount; + bool has_ycbcr_sampler = false; + + /* main image + fmask */ + uint32_t max_sampled_image_descriptors = 2; + + if (binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER && + binding->pImmutableSamplers) { + for (unsigned i = 0; i < binding->descriptorCount; ++i) { + struct radv_sampler_ycbcr_conversion *conversion = + radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler; + + if (conversion) { + has_ycbcr_sampler = true; + max_sampled_image_descriptors = MAX2(max_sampled_image_descriptors, + vk_format_get_plane_count(conversion->format)); + } + } + } switch (binding->descriptorType) { case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: @@ -115,48 +212,74 @@ VkResult radv_CreateDescriptorSetLayout( set_layout->binding[b].size = 16; alignment = 16; break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + alignment = 16; + set_layout->binding[b].size = descriptor_count; + descriptor_count = 1; + break; default: - unreachable("unknown descriptor type\n"); break; } set_layout->size = align(set_layout->size, alignment); - assert(binding->descriptorCount > 0); set_layout->binding[b].type = binding->descriptorType; - set_layout->binding[b].array_size = binding->descriptorCount; + set_layout->binding[b].array_size = descriptor_count; set_layout->binding[b].offset = set_layout->size; set_layout->binding[b].buffer_offset = buffer_count; set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count; - if (binding->pImmutableSamplers) { + if (variable_flags && binding->binding < variable_flags->bindingCount && + (variable_flags->pBindingFlags[binding->binding] & VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) { + assert(!binding->pImmutableSamplers); /* Terribly ill defined how many samplers are valid */ + assert(binding->binding == max_binding); + + set_layout->has_variable_descriptors = true; + } + + if ((binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER || + binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) && + binding->pImmutableSamplers) { set_layout->binding[b].immutable_samplers_offset = samplers_offset; - set_layout->binding[b].immutable_samplers_equal = true; + set_layout->binding[b].immutable_samplers_equal = + has_equal_immutable_samplers(binding->pImmutableSamplers, binding->descriptorCount); + set_layout->has_immutable_samplers = true; for (uint32_t i = 0; i < binding->descriptorCount; i++) memcpy(samplers + 4 * i, &radv_sampler_from_handle(binding->pImmutableSamplers[i])->state, 16); - for (uint32_t i = 1; i < binding->descriptorCount; i++) - if (memcmp(samplers + 4 * i, samplers, 16) != 0) - set_layout->binding[b].immutable_samplers_equal = false; /* Don't reserve space for the samplers if they're not accessed. */ if (set_layout->binding[b].immutable_samplers_equal) { - if (binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) + if (binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER && + max_sampled_image_descriptors <= 2) set_layout->binding[b].size -= 32; else if (binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) set_layout->binding[b].size -= 16; } samplers += 4 * binding->descriptorCount; samplers_offset += 4 * sizeof(uint32_t) * binding->descriptorCount; + + if (has_ycbcr_sampler) { + ycbcr_sampler_offsets[b] = (const char*)ycbcr_samplers - (const char*)set_layout; + for (uint32_t i = 0; i < binding->descriptorCount; i++) { + if (radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler) + ycbcr_samplers[i] = *radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler; + else + ycbcr_samplers[i].format = VK_FORMAT_UNDEFINED; + } + ycbcr_samplers += binding->descriptorCount; + } } - set_layout->size += binding->descriptorCount * set_layout->binding[b].size; - buffer_count += binding->descriptorCount * binding_buffer_count; - dynamic_offset_count += binding->descriptorCount * + set_layout->size += descriptor_count * set_layout->binding[b].size; + buffer_count += descriptor_count * binding_buffer_count; + dynamic_offset_count += descriptor_count * set_layout->binding[b].dynamic_offset_count; set_layout->shader_stages |= binding->stageFlags; } + free(bindings); + set_layout->buffer_count = buffer_count; set_layout->dynamic_offset_count = dynamic_offset_count; @@ -176,12 +299,106 @@ void radv_DestroyDescriptorSetLayout( if (!set_layout) return; - vk_free2(&device->alloc, pAllocator, set_layout); + vk_object_base_finish(&set_layout->base); + vk_free2(&device->vk.alloc, pAllocator, set_layout); +} + +void radv_GetDescriptorSetLayoutSupport(VkDevice device, + const VkDescriptorSetLayoutCreateInfo* pCreateInfo, + VkDescriptorSetLayoutSupport* pSupport) +{ + VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(pCreateInfo->pBindings, + pCreateInfo->bindingCount); + if (!bindings) { + pSupport->supported = false; + return; + } + + const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags = + vk_find_struct_const(pCreateInfo->pNext, DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO); + VkDescriptorSetVariableDescriptorCountLayoutSupport *variable_count = + vk_find_struct((void*)pCreateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT); + if (variable_count) { + variable_count->maxVariableDescriptorCount = 0; + } + + bool supported = true; + uint64_t size = 0; + for (uint32_t i = 0; i < pCreateInfo->bindingCount; i++) { + const VkDescriptorSetLayoutBinding *binding = bindings + i; + + uint64_t descriptor_size = 0; + uint64_t descriptor_alignment = 1; + uint32_t descriptor_count = binding->descriptorCount; + switch (binding->descriptorType) { + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: + break; + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER: + case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER: + case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER: + descriptor_size = 16; + descriptor_alignment = 16; + break; + case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: + case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: + case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: + descriptor_size = 64; + descriptor_alignment = 32; + break; + case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: + if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) { + descriptor_size = 64; + } else { + descriptor_size = 96; + } + descriptor_alignment = 32; + break; + case VK_DESCRIPTOR_TYPE_SAMPLER: + if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) { + descriptor_size = 16; + descriptor_alignment = 16; + } + break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + descriptor_alignment = 16; + descriptor_size = descriptor_count; + descriptor_count = 1; + break; + default: + break; + } + + if (size && !align_u64(size, descriptor_alignment)) { + supported = false; + } + size = align_u64(size, descriptor_alignment); + + uint64_t max_count = INT32_MAX; + if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) + max_count = INT32_MAX - size; + else if (descriptor_size) + max_count = (INT32_MAX - size) / descriptor_size; + + if (max_count < descriptor_count) { + supported = false; + } + if (variable_flags && binding->binding bindingCount && variable_count && + (variable_flags->pBindingFlags[binding->binding] & VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) { + variable_count->maxVariableDescriptorCount = MIN2(UINT32_MAX, max_count); + } + size += descriptor_count * descriptor_size; + } + + free(bindings); + + pSupport->supported = supported; } /* * Pipeline layouts. These have nothing to do with the pipeline. They are - * just muttiple descriptor set layouts pasted together + * just multiple descriptor set layouts pasted together. */ VkResult radv_CreatePipelineLayout( @@ -196,14 +413,18 @@ VkResult radv_CreatePipelineLayout( assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO); - layout = vk_alloc2(&device->alloc, pAllocator, sizeof(*layout), 8, + layout = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*layout), 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (layout == NULL) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); + + vk_object_base_init(&device->vk, &layout->base, + VK_OBJECT_TYPE_PIPELINE_LAYOUT); layout->num_sets = pCreateInfo->setLayoutCount; unsigned dynamic_offset_count = 0; + uint16_t dynamic_shader_stages = 0; _mesa_sha1_init(&ctx); @@ -215,16 +436,15 @@ VkResult radv_CreatePipelineLayout( layout->set[set].dynamic_offset_start = dynamic_offset_count; for (uint32_t b = 0; b < set_layout->binding_count; b++) { dynamic_offset_count += set_layout->binding[b].array_size * set_layout->binding[b].dynamic_offset_count; - if (set_layout->binding[b].immutable_samplers_offset) - _mesa_sha1_update(&ctx, radv_immutable_samplers(set_layout, set_layout->binding + b), - set_layout->binding[b].array_size * 4 * sizeof(uint32_t)); + dynamic_shader_stages |= set_layout->dynamic_shader_stages; } - _mesa_sha1_update(&ctx, set_layout->binding, - sizeof(set_layout->binding[0]) * set_layout->binding_count); + _mesa_sha1_update(&ctx, set_layout, set_layout->layout_size); } layout->dynamic_offset_count = dynamic_offset_count; + layout->dynamic_shader_stages = dynamic_shader_stages; layout->push_constant_size = 0; + for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) { const VkPushConstantRange *range = pCreateInfo->pPushConstantRanges + i; layout->push_constant_size = MAX2(layout->push_constant_size, @@ -250,7 +470,9 @@ void radv_DestroyPipelineLayout( if (!pipeline_layout) return; - vk_free2(&device->alloc, pAllocator, pipeline_layout); + + vk_object_base_finish(&pipeline_layout->base); + vk_free2(&device->vk.alloc, pAllocator, pipeline_layout); } #define EMPTY 1 @@ -259,38 +481,67 @@ static VkResult radv_descriptor_set_create(struct radv_device *device, struct radv_descriptor_pool *pool, const struct radv_descriptor_set_layout *layout, + const uint32_t *variable_count, struct radv_descriptor_set **out_set) { struct radv_descriptor_set *set; + uint32_t buffer_count = layout->buffer_count; + if (variable_count) { + unsigned stride = 1; + if (layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_SAMPLER || + layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) + stride = 0; + buffer_count = layout->binding[layout->binding_count - 1].buffer_offset + + *variable_count * stride; + } unsigned range_offset = sizeof(struct radv_descriptor_set) + - sizeof(struct radeon_winsys_bo *) * layout->buffer_count; + sizeof(struct radeon_winsys_bo *) * buffer_count; unsigned mem_size = range_offset + sizeof(struct radv_descriptor_range) * layout->dynamic_offset_count; if (pool->host_memory_base) { if (pool->host_memory_end - pool->host_memory_ptr < mem_size) - return vk_error(VK_ERROR_OUT_OF_POOL_MEMORY_KHR); + return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY); set = (struct radv_descriptor_set*)pool->host_memory_ptr; pool->host_memory_ptr += mem_size; } else { - set = vk_alloc2(&device->alloc, NULL, mem_size, 8, + set = vk_alloc2(&device->vk.alloc, NULL, mem_size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!set) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); } memset(set, 0, mem_size); + vk_object_base_init(&device->vk, &set->base, + VK_OBJECT_TYPE_DESCRIPTOR_SET); + if (layout->dynamic_offset_count) { set->dynamic_descriptors = (struct radv_descriptor_range*)((uint8_t*)set + range_offset); } set->layout = layout; - if (layout->size) { - uint32_t layout_size = align_u32(layout->size, 32); - set->size = layout->size; + set->buffer_count = buffer_count; + uint32_t layout_size = layout->size; + if (variable_count) { + assert(layout->has_variable_descriptors); + uint32_t stride = layout->binding[layout->binding_count - 1].size; + if (layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) + stride = 1; + + layout_size = layout->binding[layout->binding_count - 1].offset + + *variable_count * stride; + } + layout_size = align_u32(layout_size, 32); + if (layout_size) { + set->size = layout_size; + + if (!pool->host_memory_base && pool->entry_count == pool->max_entry_count) { + vk_free2(&device->vk.alloc, NULL, set); + return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY); + } /* try to allocate linearly first, so that we don't spend * time looking for gaps if the app only allocates & @@ -298,52 +549,58 @@ radv_descriptor_set_create(struct radv_device *device, if (pool->current_offset + layout_size <= pool->size) { set->bo = pool->bo; set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + pool->current_offset); - set->va = device->ws->buffer_get_va(set->bo) + pool->current_offset; + set->va = radv_buffer_get_va(set->bo) + pool->current_offset; + if (!pool->host_memory_base) { + pool->entries[pool->entry_count].offset = pool->current_offset; + pool->entries[pool->entry_count].size = layout_size; + pool->entries[pool->entry_count].set = set; + pool->entry_count++; + } pool->current_offset += layout_size; - list_addtail(&set->vram_list, &pool->vram_list); } else if (!pool->host_memory_base) { uint64_t offset = 0; - struct list_head *prev = &pool->vram_list; - struct radv_descriptor_set *cur; + int index; - assert(!pool->host_memory_base); - LIST_FOR_EACH_ENTRY(cur, &pool->vram_list, vram_list) { - uint64_t start = (uint8_t*)cur->mapped_ptr - pool->mapped_ptr; - if (start - offset >= layout_size) + for (index = 0; index < pool->entry_count; ++index) { + if (pool->entries[index].offset - offset >= layout_size) break; - - offset = start + cur->size; - prev = &cur->vram_list; + offset = pool->entries[index].offset + pool->entries[index].size; } if (pool->size - offset < layout_size) { - vk_free2(&device->alloc, NULL, set->dynamic_descriptors); - vk_free2(&device->alloc, NULL, set); - return vk_error(VK_ERROR_OUT_OF_POOL_MEMORY_KHR); + vk_free2(&device->vk.alloc, NULL, set); + return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY); } set->bo = pool->bo; set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + offset); - set->va = device->ws->buffer_get_va(set->bo) + offset; - list_add(&set->vram_list, prev); + set->va = radv_buffer_get_va(set->bo) + offset; + memmove(&pool->entries[index + 1], &pool->entries[index], + sizeof(pool->entries[0]) * (pool->entry_count - index)); + pool->entries[index].offset = offset; + pool->entries[index].size = layout_size; + pool->entries[index].set = set; + pool->entry_count++; } else - return vk_error(VK_ERROR_OUT_OF_POOL_MEMORY_KHR); + return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY); } - for (unsigned i = 0; i < layout->binding_count; ++i) { - if (!layout->binding[i].immutable_samplers_offset || - layout->binding[i].immutable_samplers_equal) - continue; + if (layout->has_immutable_samplers) { + for (unsigned i = 0; i < layout->binding_count; ++i) { + if (!layout->binding[i].immutable_samplers_offset || + layout->binding[i].immutable_samplers_equal) + continue; - unsigned offset = layout->binding[i].offset / 4; - if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) - offset += 16; + unsigned offset = layout->binding[i].offset / 4; + if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) + offset += radv_combined_image_descriptor_sampler_offset(layout->binding + i) / 4; - const uint32_t *samplers = (const uint32_t*)((const char*)layout + layout->binding[i].immutable_samplers_offset); - for (unsigned j = 0; j < layout->binding[i].array_size; ++j) { - memcpy(set->mapped_ptr + offset, samplers + 4 * j, 16); - offset += layout->binding[i].size / 4; - } + const uint32_t *samplers = (const uint32_t*)((const char*)layout + layout->binding[i].immutable_samplers_offset); + for (unsigned j = 0; j < layout->binding[i].array_size; ++j) { + memcpy(set->mapped_ptr + offset, samplers + 4 * j, 16); + offset += layout->binding[i].size / 4; + } + } } *out_set = set; return VK_SUCCESS; @@ -357,9 +614,19 @@ radv_descriptor_set_destroy(struct radv_device *device, { assert(!pool->host_memory_base); - if (free_bo && set->size) - list_del(&set->vram_list); - vk_free2(&device->alloc, NULL, set); + if (free_bo && set->size && !pool->host_memory_base) { + uint32_t offset = (uint8_t*)set->mapped_ptr - pool->mapped_ptr; + for (int i = 0; i < pool->entry_count; ++i) { + if (pool->entries[i].offset == offset) { + memmove(&pool->entries[i], &pool->entries[i+1], + sizeof(pool->entries[i]) * (pool->entry_count - i - 1)); + --pool->entry_count; + break; + } + } + } + vk_object_base_finish(&set->base); + vk_free2(&device->vk.alloc, NULL, set); } VkResult radv_CreateDescriptorPool( @@ -370,9 +637,24 @@ VkResult radv_CreateDescriptorPool( { RADV_FROM_HANDLE(radv_device, device, _device); struct radv_descriptor_pool *pool; - int size = sizeof(struct radv_descriptor_pool); + uint64_t size = sizeof(struct radv_descriptor_pool); uint64_t bo_size = 0, bo_count = 0, range_count = 0; + vk_foreach_struct(ext, pCreateInfo->pNext) { + switch (ext->sType) { + case VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_INLINE_UNIFORM_BLOCK_CREATE_INFO_EXT: { + const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT *info = + (const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT*)ext; + /* the sizes are 4 aligned, and we need to align to at + * most 32, which needs at most 28 bytes extra per + * binding. */ + bo_size += 28llu * info->maxInlineUniformBlockBindings; + break; + } + default: + break; + } + } for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) { if (pCreateInfo->pPoolSizes[i].type != VK_DESCRIPTOR_TYPE_SAMPLER) @@ -399,8 +681,10 @@ VkResult radv_CreateDescriptorPool( case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: bo_size += 96 * pCreateInfo->pPoolSizes[i].descriptorCount; break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + bo_size += pCreateInfo->pPoolSizes[i].descriptorCount; + break; default: - unreachable("unknown descriptor type\n"); break; } } @@ -410,15 +694,20 @@ VkResult radv_CreateDescriptorPool( host_size += sizeof(struct radeon_winsys_bo*) * bo_count; host_size += sizeof(struct radv_descriptor_range) * range_count; size += host_size; + } else { + size += sizeof(struct radv_descriptor_pool_entry) * pCreateInfo->maxSets; } - pool = vk_alloc2(&device->alloc, pAllocator, size, 8, + pool = vk_alloc2(&device->vk.alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!pool) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); memset(pool, 0, sizeof(*pool)); + vk_object_base_init(&device->vk, &pool->base, + VK_OBJECT_TYPE_DESCRIPTOR_POOL); + if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) { pool->host_memory_base = (uint8_t*)pool + sizeof(struct radv_descriptor_pool); pool->host_memory_ptr = pool->host_memory_base; @@ -426,13 +715,17 @@ VkResult radv_CreateDescriptorPool( } if (bo_size) { - pool->bo = device->ws->buffer_create(device->ws, bo_size, - 32, RADEON_DOMAIN_VRAM, 0); + pool->bo = device->ws->buffer_create(device->ws, bo_size, 32, + RADEON_DOMAIN_VRAM, + RADEON_FLAG_NO_INTERPROCESS_SHARING | + RADEON_FLAG_READ_ONLY | + RADEON_FLAG_32BIT, + RADV_BO_PRIORITY_DESCRIPTOR); pool->mapped_ptr = (uint8_t*)device->ws->buffer_map(pool->bo); } pool->size = bo_size; + pool->max_entry_count = pCreateInfo->maxSets; - list_inithead(&pool->vram_list); *pDescriptorPool = radv_descriptor_pool_to_handle(pool); return VK_SUCCESS; } @@ -449,15 +742,16 @@ void radv_DestroyDescriptorPool( return; if (!pool->host_memory_base) { - list_for_each_entry_safe(struct radv_descriptor_set, set, - &pool->vram_list, vram_list) { - radv_descriptor_set_destroy(device, pool, set, false); + for(int i = 0; i < pool->entry_count; ++i) { + radv_descriptor_set_destroy(device, pool, pool->entries[i].set, false); } } if (pool->bo) device->ws->buffer_destroy(pool->bo); - vk_free2(&device->alloc, pAllocator, pool); + + vk_object_base_finish(&pool->base); + vk_free2(&device->vk.alloc, pAllocator, pool); } VkResult radv_ResetDescriptorPool( @@ -469,14 +763,12 @@ VkResult radv_ResetDescriptorPool( RADV_FROM_HANDLE(radv_descriptor_pool, pool, descriptorPool); if (!pool->host_memory_base) { - list_for_each_entry_safe(struct radv_descriptor_set, set, - &pool->vram_list, vram_list) { - radv_descriptor_set_destroy(device, pool, set, false); + for(int i = 0; i < pool->entry_count; ++i) { + radv_descriptor_set_destroy(device, pool, pool->entries[i].set, false); } + pool->entry_count = 0; } - list_inithead(&pool->vram_list); - pool->current_offset = 0; pool->host_memory_ptr = pool->host_memory_base; @@ -493,25 +785,41 @@ VkResult radv_AllocateDescriptorSets( VkResult result = VK_SUCCESS; uint32_t i; - struct radv_descriptor_set *set; + struct radv_descriptor_set *set = NULL; + + const VkDescriptorSetVariableDescriptorCountAllocateInfo *variable_counts = + vk_find_struct_const(pAllocateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_ALLOCATE_INFO); + const uint32_t zero = 0; /* allocate a set of buffers for each shader to contain descriptors */ for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) { RADV_FROM_HANDLE(radv_descriptor_set_layout, layout, pAllocateInfo->pSetLayouts[i]); + const uint32_t *variable_count = NULL; + if (variable_counts) { + if (i < variable_counts->descriptorSetCount) + variable_count = variable_counts->pDescriptorCounts + i; + else + variable_count = &zero; + } + assert(!(layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR)); - result = radv_descriptor_set_create(device, pool, layout, &set); + result = radv_descriptor_set_create(device, pool, layout, variable_count, &set); if (result != VK_SUCCESS) break; pDescriptorSets[i] = radv_descriptor_set_to_handle(set); } - if (result != VK_SUCCESS) + if (result != VK_SUCCESS) { radv_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool, i, pDescriptorSets); + for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) { + pDescriptorSets[i] = VK_NULL_HANDLE; + } + } return result; } @@ -527,7 +835,7 @@ VkResult radv_FreeDescriptorSets( for (uint32_t i = 0; i < count; i++) { RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]); - if (set) + if (set && !pool->host_memory_base) radv_descriptor_set_destroy(device, pool, set, true); } return VK_SUCCESS; @@ -541,10 +849,15 @@ static void write_texel_buffer_descriptor(struct radv_device *device, { RADV_FROM_HANDLE(radv_buffer_view, buffer_view, _buffer_view); + if (!buffer_view) { + memset(dst, 0, 4 * 4); + return; + } + memcpy(dst, buffer_view->state, 4 * 4); if (cmd_buffer) - device->ws->cs_add_buffer(cmd_buffer->cs, buffer_view->bo, 7); + radv_cs_add_buffer(device->ws, cmd_buffer->cs, buffer_view->bo); else *buffer_list = buffer_view->bo; } @@ -556,12 +869,24 @@ static void write_buffer_descriptor(struct radv_device *device, const VkDescriptorBufferInfo *buffer_info) { RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer); - uint64_t va = device->ws->buffer_get_va(buffer->bo); + + if (!buffer) { + memset(dst, 0, 4 * 4); + return; + } + + uint64_t va = radv_buffer_get_va(buffer->bo); uint32_t range = buffer_info->range; if (buffer_info->range == VK_WHOLE_SIZE) range = buffer->size - buffer_info->offset; + /* robustBufferAccess is relaxed enough to allow this (in combination + * with the alignment/size we return from vkGetBufferMemoryRequirements) + * and this allows the shader compiler to create more efficient 8/16-bit + * buffer accesses. */ + range = align(range, 4); + va += buffer_info->offset + buffer->offset; dst[0] = va; dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32); @@ -569,28 +894,58 @@ static void write_buffer_descriptor(struct radv_device *device, dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) | S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | - S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) | - S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | - S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32); + S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W); + + if (device->physical_device->rad_info.chip_class >= GFX10) { + dst[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) | + S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | + S_008F0C_RESOURCE_LEVEL(1); + } else { + dst[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | + S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32); + } if (cmd_buffer) - device->ws->cs_add_buffer(cmd_buffer->cs, buffer->bo, 7); + radv_cs_add_buffer(device->ws, cmd_buffer->cs, buffer->bo); else *buffer_list = buffer->bo; } +static void write_block_descriptor(struct radv_device *device, + struct radv_cmd_buffer *cmd_buffer, + void *dst, + const VkWriteDescriptorSet *writeset) +{ + const VkWriteDescriptorSetInlineUniformBlockEXT *inline_ub = + vk_find_struct_const(writeset->pNext, WRITE_DESCRIPTOR_SET_INLINE_UNIFORM_BLOCK_EXT); + + memcpy(dst, inline_ub->pData, inline_ub->dataSize); +} + static void write_dynamic_buffer_descriptor(struct radv_device *device, struct radv_descriptor_range *range, struct radeon_winsys_bo **buffer_list, const VkDescriptorBufferInfo *buffer_info) { RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer); - uint64_t va = device->ws->buffer_get_va(buffer->bo); - unsigned size = buffer_info->range; + uint64_t va; + unsigned size; + + if (!buffer) + return; + + va = radv_buffer_get_va(buffer->bo); + size = buffer_info->range; if (buffer_info->range == VK_WHOLE_SIZE) size = buffer->size - buffer_info->offset; + /* robustBufferAccess is relaxed enough to allow this (in combination + * with the alignment/size we return from vkGetBufferMemoryRequirements) + * and this allows the shader compiler to create more efficient 8/16-bit + * buffer accesses. */ + size = align(size, 4); + va += buffer_info->offset + buffer->offset; range->va = va; range->size = size; @@ -601,16 +956,29 @@ static void write_dynamic_buffer_descriptor(struct radv_device *device, static void write_image_descriptor(struct radv_device *device, struct radv_cmd_buffer *cmd_buffer, - unsigned *dst, + unsigned size, unsigned *dst, struct radeon_winsys_bo **buffer_list, + VkDescriptorType descriptor_type, const VkDescriptorImageInfo *image_info) { RADV_FROM_HANDLE(radv_image_view, iview, image_info->imageView); - memcpy(dst, iview->descriptor, 8 * 4); - memcpy(dst + 8, iview->fmask_descriptor, 8 * 4); + union radv_descriptor *descriptor; + + if (!iview) { + memset(dst, 0, size); + return; + } + + if (descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) { + descriptor = &iview->storage_descriptor; + } else { + descriptor = &iview->descriptor; + } + + memcpy(dst, descriptor, size); if (cmd_buffer) - device->ws->cs_add_buffer(cmd_buffer->cs, iview->bo, 7); + radv_cs_add_buffer(device->ws, cmd_buffer->cs, iview->bo); else *buffer_list = iview->bo; } @@ -618,17 +986,21 @@ write_image_descriptor(struct radv_device *device, static void write_combined_image_sampler_descriptor(struct radv_device *device, struct radv_cmd_buffer *cmd_buffer, + unsigned sampler_offset, unsigned *dst, struct radeon_winsys_bo **buffer_list, + VkDescriptorType descriptor_type, const VkDescriptorImageInfo *image_info, bool has_sampler) { RADV_FROM_HANDLE(radv_sampler, sampler, image_info->sampler); - write_image_descriptor(device, cmd_buffer, dst, buffer_list, image_info); + write_image_descriptor(device, cmd_buffer, sampler_offset, dst, buffer_list, + descriptor_type, image_info); /* copy over sampler state */ - if (has_sampler) - memcpy(dst + 16, sampler->state, 16); + if (has_sampler) { + memcpy(dst + sampler_offset / sizeof(*dst), sampler->state, 16); + } } static void @@ -668,6 +1040,12 @@ void radv_update_descriptor_sets( const uint32_t *samplers = radv_immutable_samplers(set->layout, binding_layout); ptr += binding_layout->offset / 4; + + if (writeset->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) { + write_block_descriptor(device, cmd_buffer, (uint8_t*)ptr + writeset->dstArrayElement, writeset); + continue; + } + ptr += binding_layout->size * writeset->dstArrayElement / 4; buffer_list += binding_layout->buffer_offset; buffer_list += writeset->dstArrayElement; @@ -695,18 +1073,23 @@ void radv_update_descriptor_sets( case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: - write_image_descriptor(device, cmd_buffer, ptr, buffer_list, + write_image_descriptor(device, cmd_buffer, 64, ptr, buffer_list, + writeset->descriptorType, writeset->pImageInfo + j); break; - case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: - write_combined_image_sampler_descriptor(device, cmd_buffer, ptr, buffer_list, + case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: { + unsigned sampler_offset = radv_combined_image_descriptor_sampler_offset(binding_layout); + write_combined_image_sampler_descriptor(device, cmd_buffer, sampler_offset, + ptr, buffer_list, + writeset->descriptorType, writeset->pImageInfo + j, !binding_layout->immutable_samplers_offset); if (copy_immutable_samplers) { const unsigned idx = writeset->dstArrayElement + j; - memcpy(ptr + 16, samplers + 4 * idx, 16); + memcpy((char*)ptr + sampler_offset, samplers + 4 * idx, 16); } break; + } case VK_DESCRIPTOR_TYPE_SAMPLER: if (!binding_layout->immutable_samplers_offset) { write_sampler_descriptor(device, ptr, @@ -717,7 +1100,6 @@ void radv_update_descriptor_sets( } break; default: - unreachable("unimplemented descriptor type"); break; } ptr += binding_layout->size / 4; @@ -725,8 +1107,69 @@ void radv_update_descriptor_sets( } } - if (descriptorCopyCount) - radv_finishme("copy descriptors"); + + for (i = 0; i < descriptorCopyCount; i++) { + const VkCopyDescriptorSet *copyset = &pDescriptorCopies[i]; + RADV_FROM_HANDLE(radv_descriptor_set, src_set, + copyset->srcSet); + RADV_FROM_HANDLE(radv_descriptor_set, dst_set, + copyset->dstSet); + const struct radv_descriptor_set_binding_layout *src_binding_layout = + src_set->layout->binding + copyset->srcBinding; + const struct radv_descriptor_set_binding_layout *dst_binding_layout = + dst_set->layout->binding + copyset->dstBinding; + uint32_t *src_ptr = src_set->mapped_ptr; + uint32_t *dst_ptr = dst_set->mapped_ptr; + struct radeon_winsys_bo **src_buffer_list = src_set->descriptors; + struct radeon_winsys_bo **dst_buffer_list = dst_set->descriptors; + + src_ptr += src_binding_layout->offset / 4; + dst_ptr += dst_binding_layout->offset / 4; + + if (src_binding_layout->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) { + src_ptr += copyset->srcArrayElement / 4; + dst_ptr += copyset->dstArrayElement / 4; + + memcpy(dst_ptr, src_ptr, copyset->descriptorCount); + continue; + } + + src_ptr += src_binding_layout->size * copyset->srcArrayElement / 4; + dst_ptr += dst_binding_layout->size * copyset->dstArrayElement / 4; + + src_buffer_list += src_binding_layout->buffer_offset; + src_buffer_list += copyset->srcArrayElement; + + dst_buffer_list += dst_binding_layout->buffer_offset; + dst_buffer_list += copyset->dstArrayElement; + + for (j = 0; j < copyset->descriptorCount; ++j) { + switch (src_binding_layout->type) { + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: { + unsigned src_idx = copyset->srcArrayElement + j; + unsigned dst_idx = copyset->dstArrayElement + j; + struct radv_descriptor_range *src_range, *dst_range; + src_idx += src_binding_layout->dynamic_offset_offset; + dst_idx += dst_binding_layout->dynamic_offset_offset; + + src_range = src_set->dynamic_descriptors + src_idx; + dst_range = dst_set->dynamic_descriptors + dst_idx; + *dst_range = *src_range; + break; + } + default: + memcpy(dst_ptr, src_ptr, src_binding_layout->size); + } + src_ptr += src_binding_layout->size / 4; + dst_ptr += dst_binding_layout->size / 4; + + if (src_binding_layout->type != VK_DESCRIPTOR_TYPE_SAMPLER) { + /* Sampler descriptors don't have a buffer list. */ + dst_buffer_list[j] = src_buffer_list[j]; + } + } + } } void radv_UpdateDescriptorSets( @@ -742,10 +1185,10 @@ void radv_UpdateDescriptorSets( descriptorCopyCount, pDescriptorCopies); } -VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, - const VkDescriptorUpdateTemplateCreateInfoKHR *pCreateInfo, - const VkAllocationCallbacks *pAllocator, - VkDescriptorUpdateTemplateKHR *pDescriptorUpdateTemplate) +VkResult radv_CreateDescriptorUpdateTemplate(VkDevice _device, + const VkDescriptorUpdateTemplateCreateInfo *pCreateInfo, + const VkAllocationCallbacks *pAllocator, + VkDescriptorUpdateTemplate *pDescriptorUpdateTemplate) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout, pCreateInfo->descriptorSetLayout); @@ -755,14 +1198,29 @@ VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, struct radv_descriptor_update_template *templ; uint32_t i; - templ = vk_alloc2(&device->alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + templ = vk_alloc2(&device->vk.alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!templ) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); + + vk_object_base_init(&device->vk, &templ->base, + VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE); templ->entry_count = entry_count; + if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) { + RADV_FROM_HANDLE(radv_pipeline_layout, pipeline_layout, pCreateInfo->pipelineLayout); + + /* descriptorSetLayout should be ignored for push descriptors + * and instead it refers to pipelineLayout and set. + */ + assert(pCreateInfo->set < MAX_SETS); + set_layout = pipeline_layout->set[pCreateInfo->set].layout; + + templ->bind_point = pCreateInfo->pipelineBindPoint; + } + for (i = 0; i < entry_count; i++) { - const VkDescriptorUpdateTemplateEntryKHR *entry = &pCreateInfo->pDescriptorUpdateEntries[i]; + const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i]; const struct radv_descriptor_set_binding_layout *binding_layout = set_layout->binding + entry->dstBinding; const uint32_t buffer_offset = binding_layout->buffer_offset + entry->dstArrayElement; @@ -775,7 +1233,7 @@ VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, switch (entry->descriptorType) { case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: - assert(pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET_KHR); + assert(pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET); dst_offset = binding_layout->dynamic_offset_offset + entry->dstArrayElement; dst_stride = 0; /* Not used */ break; @@ -792,7 +1250,12 @@ VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, default: break; } - dst_offset = binding_layout->offset / 4 + binding_layout->size * entry->dstArrayElement / 4; + dst_offset = binding_layout->offset / 4; + if (entry->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) + dst_offset += entry->dstArrayElement / 4; + else + dst_offset += binding_layout->size * entry->dstArrayElement / 4; + dst_stride = binding_layout->size / 4; break; } @@ -806,6 +1269,7 @@ VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, .dst_stride = dst_stride, .buffer_offset = buffer_offset, .has_sampler = !binding_layout->immutable_samplers_offset, + .sampler_offset = radv_combined_image_descriptor_sampler_offset(binding_layout), .immutable_samplers = immutable_samplers }; } @@ -814,9 +1278,9 @@ VkResult radv_CreateDescriptorUpdateTemplateKHR(VkDevice _device, return VK_SUCCESS; } -void radv_DestroyDescriptorUpdateTemplateKHR(VkDevice _device, - VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate, - const VkAllocationCallbacks *pAllocator) +void radv_DestroyDescriptorUpdateTemplate(VkDevice _device, + VkDescriptorUpdateTemplate descriptorUpdateTemplate, + const VkAllocationCallbacks *pAllocator) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_descriptor_update_template, templ, descriptorUpdateTemplate); @@ -824,13 +1288,14 @@ void radv_DestroyDescriptorUpdateTemplateKHR(VkDevice _device, if (!templ) return; - vk_free2(&device->alloc, pAllocator, templ); + vk_object_base_finish(&templ->base); + vk_free2(&device->vk.alloc, pAllocator, templ); } void radv_update_descriptor_set_with_template(struct radv_device *device, struct radv_cmd_buffer *cmd_buffer, struct radv_descriptor_set *set, - VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate, + VkDescriptorUpdateTemplate descriptorUpdateTemplate, const void *pData) { RADV_FROM_HANDLE(radv_descriptor_update_template, templ, descriptorUpdateTemplate); @@ -842,6 +1307,11 @@ void radv_update_descriptor_set_with_template(struct radv_device *device, const uint8_t *pSrc = ((const uint8_t *) pData) + templ->entry[i].src_offset; uint32_t j; + if (templ->entry[i].descriptor_type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) { + memcpy((uint8_t*)pDst, pSrc, templ->entry[i].descriptor_count); + continue; + } + for (j = 0; j < templ->entry[i].descriptor_count; ++j) { switch (templ->entry[i].descriptor_type) { case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: @@ -865,15 +1335,18 @@ void radv_update_descriptor_set_with_template(struct radv_device *device, case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: - write_image_descriptor(device, cmd_buffer, pDst, buffer_list, + write_image_descriptor(device, cmd_buffer, 64, pDst, buffer_list, + templ->entry[i].descriptor_type, (struct VkDescriptorImageInfo *) pSrc); break; case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: - write_combined_image_sampler_descriptor(device, cmd_buffer, pDst, buffer_list, + write_combined_image_sampler_descriptor(device, cmd_buffer, templ->entry[i].sampler_offset, + pDst, buffer_list, templ->entry[i].descriptor_type, (struct VkDescriptorImageInfo *) pSrc, templ->entry[i].has_sampler); - if (templ->entry[i].immutable_samplers) - memcpy(pDst + 16, templ->entry[i].immutable_samplers + 4 * j, 16); + if (templ->entry[i].immutable_samplers) { + memcpy((char*)pDst + templ->entry[i].sampler_offset, templ->entry[i].immutable_samplers + 4 * j, 16); + } break; case VK_DESCRIPTOR_TYPE_SAMPLER: if (templ->entry[i].has_sampler) @@ -883,7 +1356,6 @@ void radv_update_descriptor_set_with_template(struct radv_device *device, memcpy(pDst, templ->entry[i].immutable_samplers + 4 * j, 16); break; default: - unreachable("unimplemented descriptor type"); break; } pSrc += templ->entry[i].src_stride; @@ -893,13 +1365,58 @@ void radv_update_descriptor_set_with_template(struct radv_device *device, } } -void radv_UpdateDescriptorSetWithTemplateKHR(VkDevice _device, - VkDescriptorSet descriptorSet, - VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate, - const void *pData) +void radv_UpdateDescriptorSetWithTemplate(VkDevice _device, + VkDescriptorSet descriptorSet, + VkDescriptorUpdateTemplate descriptorUpdateTemplate, + const void *pData) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_descriptor_set, set, descriptorSet); radv_update_descriptor_set_with_template(device, NULL, set, descriptorUpdateTemplate, pData); } + + +VkResult radv_CreateSamplerYcbcrConversion(VkDevice _device, + const VkSamplerYcbcrConversionCreateInfo* pCreateInfo, + const VkAllocationCallbacks* pAllocator, + VkSamplerYcbcrConversion* pYcbcrConversion) +{ + RADV_FROM_HANDLE(radv_device, device, _device); + struct radv_sampler_ycbcr_conversion *conversion = NULL; + + conversion = vk_zalloc2(&device->vk.alloc, pAllocator, sizeof(*conversion), 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + + if (conversion == NULL) + return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); + + vk_object_base_init(&device->vk, &conversion->base, + VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION); + + conversion->format = pCreateInfo->format; + conversion->ycbcr_model = pCreateInfo->ycbcrModel; + conversion->ycbcr_range = pCreateInfo->ycbcrRange; + conversion->components = pCreateInfo->components; + conversion->chroma_offsets[0] = pCreateInfo->xChromaOffset; + conversion->chroma_offsets[1] = pCreateInfo->yChromaOffset; + conversion->chroma_filter = pCreateInfo->chromaFilter; + + *pYcbcrConversion = radv_sampler_ycbcr_conversion_to_handle(conversion); + return VK_SUCCESS; +} + + +void radv_DestroySamplerYcbcrConversion(VkDevice _device, + VkSamplerYcbcrConversion ycbcrConversion, + const VkAllocationCallbacks* pAllocator) +{ + RADV_FROM_HANDLE(radv_device, device, _device); + RADV_FROM_HANDLE(radv_sampler_ycbcr_conversion, ycbcr_conversion, ycbcrConversion); + + if (!ycbcr_conversion) + return; + + vk_object_base_finish(&ycbcr_conversion->base); + vk_free2(&device->vk.alloc, pAllocator, ycbcr_conversion); +}