X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fintel%2Fvulkan%2Fanv_descriptor_set.c;h=cb6fa40b36894c5a328ef0c51f31cd455dae4047;hb=faa339e666b4d0c4f88f2588cd84176e2b19ec0f;hp=3f0e32355003271c097521c50a5ca7e507dcc00e;hpb=c520f4dec9cbedd4132143f52411df18f97869e6;p=mesa.git diff --git a/src/intel/vulkan/anv_descriptor_set.c b/src/intel/vulkan/anv_descriptor_set.c index 3f0e3235500..cb6fa40b368 100644 --- a/src/intel/vulkan/anv_descriptor_set.c +++ b/src/intel/vulkan/anv_descriptor_set.c @@ -28,6 +28,7 @@ #include #include "util/mesa-sha1.h" +#include "vk_util.h" #include "anv_private.h" @@ -44,15 +45,24 @@ anv_descriptor_data_for_type(const struct anv_physical_device *device, switch (type) { case VK_DESCRIPTOR_TYPE_SAMPLER: data = ANV_DESCRIPTOR_SAMPLER_STATE; + if (device->has_bindless_samplers) + data |= ANV_DESCRIPTOR_SAMPLED_IMAGE; break; case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: data = ANV_DESCRIPTOR_SURFACE_STATE | ANV_DESCRIPTOR_SAMPLER_STATE; + if (device->has_bindless_images || device->has_bindless_samplers) + data |= ANV_DESCRIPTOR_SAMPLED_IMAGE; break; case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER: + data = ANV_DESCRIPTOR_SURFACE_STATE; + if (device->has_bindless_images) + data |= ANV_DESCRIPTOR_SAMPLED_IMAGE; + break; + case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: data = ANV_DESCRIPTOR_SURFACE_STATE; break; @@ -62,6 +72,8 @@ anv_descriptor_data_for_type(const struct anv_physical_device *device, data = ANV_DESCRIPTOR_SURFACE_STATE; if (device->info.gen < 9) data |= ANV_DESCRIPTOR_IMAGE_PARAM; + if (device->has_bindless_images) + data |= ANV_DESCRIPTOR_STORAGE_IMAGE; break; case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: @@ -75,24 +87,88 @@ anv_descriptor_data_for_type(const struct anv_physical_device *device, data = ANV_DESCRIPTOR_SURFACE_STATE; break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + data = ANV_DESCRIPTOR_INLINE_UNIFORM; + break; + default: unreachable("Unsupported descriptor type"); } + /* On gen8 and above when we have softpin enabled, we also need to push + * SSBO address ranges so that we can use A64 messages in the shader. + */ + if (device->has_a64_buffer_access && + (type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER || + type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC)) + data |= ANV_DESCRIPTOR_ADDRESS_RANGE; + + /* On Ivy Bridge and Bay Trail, we need swizzles textures in the shader + * Do not handle VK_DESCRIPTOR_TYPE_STORAGE_IMAGE and + * VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT because they already must + * have identity swizzle. + */ + if (device->info.gen == 7 && !device->info.is_haswell && + (type == VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE || + type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)) + data |= ANV_DESCRIPTOR_TEXTURE_SWIZZLE; + return data; } static unsigned anv_descriptor_data_size(enum anv_descriptor_data data) { - return 0; + unsigned size = 0; + + if (data & ANV_DESCRIPTOR_SAMPLED_IMAGE) + size += sizeof(struct anv_sampled_image_descriptor); + + if (data & ANV_DESCRIPTOR_STORAGE_IMAGE) + size += sizeof(struct anv_storage_image_descriptor); + + if (data & ANV_DESCRIPTOR_IMAGE_PARAM) + size += BRW_IMAGE_PARAM_SIZE * 4; + + if (data & ANV_DESCRIPTOR_ADDRESS_RANGE) + size += sizeof(struct anv_address_range_descriptor); + + if (data & ANV_DESCRIPTOR_TEXTURE_SWIZZLE) + size += sizeof(struct anv_texture_swizzle_descriptor); + + return size; +} + +static bool +anv_needs_descriptor_buffer(VkDescriptorType desc_type, + enum anv_descriptor_data desc_data) +{ + if (desc_type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT || + anv_descriptor_data_size(desc_data) > 0) + return true; + return false; } /** Returns the size in bytes of each descriptor with the given layout */ unsigned anv_descriptor_size(const struct anv_descriptor_set_binding_layout *layout) { - return anv_descriptor_data_size(layout->data); + if (layout->data & ANV_DESCRIPTOR_INLINE_UNIFORM) { + assert(layout->data == ANV_DESCRIPTOR_INLINE_UNIFORM); + return layout->array_size; + } + + unsigned size = anv_descriptor_data_size(layout->data); + + /* For multi-planar bindings, we make every descriptor consume the maximum + * number of planes so we don't have to bother with walking arrays and + * adding things up every time. Fortunately, YCbCr samplers aren't all + * that common and likely won't be in the middle of big arrays. + */ + if (layout->max_plane_count > 1) + size *= layout->max_plane_count; + + return size; } /** Returns the size in bytes of each descriptor of the given type @@ -106,25 +182,96 @@ unsigned anv_descriptor_type_size(const struct anv_physical_device *pdevice, VkDescriptorType type) { + assert(type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT && + type != VK_DESCRIPTOR_TYPE_SAMPLER && + type != VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE && + type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); + return anv_descriptor_data_size(anv_descriptor_data_for_type(pdevice, type)); } +static bool +anv_descriptor_data_supports_bindless(const struct anv_physical_device *pdevice, + enum anv_descriptor_data data, + bool sampler) +{ + if (data & ANV_DESCRIPTOR_ADDRESS_RANGE) { + assert(pdevice->has_a64_buffer_access); + return true; + } + + if (data & ANV_DESCRIPTOR_SAMPLED_IMAGE) { + assert(pdevice->has_bindless_images || pdevice->has_bindless_samplers); + return sampler ? pdevice->has_bindless_samplers : + pdevice->has_bindless_images; + } + + if (data & ANV_DESCRIPTOR_STORAGE_IMAGE) { + assert(pdevice->has_bindless_images); + return true; + } + + return false; +} + +bool +anv_descriptor_supports_bindless(const struct anv_physical_device *pdevice, + const struct anv_descriptor_set_binding_layout *binding, + bool sampler) +{ + return anv_descriptor_data_supports_bindless(pdevice, binding->data, + sampler); +} + +bool +anv_descriptor_requires_bindless(const struct anv_physical_device *pdevice, + const struct anv_descriptor_set_binding_layout *binding, + bool sampler) +{ + if (pdevice->always_use_bindless) + return anv_descriptor_supports_bindless(pdevice, binding, sampler); + + static const VkDescriptorBindingFlagBitsEXT flags_requiring_bindless = + VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT | + VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT | + VK_DESCRIPTOR_BINDING_PARTIALLY_BOUND_BIT_EXT; + + return (binding->flags & flags_requiring_bindless) != 0; +} + void anv_GetDescriptorSetLayoutSupport( - VkDevice device, + VkDevice _device, const VkDescriptorSetLayoutCreateInfo* pCreateInfo, VkDescriptorSetLayoutSupport* pSupport) { + ANV_FROM_HANDLE(anv_device, device, _device); + const struct anv_physical_device *pdevice = device->physical; + uint32_t surface_count[MESA_SHADER_STAGES] = { 0, }; + bool needs_descriptor_buffer = false; for (uint32_t b = 0; b < pCreateInfo->bindingCount; b++) { const VkDescriptorSetLayoutBinding *binding = &pCreateInfo->pBindings[b]; + enum anv_descriptor_data desc_data = + anv_descriptor_data_for_type(pdevice, binding->descriptorType); + + if (anv_needs_descriptor_buffer(binding->descriptorType, desc_data)) + needs_descriptor_buffer = true; + switch (binding->descriptorType) { case VK_DESCRIPTOR_TYPE_SAMPLER: /* There is no real limit on samplers */ break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + /* Inline uniforms don't use a binding */ + break; + case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: + if (anv_descriptor_data_supports_bindless(pdevice, desc_data, false)) + break; + if (binding->pImmutableSamplers) { for (uint32_t i = 0; i < binding->descriptorCount; i++) { ANV_FROM_HANDLE(anv_sampler, sampler, @@ -139,18 +286,26 @@ void anv_GetDescriptorSetLayoutSupport( break; default: + if (anv_descriptor_data_supports_bindless(pdevice, desc_data, false)) + break; + anv_foreach_stage(s, binding->stageFlags) surface_count[s] += binding->descriptorCount; break; } } + for (unsigned s = 0; s < MESA_SHADER_STAGES; s++) { + if (needs_descriptor_buffer) + surface_count[s] += 1; + } + bool supported = true; for (unsigned s = 0; s < MESA_SHADER_STAGES; s++) { - /* Our maximum binding table size is 250 and we need to reserve 8 for - * render targets. 240 is a nice round number. + /* Our maximum binding table size is 240 and we need to reserve 8 for + * render targets. */ - if (surface_count[s] >= 240) + if (surface_count[s] > MAX_BINDING_TABLE_SIZE - MAX_RTS) supported = false; } @@ -203,11 +358,13 @@ VkResult anv_CreateDescriptorSetLayout( anv_multialloc_add(&ma, &bindings, max_binding + 1); anv_multialloc_add(&ma, &samplers, immutable_sampler_count); - if (!anv_multialloc_alloc(&ma, &device->alloc, + if (!anv_multialloc_alloc(&ma, &device->vk.alloc, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE)) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); memset(set_layout, 0, sizeof(*set_layout)); + vk_object_base_init(&device->vk, &set_layout->base, + VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT); set_layout->ref_cnt = 1; set_layout->binding_count = max_binding + 1; @@ -215,7 +372,9 @@ VkResult anv_CreateDescriptorSetLayout( /* Initialize all binding_layout entries to -1 */ memset(&set_layout->binding[b], -1, sizeof(set_layout->binding[b])); + set_layout->binding[b].flags = 0; set_layout->binding[b].data = 0; + set_layout->binding[b].max_plane_count = 0; set_layout->binding[b].array_size = 0; set_layout->binding[b].immutable_samplers = NULL; } @@ -230,34 +389,46 @@ VkResult anv_CreateDescriptorSetLayout( for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) { const VkDescriptorSetLayoutBinding *binding = &pCreateInfo->pBindings[j]; uint32_t b = binding->binding; - /* We temporarily store the pointer to the binding in the + /* We temporarily store pCreateInfo->pBindings[] index (plus one) in the * immutable_samplers pointer. This provides us with a quick-and-dirty * way to sort the bindings by binding number. */ - set_layout->binding[b].immutable_samplers = (void *)binding; + set_layout->binding[b].immutable_samplers = (void *)(uintptr_t)(j + 1); } - for (uint32_t b = 0; b <= max_binding; b++) { - const VkDescriptorSetLayoutBinding *binding = - (void *)set_layout->binding[b].immutable_samplers; + const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *binding_flags_info = + vk_find_struct_const(pCreateInfo->pNext, + DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT); - if (binding == NULL) - continue; - - /* We temporarily stashed the pointer to the binding in the - * immutable_samplers pointer. Now that we've pulled it back out - * again, we reset immutable_samplers to NULL. + for (uint32_t b = 0; b <= max_binding; b++) { + /* We stashed the pCreateInfo->pBindings[] index (plus one) in the + * immutable_samplers pointer. Check for NULL (empty binding) and then + * reset it and compute the index. */ + if (set_layout->binding[b].immutable_samplers == NULL) + continue; + const uint32_t info_idx = + (uintptr_t)(void *)set_layout->binding[b].immutable_samplers - 1; set_layout->binding[b].immutable_samplers = NULL; + const VkDescriptorSetLayoutBinding *binding = + &pCreateInfo->pBindings[info_idx]; + if (binding->descriptorCount == 0) continue; #ifndef NDEBUG set_layout->binding[b].type = binding->descriptorType; #endif + + if (binding_flags_info && binding_flags_info->bindingCount > 0) { + assert(binding_flags_info->bindingCount == pCreateInfo->bindingCount); + set_layout->binding[b].flags = + binding_flags_info->pBindingFlags[info_idx]; + } + set_layout->binding[b].data = - anv_descriptor_data_for_type(&device->instance->physicalDevice, + anv_descriptor_data_for_type(device->physical, binding->descriptorType); set_layout->binding[b].array_size = binding->descriptorCount; set_layout->binding[b].descriptor_index = set_layout->size; @@ -271,15 +442,26 @@ VkResult anv_CreateDescriptorSetLayout( switch (binding->descriptorType) { case VK_DESCRIPTOR_TYPE_SAMPLER: case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: + set_layout->binding[b].max_plane_count = 1; if (binding->pImmutableSamplers) { set_layout->binding[b].immutable_samplers = samplers; samplers += binding->descriptorCount; - for (uint32_t i = 0; i < binding->descriptorCount; i++) - set_layout->binding[b].immutable_samplers[i] = - anv_sampler_from_handle(binding->pImmutableSamplers[i]); + for (uint32_t i = 0; i < binding->descriptorCount; i++) { + ANV_FROM_HANDLE(anv_sampler, sampler, + binding->pImmutableSamplers[i]); + + set_layout->binding[b].immutable_samplers[i] = sampler; + if (set_layout->binding[b].max_plane_count < sampler->n_planes) + set_layout->binding[b].max_plane_count = sampler->n_planes; + } } break; + + case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: + set_layout->binding[b].max_plane_count = 1; + break; + default: break; } @@ -288,16 +470,34 @@ VkResult anv_CreateDescriptorSetLayout( case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: set_layout->binding[b].dynamic_offset_index = dynamic_offset_count; + anv_foreach_stage(s, binding->stageFlags) { + STATIC_ASSERT(MAX_DYNAMIC_BUFFERS <= + sizeof(set_layout->stage_dynamic_offsets[s]) * 8); + set_layout->stage_dynamic_offsets[s] |= + BITFIELD_RANGE(set_layout->binding[b].dynamic_offset_index, + binding->descriptorCount); + } dynamic_offset_count += binding->descriptorCount; + assert(dynamic_offset_count < MAX_DYNAMIC_BUFFERS); break; default: break; } - set_layout->binding[b].descriptor_offset = descriptor_buffer_size; - descriptor_buffer_size += anv_descriptor_size(&set_layout->binding[b]) * - binding->descriptorCount; + if (binding->descriptorType == + VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) { + /* Inline uniform blocks are specified to use the descriptor array + * size as the size in bytes of the block. + */ + descriptor_buffer_size = align_u32(descriptor_buffer_size, 32); + set_layout->binding[b].descriptor_offset = descriptor_buffer_size; + descriptor_buffer_size += binding->descriptorCount; + } else { + set_layout->binding[b].descriptor_offset = descriptor_buffer_size; + descriptor_buffer_size += anv_descriptor_size(&set_layout->binding[b]) * + binding->descriptorCount; + } set_layout->shader_stages |= binding->stageFlags; } @@ -311,6 +511,15 @@ VkResult anv_CreateDescriptorSetLayout( return VK_SUCCESS; } +void +anv_descriptor_set_layout_destroy(struct anv_device *device, + struct anv_descriptor_set_layout *layout) +{ + assert(layout->ref_cnt == 0); + vk_object_base_finish(&layout->base); + vk_free(&device->vk.alloc, layout); +} + void anv_DestroyDescriptorSetLayout( VkDevice _device, VkDescriptorSetLayout _set_layout, @@ -343,7 +552,9 @@ static void sha1_update_descriptor_set_binding_layout(struct mesa_sha1 *ctx, const struct anv_descriptor_set_binding_layout *layout) { + SHA1_UPDATE_VALUE(ctx, layout->flags); SHA1_UPDATE_VALUE(ctx, layout->data); + SHA1_UPDATE_VALUE(ctx, layout->max_plane_count); SHA1_UPDATE_VALUE(ctx, layout->array_size); SHA1_UPDATE_VALUE(ctx, layout->descriptor_index); SHA1_UPDATE_VALUE(ctx, layout->dynamic_offset_index); @@ -387,11 +598,13 @@ VkResult anv_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); + vk_object_base_init(&device->vk, &layout->base, + VK_OBJECT_TYPE_PIPELINE_LAYOUT); layout->num_sets = pCreateInfo->setLayoutCount; unsigned dynamic_offset_count = 0; @@ -410,6 +623,7 @@ VkResult anv_CreatePipelineLayout( dynamic_offset_count += set_layout->binding[b].array_size; } } + assert(dynamic_offset_count < MAX_DYNAMIC_BUFFERS); struct mesa_sha1 ctx; _mesa_sha1_init(&ctx); @@ -440,7 +654,8 @@ void anv_DestroyPipelineLayout( for (uint32_t i = 0; i < pipeline_layout->num_sets; i++) anv_descriptor_set_layout_unref(device, pipeline_layout->set[i].layout); - vk_free2(&device->alloc, pAllocator, pipeline_layout); + vk_object_base_finish(&pipeline_layout->base); + vk_free2(&device->vk.alloc, pAllocator, pipeline_layout); } /* @@ -472,12 +687,16 @@ VkResult anv_CreateDescriptorPool( ANV_FROM_HANDLE(anv_device, device, _device); struct anv_descriptor_pool *pool; + const VkDescriptorPoolInlineUniformBlockCreateInfoEXT *inline_info = + vk_find_struct_const(pCreateInfo->pNext, + DESCRIPTOR_POOL_INLINE_UNIFORM_BLOCK_CREATE_INFO_EXT); + uint32_t descriptor_count = 0; uint32_t buffer_view_count = 0; uint32_t descriptor_bo_size = 0; for (uint32_t i = 0; i < pCreateInfo->poolSizeCount; i++) { enum anv_descriptor_data desc_data = - anv_descriptor_data_for_type(&device->instance->physicalDevice, + anv_descriptor_data_for_type(device->physical, pCreateInfo->pPoolSizes[i].type); if (desc_data & ANV_DESCRIPTOR_BUFFER_VIEW) @@ -485,6 +704,23 @@ VkResult anv_CreateDescriptorPool( unsigned desc_data_size = anv_descriptor_data_size(desc_data) * pCreateInfo->pPoolSizes[i].descriptorCount; + + /* Combined image sampler descriptors can take up to 3 slots if they + * hold a YCbCr image. + */ + if (pCreateInfo->pPoolSizes[i].type == + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) + desc_data_size *= 3; + + if (pCreateInfo->pPoolSizes[i].type == + VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) { + /* Inline uniform blocks are specified to use the descriptor array + * size as the size in bytes of the block. + */ + assert(inline_info); + desc_data_size += pCreateInfo->pPoolSizes[i].descriptorCount; + } + descriptor_bo_size += desc_data_size; descriptor_count += pCreateInfo->pPoolSizes[i].descriptorCount; @@ -498,6 +734,9 @@ VkResult anv_CreateDescriptorPool( * of them to 32B. */ descriptor_bo_size += 32 * pCreateInfo->maxSets; + /* We align inline uniform blocks to 32B */ + if (inline_info) + descriptor_bo_size += 32 * inline_info->maxInlineUniformBlockBindings; descriptor_bo_size = ALIGN(descriptor_bo_size, 4096); const size_t pool_size = @@ -506,46 +745,40 @@ VkResult anv_CreateDescriptorPool( buffer_view_count * sizeof(struct anv_buffer_view); const size_t total_size = sizeof(*pool) + pool_size; - pool = vk_alloc2(&device->alloc, pAllocator, total_size, 8, + pool = vk_alloc2(&device->vk.alloc, pAllocator, total_size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!pool) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + vk_object_base_init(&device->vk, &pool->base, + VK_OBJECT_TYPE_DESCRIPTOR_POOL); pool->size = pool_size; pool->next = 0; pool->free_list = EMPTY; if (descriptor_bo_size > 0) { - VkResult result = anv_bo_init_new(&pool->bo, device, descriptor_bo_size); + VkResult result = anv_device_alloc_bo(device, + descriptor_bo_size, + ANV_BO_ALLOC_MAPPED | + ANV_BO_ALLOC_SNOOPED, + 0 /* explicit_address */, + &pool->bo); if (result != VK_SUCCESS) { - vk_free2(&device->alloc, pAllocator, pool); + vk_free2(&device->vk.alloc, pAllocator, pool); return result; } - anv_gem_set_caching(device, pool->bo.gem_handle, I915_CACHING_CACHED); - - pool->bo.map = anv_gem_mmap(device, pool->bo.gem_handle, 0, - descriptor_bo_size, 0); - if (pool->bo.map == NULL) { - anv_gem_close(device, pool->bo.gem_handle); - vk_free2(&device->alloc, pAllocator, pool); - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - } - - if (device->instance->physicalDevice.use_softpin) { - pool->bo.flags |= EXEC_OBJECT_PINNED; - anv_vma_alloc(device, &pool->bo); - } - util_vma_heap_init(&pool->bo_heap, POOL_HEAP_OFFSET, descriptor_bo_size); } else { - pool->bo.size = 0; + pool->bo = NULL; } anv_state_stream_init(&pool->surface_state_stream, &device->surface_state_pool, 4096); pool->surface_state_free_list = NULL; + list_inithead(&pool->desc_sets); + *pDescriptorPool = anv_descriptor_pool_to_handle(pool); return VK_SUCCESS; @@ -562,13 +795,17 @@ void anv_DestroyDescriptorPool( if (!pool) return; - if (pool->bo.size) { - anv_gem_munmap(pool->bo.map, pool->bo.size); - anv_vma_free(device, &pool->bo); - anv_gem_close(device, pool->bo.gem_handle); + list_for_each_entry_safe(struct anv_descriptor_set, set, + &pool->desc_sets, pool_link) { + anv_descriptor_set_layout_unref(device, set->layout); } + + if (pool->bo) + anv_device_release_bo(device, pool->bo); anv_state_stream_finish(&pool->surface_state_stream); - vk_free2(&device->alloc, pAllocator, pool); + + vk_object_base_finish(&pool->base); + vk_free2(&device->vk.alloc, pAllocator, pool); } VkResult anv_ResetDescriptorPool( @@ -579,12 +816,18 @@ VkResult anv_ResetDescriptorPool( ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_descriptor_pool, pool, descriptorPool); + list_for_each_entry_safe(struct anv_descriptor_set, set, + &pool->desc_sets, pool_link) { + anv_descriptor_set_layout_unref(device, set->layout); + } + list_inithead(&pool->desc_sets); + pool->next = 0; pool->free_list = EMPTY; - if (pool->bo.size) { + if (pool->bo) { util_vma_heap_finish(&pool->bo_heap); - util_vma_heap_init(&pool->bo_heap, POOL_HEAP_OFFSET, pool->bo.size); + util_vma_heap_init(&pool->bo_heap, POOL_HEAP_OFFSET, pool->bo->size); } anv_state_stream_finish(&pool->surface_state_stream); @@ -704,9 +947,9 @@ anv_descriptor_set_create(struct anv_device *device, /* Align the size to 32 so that alignment gaps don't cause extra holes * in the heap which can lead to bad performance. */ + uint32_t set_buffer_size = ALIGN(layout->descriptor_buffer_size, 32); uint64_t pool_vma_offset = - util_vma_heap_alloc(&pool->bo_heap, - ALIGN(layout->descriptor_buffer_size, 32), 32); + util_vma_heap_alloc(&pool->bo_heap, set_buffer_size, 32); if (pool_vma_offset == 0) { anv_descriptor_pool_free_set(pool, set); return vk_error(VK_ERROR_FRAGMENTED_POOL); @@ -714,14 +957,14 @@ anv_descriptor_set_create(struct anv_device *device, assert(pool_vma_offset >= POOL_HEAP_OFFSET && pool_vma_offset - POOL_HEAP_OFFSET <= INT32_MAX); set->desc_mem.offset = pool_vma_offset - POOL_HEAP_OFFSET; - set->desc_mem.alloc_size = layout->descriptor_buffer_size; - set->desc_mem.map = pool->bo.map + set->desc_mem.offset; + set->desc_mem.alloc_size = set_buffer_size; + set->desc_mem.map = pool->bo->map + set->desc_mem.offset; set->desc_surface_state = anv_descriptor_pool_alloc_state(pool); anv_fill_buffer_surface_state(device, set->desc_surface_state, ISL_FORMAT_R32G32B32A32_FLOAT, (struct anv_address) { - .bo = &pool->bo, + .bo = pool->bo, .offset = set->desc_mem.offset, }, layout->descriptor_buffer_size, 1); @@ -730,6 +973,8 @@ anv_descriptor_set_create(struct anv_device *device, set->desc_surface_state = ANV_STATE_NULL; } + vk_object_base_init(&device->vk, &set->base, + VK_OBJECT_TYPE_DESCRIPTOR_SET); set->pool = pool; set->layout = layout; anv_descriptor_set_layout_ref(layout); @@ -753,11 +998,15 @@ anv_descriptor_set_create(struct anv_device *device, * UpdateDescriptorSets if needed. However, if the descriptor * set has an immutable sampler, UpdateDescriptorSets may never * touch it, so we need to make sure it's 100% valid now. + * + * We don't need to actually provide a sampler because the helper + * will always write in the immutable sampler regardless of what + * is in the sampler parameter. */ - desc[i] = (struct anv_descriptor) { - .type = VK_DESCRIPTOR_TYPE_SAMPLER, - .sampler = layout->binding[b].immutable_samplers[i], - }; + VkDescriptorImageInfo info = { }; + anv_descriptor_set_write_image_view(device, set, &info, + VK_DESCRIPTOR_TYPE_SAMPLER, + b, i); } } desc += layout->binding[b].array_size; @@ -769,6 +1018,8 @@ anv_descriptor_set_create(struct anv_device *device, anv_descriptor_pool_alloc_state(pool); } + list_addtail(&set->pool_link, &pool->desc_sets); + *out_set = set; return VK_SUCCESS; @@ -791,6 +1042,9 @@ anv_descriptor_set_destroy(struct anv_device *device, for (uint32_t b = 0; b < set->buffer_view_count; b++) anv_descriptor_pool_free_state(pool, set->buffer_views[b].surface_state); + list_del(&set->pool_link); + + vk_object_base_finish(&set->base); anv_descriptor_pool_free_set(pool, set); } @@ -845,6 +1099,36 @@ VkResult anv_FreeDescriptorSets( return VK_SUCCESS; } +static void +anv_descriptor_set_write_image_param(uint32_t *param_desc_map, + const struct brw_image_param *param) +{ +#define WRITE_PARAM_FIELD(field, FIELD) \ + for (unsigned i = 0; i < ARRAY_SIZE(param->field); i++) \ + param_desc_map[BRW_IMAGE_PARAM_##FIELD##_OFFSET + i] = param->field[i] + + WRITE_PARAM_FIELD(offset, OFFSET); + WRITE_PARAM_FIELD(size, SIZE); + WRITE_PARAM_FIELD(stride, STRIDE); + WRITE_PARAM_FIELD(tiling, TILING); + WRITE_PARAM_FIELD(swizzling, SWIZZLING); + WRITE_PARAM_FIELD(size, SIZE); + +#undef WRITE_PARAM_FIELD +} + +static uint32_t +anv_surface_state_to_handle(struct anv_state state) +{ + /* Bits 31:12 of the bindless surface offset in the extended message + * descriptor is bits 25:6 of the byte-based address. + */ + assert(state.offset >= 0); + uint32_t offset = state.offset; + assert((offset & 0x3f) == 0 && offset < (1 << 26)); + return offset << 6; +} + void anv_descriptor_set_write_image_view(struct anv_device *device, struct anv_descriptor_set *set, @@ -860,16 +1144,24 @@ anv_descriptor_set_write_image_view(struct anv_device *device, struct anv_image_view *image_view = NULL; struct anv_sampler *sampler = NULL; - assert(type == bind_layout->type); + /* We get called with just VK_DESCRIPTOR_TYPE_SAMPLER as part of descriptor + * set initialization to set the bindless samplers. + */ + assert(type == bind_layout->type || + type == VK_DESCRIPTOR_TYPE_SAMPLER); switch (type) { case VK_DESCRIPTOR_TYPE_SAMPLER: - sampler = anv_sampler_from_handle(info->sampler); + sampler = bind_layout->immutable_samplers ? + bind_layout->immutable_samplers[element] : + anv_sampler_from_handle(info->sampler); break; case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: image_view = anv_image_view_from_handle(info->imageView); - sampler = anv_sampler_from_handle(info->sampler); + sampler = bind_layout->immutable_samplers ? + bind_layout->immutable_samplers[element] : + anv_sampler_from_handle(info->sampler); break; case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: @@ -882,19 +1174,87 @@ anv_descriptor_set_write_image_view(struct anv_device *device, unreachable("invalid descriptor type"); } - /* If this descriptor has an immutable sampler, we don't want to stomp on - * it. - */ - sampler = bind_layout->immutable_samplers ? - bind_layout->immutable_samplers[element] : - sampler; - *desc = (struct anv_descriptor) { .type = type, .layout = info->imageLayout, .image_view = image_view, .sampler = sampler, }; + + void *desc_map = set->desc_mem.map + bind_layout->descriptor_offset + + element * anv_descriptor_size(bind_layout); + memset(desc_map, 0, anv_descriptor_size(bind_layout)); + + if (bind_layout->data & ANV_DESCRIPTOR_SAMPLED_IMAGE) { + struct anv_sampled_image_descriptor desc_data[3]; + memset(desc_data, 0, sizeof(desc_data)); + + if (image_view) { + for (unsigned p = 0; p < image_view->n_planes; p++) { + struct anv_surface_state sstate = + (desc->layout == VK_IMAGE_LAYOUT_GENERAL) ? + image_view->planes[p].general_sampler_surface_state : + image_view->planes[p].optimal_sampler_surface_state; + desc_data[p].image = anv_surface_state_to_handle(sstate.state); + } + } + + if (sampler) { + for (unsigned p = 0; p < sampler->n_planes; p++) + desc_data[p].sampler = sampler->bindless_state.offset + p * 32; + } + + /* We may have max_plane_count < 0 if this isn't a sampled image but it + * can be no more than the size of our array of handles. + */ + assert(bind_layout->max_plane_count <= ARRAY_SIZE(desc_data)); + memcpy(desc_map, desc_data, + MAX2(1, bind_layout->max_plane_count) * sizeof(desc_data[0])); + } + + if (image_view == NULL) + return; + + if (bind_layout->data & ANV_DESCRIPTOR_STORAGE_IMAGE) { + assert(!(bind_layout->data & ANV_DESCRIPTOR_IMAGE_PARAM)); + assert(image_view->n_planes == 1); + struct anv_storage_image_descriptor desc_data = { + .read_write = anv_surface_state_to_handle( + image_view->planes[0].storage_surface_state.state), + .write_only = anv_surface_state_to_handle( + image_view->planes[0].writeonly_storage_surface_state.state), + }; + memcpy(desc_map, &desc_data, sizeof(desc_data)); + } + + if (bind_layout->data & ANV_DESCRIPTOR_IMAGE_PARAM) { + /* Storage images can only ever have one plane */ + assert(image_view->n_planes == 1); + const struct brw_image_param *image_param = + &image_view->planes[0].storage_image_param; + + anv_descriptor_set_write_image_param(desc_map, image_param); + } + + if (bind_layout->data & ANV_DESCRIPTOR_TEXTURE_SWIZZLE) { + assert(!(bind_layout->data & ANV_DESCRIPTOR_SAMPLED_IMAGE)); + assert(image_view); + struct anv_texture_swizzle_descriptor desc_data[3]; + memset(desc_data, 0, sizeof(desc_data)); + + for (unsigned p = 0; p < image_view->n_planes; p++) { + desc_data[p] = (struct anv_texture_swizzle_descriptor) { + .swizzle = { + (uint8_t)image_view->planes[p].isl.swizzle.r, + (uint8_t)image_view->planes[p].isl.swizzle.g, + (uint8_t)image_view->planes[p].isl.swizzle.b, + (uint8_t)image_view->planes[p].isl.swizzle.a, + }, + }; + } + memcpy(desc_map, desc_data, + MAX2(1, bind_layout->max_plane_count) * sizeof(desc_data[0])); + } } void @@ -912,10 +1272,42 @@ anv_descriptor_set_write_buffer_view(struct anv_device *device, assert(type == bind_layout->type); + void *desc_map = set->desc_mem.map + bind_layout->descriptor_offset + + element * anv_descriptor_size(bind_layout); + + if (buffer_view == NULL) { + *desc = (struct anv_descriptor) { .type = type, }; + memset(desc_map, 0, anv_descriptor_size(bind_layout)); + return; + } + *desc = (struct anv_descriptor) { .type = type, .buffer_view = buffer_view, }; + + if (bind_layout->data & ANV_DESCRIPTOR_SAMPLED_IMAGE) { + struct anv_sampled_image_descriptor desc_data = { + .image = anv_surface_state_to_handle(buffer_view->surface_state), + }; + memcpy(desc_map, &desc_data, sizeof(desc_data)); + } + + if (bind_layout->data & ANV_DESCRIPTOR_STORAGE_IMAGE) { + assert(!(bind_layout->data & ANV_DESCRIPTOR_IMAGE_PARAM)); + struct anv_storage_image_descriptor desc_data = { + .read_write = anv_surface_state_to_handle( + buffer_view->storage_surface_state), + .write_only = anv_surface_state_to_handle( + buffer_view->writeonly_storage_surface_state), + }; + memcpy(desc_map, &desc_data, sizeof(desc_data)); + } + + if (bind_layout->data & ANV_DESCRIPTOR_IMAGE_PARAM) { + anv_descriptor_set_write_image_param(desc_map, + &buffer_view->storage_image_param); + } } void @@ -936,6 +1328,25 @@ anv_descriptor_set_write_buffer(struct anv_device *device, assert(type == bind_layout->type); + void *desc_map = set->desc_mem.map + bind_layout->descriptor_offset + + element * anv_descriptor_size(bind_layout); + + if (buffer == NULL) { + *desc = (struct anv_descriptor) { .type = type, }; + memset(desc_map, 0, anv_descriptor_size(bind_layout)); + return; + } + + struct anv_address bind_addr = anv_address_add(buffer->address, offset); + uint64_t bind_range = anv_buffer_get_range(buffer, offset, range); + + /* We report a bounds checking alignment of 32B for the sake of block + * messages which read an entire register worth at a time. + */ + if (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER || + type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) + bind_range = align_u64(bind_range, ANV_UBO_ALIGNMENT); + if (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC || type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) { *desc = (struct anv_descriptor) { @@ -950,8 +1361,8 @@ anv_descriptor_set_write_buffer(struct anv_device *device, &set->buffer_views[bind_layout->buffer_view_index + element]; bview->format = anv_isl_format_for_descriptor_type(type); - bview->range = anv_buffer_get_range(buffer, offset, range); - bview->address = anv_address_add(buffer->address, offset); + bview->range = bind_range; + bview->address = bind_addr; /* If we're writing descriptors through a push command, we need to * allocate the surface state from the command buffer. Otherwise it will @@ -961,14 +1372,39 @@ anv_descriptor_set_write_buffer(struct anv_device *device, bview->surface_state = anv_state_stream_alloc(alloc_stream, 64, 64); anv_fill_buffer_surface_state(device, bview->surface_state, - bview->format, - bview->address, bview->range, 1); + bview->format, bind_addr, bind_range, 1); *desc = (struct anv_descriptor) { .type = type, .buffer_view = bview, }; } + + if (bind_layout->data & ANV_DESCRIPTOR_ADDRESS_RANGE) { + struct anv_address_range_descriptor desc_data = { + .address = anv_address_physical(bind_addr), + .range = bind_range, + }; + memcpy(desc_map, &desc_data, sizeof(desc_data)); + } +} + +void +anv_descriptor_set_write_inline_uniform_data(struct anv_device *device, + struct anv_descriptor_set *set, + uint32_t binding, + const void *data, + size_t offset, + size_t size) +{ + const struct anv_descriptor_set_binding_layout *bind_layout = + &set->layout->binding[binding]; + + assert(bind_layout->data & ANV_DESCRIPTOR_INLINE_UNIFORM); + + void *desc_map = set->desc_mem.map + bind_layout->descriptor_offset; + + memcpy(desc_map + offset, data, size); } void anv_UpdateDescriptorSets( @@ -1018,9 +1454,7 @@ void anv_UpdateDescriptorSets( case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: for (uint32_t j = 0; j < write->descriptorCount; j++) { - assert(write->pBufferInfo[j].buffer); ANV_FROM_HANDLE(anv_buffer, buffer, write->pBufferInfo[j].buffer); - assert(buffer); anv_descriptor_set_write_buffer(device, set, NULL, @@ -1033,6 +1467,19 @@ void anv_UpdateDescriptorSets( } break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: { + const VkWriteDescriptorSetInlineUniformBlockEXT *inline_write = + vk_find_struct_const(write->pNext, + WRITE_DESCRIPTOR_SET_INLINE_UNIFORM_BLOCK_EXT); + assert(inline_write->dataSize == write->descriptorCount); + anv_descriptor_set_write_inline_uniform_data(device, set, + write->dstBinding, + inline_write->pData, + write->dstArrayElement, + inline_write->dataSize); + break; + } + default: break; } @@ -1055,17 +1502,26 @@ void anv_UpdateDescriptorSets( &dst->descriptors[dst_layout->descriptor_index]; dst_desc += copy->dstArrayElement; - for (uint32_t j = 0; j < copy->descriptorCount; j++) - dst_desc[j] = src_desc[j]; - - unsigned desc_size = anv_descriptor_size(src_layout); - if (desc_size > 0) { - assert(desc_size == anv_descriptor_size(dst_layout)); + if (src_layout->data & ANV_DESCRIPTOR_INLINE_UNIFORM) { + assert(src_layout->data == ANV_DESCRIPTOR_INLINE_UNIFORM); memcpy(dst->desc_mem.map + dst_layout->descriptor_offset + - copy->dstArrayElement * desc_size, + copy->dstArrayElement, src->desc_mem.map + src_layout->descriptor_offset + - copy->srcArrayElement * desc_size, - copy->descriptorCount * desc_size); + copy->srcArrayElement, + copy->descriptorCount); + } else { + for (uint32_t j = 0; j < copy->descriptorCount; j++) + dst_desc[j] = src_desc[j]; + + unsigned desc_size = anv_descriptor_size(src_layout); + if (desc_size > 0) { + assert(desc_size == anv_descriptor_size(dst_layout)); + memcpy(dst->desc_mem.map + dst_layout->descriptor_offset + + copy->dstArrayElement * desc_size, + src->desc_mem.map + src_layout->descriptor_offset + + copy->srcArrayElement * desc_size, + copy->descriptorCount * desc_size); + } } } } @@ -1135,6 +1591,14 @@ anv_descriptor_set_write_template(struct anv_device *device, } break; + case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: + anv_descriptor_set_write_inline_uniform_data(device, set, + entry->binding, + data + entry->offset, + entry->array_element, + entry->array_count); + break; + default: break; } @@ -1152,11 +1616,13 @@ VkResult anv_CreateDescriptorUpdateTemplate( size_t size = sizeof(*template) + pCreateInfo->descriptorUpdateEntryCount * sizeof(template->entries[0]); - template = vk_alloc2(&device->alloc, pAllocator, size, 8, + template = vk_alloc2(&device->vk.alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (template == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + vk_object_base_init(&device->vk, &template->base, + VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE); template->bind_point = pCreateInfo->pipelineBindPoint; if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET) @@ -1192,7 +1658,8 @@ void anv_DestroyDescriptorUpdateTemplate( ANV_FROM_HANDLE(anv_descriptor_update_template, template, descriptorUpdateTemplate); - vk_free2(&device->alloc, pAllocator, template); + vk_object_base_finish(&template->base); + vk_free2(&device->vk.alloc, pAllocator, template); } void anv_UpdateDescriptorSetWithTemplate(