Use EXT_descriptor_indexing's UPDATE_AFTER_BIND feature to support semantics required by RS 1.0 VOLATILE descriptors. We implement this by deferring all updates of desciptor sets until Submit time.
This is fine, as command buffers cannot be executed simultaneously on D3D12, so at Submit time, we know that the command buffer is not being executed on the GPU, and updating descriptors for multiple submissions is correct.
If EXT_descriptor_indexing is not available, the fallback is the older method, which matches RS 1.1 STATIC descriptor model.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no --- libs/vkd3d/command.c | 358 +++++++++++++++++++++++++++++-------- libs/vkd3d/device.c | 20 ++- libs/vkd3d/state.c | 22 +++ libs/vkd3d/vkd3d_private.h | 19 +- tests/d3d12.c | 2 +- 5 files changed, 338 insertions(+), 83 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 0532ec0..3f78d17 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1341,7 +1341,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( { pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL; - pool_desc.flags = 0; + pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; pool_desc.maxSets = 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes; @@ -1865,6 +1865,10 @@ static void d3d12_command_list_invalidate_bindings(struct d3d12_command_list *li if (!state) return;
+ /* Each pipeline state has its own set layout for UAV counters + * based on their implicit usage in the shader. + * Binding a different pipeline state means having to re-remit + * UAV counters in a new descriptor set (and layout). */ if (state->uav_counter_mask) { struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[state->vk_bind_point]; @@ -2196,6 +2200,7 @@ static ULONG STDMETHODCALLTYPE d3d12_command_list_Release(ID3D12GraphicsCommandL if (list->allocator) d3d12_command_allocator_free_command_buffer(list->allocator, list);
+ vkd3d_free(list->descriptor_updates); vkd3d_free(list);
d3d12_device_release(device); @@ -2334,6 +2339,9 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list,
list->state = NULL;
+ /* Recycle deferred descriptor update memory if possible. */ + list->descriptor_updates_count = 0; + memset(list->so_counter_buffers, 0, sizeof(list->so_counter_buffers)); memset(list->so_counter_buffer_offsets, 0, sizeof(list->so_counter_buffer_offsets));
@@ -2534,15 +2542,54 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * time in between. Thus, the contents must not be altered (overwritten * by an update command, or freed) between when the command is recorded * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, root_signature->vk_set_layout); + bindings->in_use = false;
bindings->descriptor_table_dirty_mask |= bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask |= bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask; }
+static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point) +{ + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + + if (bindings->uav_counter_descriptor_set && !bindings->uav_counter_in_use) + return; + + /* We cannot modify bound descriptor sets. We need a new descriptor set if + * we are about to update resource bindings. + * + * The Vulkan spec says: + * + * "The descriptor set contents bound by a call to + * vkCmdBindDescriptorSets may be consumed during host execution of the + * command, or during shader execution of the resulting draws, or any + * time in between. Thus, the contents must not be altered (overwritten + * by an update command, or freed) between when the command is recorded + * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. + */ + + if (list->state->uav_counter_mask) + { + bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, + list->state->vk_set_layout); + } + else + bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; + + bindings->uav_counter_in_use = false; +} + static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_descriptor_write, VkDescriptorImageInfo *vk_image_info, const struct d3d12_desc *descriptor, uint32_t descriptor_range_magic, VkDescriptorSet vk_descriptor_set, @@ -2611,26 +2658,61 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des return true; }
-static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +static void d3d12_command_list_defer_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, uint64_t table_mask, bool uav) { + const struct d3d12_desc *base_descriptor; + unsigned i; + struct d3d12_deferred_descriptor_set_update *update; struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; - const struct d3d12_root_signature *root_signature = bindings->root_signature; + + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + { + if (table_mask & ((uint64_t)1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + { + vkd3d_array_reserve((void **)&list->descriptor_updates, &list->descriptor_updates_size, + list->descriptor_updates_count + 1, sizeof(*list->descriptor_updates)); + update = &list->descriptor_updates[list->descriptor_updates_count]; + + update->base_descriptor = base_descriptor; + update->index = i; + update->root_signature = bindings->root_signature; + update->descriptor_set = uav ? bindings->uav_counter_descriptor_set : bindings->descriptor_set; + update->uav = uav; + list->descriptor_updates_count++; + } + else + WARN("Descriptor table %u is not set.\n", i); + } + } +} + +static void d3d12_command_list_resolve_descriptor_table_uav(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + const struct d3d12_root_signature *root_signature = update->root_signature; const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - struct VkDescriptorImageInfo image_infos[24], *current_image_info; const struct d3d12_root_descriptor_table *descriptor_table; const struct d3d12_root_descriptor_table_range *range; VkDevice vk_device = list->device->vk_device; - unsigned int i, j, descriptor_count; - struct d3d12_desc *descriptor; - - descriptor_table = root_signature_get_descriptor_table(root_signature, index); + unsigned int i, j; + unsigned int uav_counter_count; + const struct d3d12_desc *descriptor; + VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkDescriptorSet vk_descriptor_set; + const struct d3d12_desc *base_descriptor = update->base_descriptor;
+ descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); descriptor = base_descriptor; - descriptor_count = 0; - current_descriptor_write = descriptor_writes; - current_image_info = image_infos; + + vk_descriptor_set = update->descriptor_set; + if (!vk_descriptor_set) + return; + + /* FIXME: There should be a smarter way than scanning through all the descriptor table ranges for this. */ for (i = 0; i < descriptor_table->range_count; ++i) { range = &descriptor_table->ranges[i]; @@ -2644,20 +2726,74 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list { unsigned int register_idx = range->base_register_idx + j;
- /* Track UAV counters. */ + /* Fish out UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV - && register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views)) + && register_idx < ARRAY_SIZE(vk_uav_counter_views)) { VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV - ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; - if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view) - bindings->uav_counter_dirty_mask |= 1u << register_idx; - bindings->vk_uav_counter_views[register_idx] = vk_counter_view; + ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; + vk_uav_counter_views[register_idx] = vk_counter_view; } + } + } + + uav_counter_count = vkd3d_popcount(list->state->uav_counter_mask); + assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes)); + + for (i = 0; i < uav_counter_count; ++i) + { + const struct vkd3d_shader_uav_counter_binding *uav_counter = &list->state->uav_counters[i]; + assert(vk_uav_counter_views[uav_counter->register_index]); + + vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; + vk_descriptor_writes[i].pNext = NULL; + vk_descriptor_writes[i].dstSet = vk_descriptor_set; + vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; + vk_descriptor_writes[i].dstArrayElement = 0; + vk_descriptor_writes[i].descriptorCount = 1; + vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; + vk_descriptor_writes[i].pImageInfo = NULL; + vk_descriptor_writes[i].pBufferInfo = NULL; + vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; + } + + VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL)); +} + +static void d3d12_command_list_resolve_descriptor_table_normal(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + const struct d3d12_root_descriptor_table *descriptor_table; + struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; + struct VkDescriptorImageInfo image_infos[24], *current_image_info; + const struct d3d12_root_signature *root_signature = update->root_signature; + const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; + const struct d3d12_root_descriptor_table_range *range; + VkDevice vk_device = list->device->vk_device; + unsigned int i, j, descriptor_count; + const struct d3d12_desc *descriptor; + const struct d3d12_desc *base_descriptor = update->base_descriptor; + + descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); + + descriptor = update->base_descriptor; + descriptor_count = 0; + current_descriptor_write = descriptor_writes; + current_image_info = image_infos; + for (i = 0; i < descriptor_table->range_count; ++i) + { + range = &descriptor_table->ranges[i];
+ if (range->offset != D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND) + { + descriptor = base_descriptor + range->offset; + } + + for (j = 0; j < range->descriptor_count; ++j, ++descriptor) + { if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write, - current_image_info, descriptor, range->descriptor_magic, - bindings->descriptor_set, range->binding, j)) + current_image_info, descriptor, range->descriptor_magic, + update->descriptor_set, range->binding, j)) continue;
++descriptor_count; @@ -2677,6 +2813,52 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list VK_CALL(vkUpdateDescriptorSets(vk_device, descriptor_count, descriptor_writes, 0, NULL)); }
+static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + if (update->uav) + d3d12_command_list_resolve_descriptor_table_uav(list, update); + else + d3d12_command_list_resolve_descriptor_table_normal(list, update); +} + +static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) +{ + unsigned i; + for (i = 0; i < list->descriptor_updates_count; i++) + d3d12_command_list_resolve_descriptor_table(list, &list->descriptor_updates[i]); +} + +static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature; + + update.descriptor_set = bindings->descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = false; + d3d12_command_list_resolve_descriptor_table_normal(list, &update); +} + +static void d3d12_command_list_update_uav_counter_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature; + + update.descriptor_set = bindings->uav_counter_descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = true; + d3d12_command_list_resolve_descriptor_table_uav(list, &update); +} + static bool vk_write_descriptor_set_from_root_descriptor(VkWriteDescriptorSet *vk_descriptor_write, const struct d3d12_root_parameter *root_parameter, VkDescriptorSet vk_descriptor_set, VkBufferView *vk_buffer_view, const VkDescriptorBufferInfo *vk_buffer_info) @@ -2781,55 +2963,6 @@ done: vkd3d_free(buffer_infos); }
-static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point) -{ - VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; - struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - const struct d3d12_pipeline_state *state = list->state; - VkDevice vk_device = list->device->vk_device; - VkDescriptorSet vk_descriptor_set; - unsigned int uav_counter_count; - unsigned int i; - - if (!state || !(state->uav_counter_mask & bindings->uav_counter_dirty_mask)) - return; - - uav_counter_count = vkd3d_popcount(state->uav_counter_mask); - assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes)); - - vk_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, state->vk_set_layout); - if (!vk_descriptor_set) - return; - - for (i = 0; i < uav_counter_count; ++i) - { - const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i]; - const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views; - - assert(vk_uav_counter_views[uav_counter->register_index]); - - vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; - vk_descriptor_writes[i].pNext = NULL; - vk_descriptor_writes[i].dstSet = vk_descriptor_set; - vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; - vk_descriptor_writes[i].dstArrayElement = 0; - vk_descriptor_writes[i].descriptorCount = 1; - vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - vk_descriptor_writes[i].pImageInfo = NULL; - vk_descriptor_writes[i].pBufferInfo = NULL; - vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; - } - - VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL)); - - VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, - state->vk_pipeline_layout, state->set_index, 1, &vk_descriptor_set, 0, NULL)); - - bindings->uav_counter_dirty_mask = 0; -} - static void d3d12_command_list_update_descriptors(struct d3d12_command_list *list, VkPipelineBindPoint bind_point) { @@ -2842,31 +2975,86 @@ static void d3d12_command_list_update_descriptors(struct d3d12_command_list *lis if (!rs || !rs->vk_set_layout) return;
+ if ((bindings->descriptor_table_active_mask | bindings->push_descriptor_dirty_mask | + (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)) == 0) + { + /* Nothing is dirty, so just return early. */ + return; + } + if (bindings->descriptor_table_dirty_mask || bindings->push_descriptor_dirty_mask) d3d12_command_list_prepare_descriptors(list, bind_point); + if (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask) + d3d12_command_list_prepare_uav_counter_descriptors(list, bind_point);
- for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + if (list->device->vk_info.EXT_descriptor_indexing) + { + d3d12_command_list_defer_update_descriptor_table(list, bind_point, bindings->descriptor_table_dirty_mask, + false); + } + else { - if (bindings->descriptor_table_dirty_mask & ((uint64_t)1 << i)) + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) { - if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) - d3d12_command_list_update_descriptor_table(list, bind_point, i, base_descriptor); - else - WARN("Descriptor table %u is not set.\n", i); + if (bindings->descriptor_table_dirty_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + d3d12_command_list_update_descriptor_table(list, bind_point, i, base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } } } bindings->descriptor_table_dirty_mask = 0;
+ /* Need to go through all descriptor tables here in the root signature, + * not just descriptor_table_dirty_mask. Binding a different shader may not invalidate descriptor tables, + * but it may invalidate the UAV counter set. */ + if (bindings->uav_counter_dirty_mask) + { + if (list->device->vk_info.EXT_descriptor_indexing) + { + d3d12_command_list_defer_update_descriptor_table(list, bind_point, bindings->descriptor_table_active_mask, + true); + } + else + { + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); i++) + { + if (bindings->descriptor_table_active_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + d3d12_command_list_update_uav_counter_descriptor_table(list, bind_point, i, base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } + } + } + } + bindings->uav_counter_dirty_mask = 0; + d3d12_command_list_update_push_descriptors(list, bind_point);
- if (bindings->descriptor_set) + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->descriptor_set && bindings->descriptor_set != bindings->descriptor_set_bound) { VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, rs->vk_pipeline_layout, rs->main_set, 1, &bindings->descriptor_set, 0, NULL)); bindings->in_use = true; + bindings->descriptor_set_bound = bindings->descriptor_set; }
- d3d12_command_list_update_uav_counter_descriptors(list, bind_point); + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->uav_counter_descriptor_set && + bindings->uav_counter_descriptor_set != bindings->uav_counter_descriptor_set_bound) + { + VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, + list->state->vk_pipeline_layout, list->state->set_index, 1, &bindings->uav_counter_descriptor_set, 0, NULL)); + bindings->uav_counter_in_use = true; + bindings->uav_counter_descriptor_set_bound = bindings->uav_counter_descriptor_set; + } }
static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list) @@ -4032,6 +4220,10 @@ static void d3d12_command_list_set_root_signature(struct d3d12_command_list *lis bindings->descriptor_set = VK_NULL_HANDLE; bindings->descriptor_table_dirty_mask = bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask = bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask; + + bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; + bindings->descriptor_set_bound = VK_NULL_HANDLE; + bindings->uav_counter_descriptor_set_bound = VK_NULL_HANDLE; }
static void STDMETHODCALLTYPE d3d12_command_list_SetComputeRootSignature(ID3D12GraphicsCommandList1 *iface, @@ -5405,6 +5597,9 @@ static HRESULT d3d12_command_list_init(struct d3d12_command_list *list, struct d d3d12_device_add_ref(list->device = device);
list->allocator = allocator; + list->descriptor_updates = NULL; + list->descriptor_updates_count = 0; + list->descriptor_updates_size = 0;
if (SUCCEEDED(hr = d3d12_command_allocator_allocate_command_buffer(allocator, list))) { @@ -5642,6 +5837,13 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm return; }
+ /* Descriptors in root signature in 1.0 are VOLATILE by default, so + * the descriptor heap only need to be valid right before we submit them to the GPU. + * If we have EXT_descriptor_indexing enabled with UpdateAfterBind, we update + * descriptor sets here rather than while we're recording the command buffer. + * For each submission of the command buffer, we can modify the descriptor heap as we please. */ + d3d12_command_list_resolve_descriptor_tables(cmd_list); + buffers[i] = cmd_list->vk_command_buffer; }
diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index 13ebc70..e159c1f 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -1436,11 +1436,25 @@ static HRESULT vkd3d_init_device_caps(struct d3d12_device *device, }
if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing - && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind - || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingSampledImageUpdateAfterBind + && descriptor_indexing->descriptorBindingStorageImageUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind) + { + TRACE("Enabling VK_EXT_descriptor_indexing for volatile descriptor updates.\n"); + } + else + { + WARN("VK_EXT_descriptor indexing not supported in sufficient capacity. Volatile descriptor updates will not work.\n"); + vulkan_info->EXT_descriptor_indexing = false; + } + + if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing + && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind || descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind + || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind || descriptor_indexing->descriptorBindingStorageTexelBufferUpdateAfterBind) - && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) + && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) { WARN("Disabling robust buffer access for the update after bind feature.\n"); features->robustBufferAccess = VK_FALSE; diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index a321fa4..50cff1f 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -745,21 +745,43 @@ static HRESULT vkd3d_create_descriptor_set_layout(struct d3d12_device *device, VkDescriptorSetLayoutCreateFlags flags, unsigned int binding_count, const VkDescriptorSetLayoutBinding *bindings, VkDescriptorSetLayout *set_layout) { + unsigned int i; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkDescriptorSetLayoutCreateInfo set_desc; VkResult vr; + VkDescriptorSetLayoutBindingFlagsCreateInfoEXT flags_info; + VkDescriptorBindingFlagsEXT *binding_flags = NULL;
set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO; set_desc.pNext = NULL; set_desc.flags = flags; set_desc.bindingCount = binding_count; set_desc.pBindings = bindings; + + if (!(flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR) && device->vk_info.EXT_descriptor_indexing) + { + set_desc.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT; + flags_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT; + flags_info.pNext = NULL; + flags_info.bindingCount = binding_count; + binding_flags = vkd3d_malloc(sizeof(*binding_flags) * binding_count); + for (i = 0; i < binding_count; i++) + { + binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT | + VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT; + } + flags_info.pBindingFlags = binding_flags; + set_desc.pNext = &flags_info; + } + if ((vr = VK_CALL(vkCreateDescriptorSetLayout(device->vk_device, &set_desc, NULL, set_layout))) < 0) { WARN("Failed to create Vulkan descriptor set layout, vr %d.\n", vr); + vkd3d_free(binding_flags); return hresult_from_vk_result(vr); }
+ vkd3d_free(binding_flags); return S_OK; }
diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2d62fda..58e9aa9 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -888,13 +888,16 @@ struct vkd3d_pipeline_bindings const struct d3d12_root_signature *root_signature;
VkDescriptorSet descriptor_set; + VkDescriptorSet uav_counter_descriptor_set; + VkDescriptorSet descriptor_set_bound; + VkDescriptorSet uav_counter_descriptor_set_bound; bool in_use; + bool uav_counter_in_use;
D3D12_GPU_DESCRIPTOR_HANDLE descriptor_tables[D3D12_MAX_ROOT_COST]; uint64_t descriptor_table_dirty_mask; uint64_t descriptor_table_active_mask;
- VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; uint8_t uav_counter_dirty_mask;
/* Needed when VK_KHR_push_descriptor is not available. */ @@ -903,6 +906,16 @@ struct vkd3d_pipeline_bindings uint32_t push_descriptor_active_mask; };
+struct d3d12_deferred_descriptor_set_update +{ + const struct d3d12_desc *base_descriptor; + unsigned int index; + + const struct d3d12_root_signature *root_signature; + VkDescriptorSet descriptor_set; + bool uav; +}; + /* ID3D12CommandList */ struct d3d12_command_list { @@ -946,6 +959,10 @@ struct d3d12_command_list VkBuffer so_counter_buffers[D3D12_SO_BUFFER_SLOT_COUNT]; VkDeviceSize so_counter_buffer_offsets[D3D12_SO_BUFFER_SLOT_COUNT];
+ struct d3d12_deferred_descriptor_set_update *descriptor_updates; + size_t descriptor_updates_size; + size_t descriptor_updates_count; + struct vkd3d_private_store private_store; };
diff --git a/tests/d3d12.c b/tests/d3d12.c index 858cf99..16a048c 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -15799,7 +15799,7 @@ static void test_update_descriptor_heap_after_closing_command_list(void) D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE); get_texture_readback_with_command_list(context.render_target, 0, &rb, queue, command_list); value = get_readback_uint(&rb, 0, 0, 0); - todo ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); + ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); release_resource_readback(&rb);
ID3D12DescriptorHeap_Release(cpu_heap);
The resource index is not found in idx[0], but idx[1] when using SM 5.1, and register space information depends on resource type.
Pass down space/index information directly to resource declaration functions instead rather than relying on this information being in reg.idx[0].
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no --- include/vkd3d_shader.h | 3 ++ libs/vkd3d-shader/dxbc.c | 31 ++++++++++++ libs/vkd3d-shader/spirv.c | 63 ++++++++++++++---------- libs/vkd3d-shader/vkd3d_shader_private.h | 5 ++ libs/vkd3d/state.c | 24 +++++---- libs/vkd3d/vkd3d_private.h | 1 + 6 files changed, 87 insertions(+), 40 deletions(-)
diff --git a/include/vkd3d_shader.h b/include/vkd3d_shader.h index 6b4d3f5..8f43d66 100644 --- a/include/vkd3d_shader.h +++ b/include/vkd3d_shader.h @@ -138,6 +138,7 @@ struct vkd3d_shader_parameter struct vkd3d_shader_resource_binding { enum vkd3d_shader_descriptor_type type; + unsigned int register_space; unsigned int register_index; enum vkd3d_shader_visibility shader_visibility; unsigned int flags; /* vkd3d_shader_binding_flags */ @@ -159,6 +160,7 @@ struct vkd3d_shader_combined_resource_sampler
struct vkd3d_shader_uav_counter_binding { + unsigned int register_space; unsigned int register_index; /* u# */ enum vkd3d_shader_visibility shader_visibility;
@@ -168,6 +170,7 @@ struct vkd3d_shader_uav_counter_binding
struct vkd3d_shader_push_constant_buffer { + unsigned int register_space; unsigned int register_index; enum vkd3d_shader_visibility shader_visibility;
diff --git a/libs/vkd3d-shader/dxbc.c b/libs/vkd3d-shader/dxbc.c index 98c51e4..f9f6614 100644 --- a/libs/vkd3d-shader/dxbc.c +++ b/libs/vkd3d-shader/dxbc.c @@ -624,6 +624,10 @@ static void shader_sm4_read_dcl_resource(struct vkd3d_shader_instruction *ins, ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT;
shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.semantic.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.semantic.register_index = ins->declaration.semantic.reg.reg.idx[1].offset; + else + ins->declaration.semantic.register_index = ins->declaration.semantic.reg.reg.idx[0].offset; }
static void shader_sm4_read_dcl_constant_buffer(struct vkd3d_shader_instruction *ins, @@ -647,9 +651,14 @@ static void shader_sm4_read_dcl_constant_buffer(struct vkd3d_shader_instruction return; }
+ /* Apparently both idx[1] and idx[2] contain the range for indexed resources, + * but only support plain CBV here and ignore idx[2]. */ + ins->declaration.cb.register_index = ins->declaration.cb.src.reg.idx[1].offset; ins->declaration.cb.size = *tokens++; shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.cb.register_space); } + else + ins->declaration.cb.register_index = ins->declaration.cb.src.reg.idx[0].offset; }
static void shader_sm4_read_dcl_sampler(struct vkd3d_shader_instruction *ins, @@ -663,6 +672,10 @@ static void shader_sm4_read_dcl_sampler(struct vkd3d_shader_instruction *ins, FIXME("Unhandled sampler mode %#x.\n", ins->flags); shader_sm4_read_src_param(priv, &tokens, end, VKD3D_DATA_SAMPLER, &ins->declaration.sampler.src); shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.sampler.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.sampler.register_index = ins->declaration.sampler.src.reg.idx[1].offset; + else + ins->declaration.sampler.register_index = ins->declaration.sampler.src.reg.idx[0].offset; }
static void shader_sm4_read_dcl_index_range(struct vkd3d_shader_instruction *ins, @@ -863,6 +876,10 @@ static void shader_sm5_read_dcl_uav_raw(struct vkd3d_shader_instruction *ins, shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_UAV, &ins->declaration.raw_resource.dst); ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT; shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.raw_resource.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.raw_resource.register_index = ins->declaration.raw_resource.dst.reg.idx[1].offset; + else + ins->declaration.raw_resource.register_index = ins->declaration.raw_resource.dst.reg.idx[0].offset; }
static void shader_sm5_read_dcl_uav_structured(struct vkd3d_shader_instruction *ins, @@ -874,9 +891,14 @@ static void shader_sm5_read_dcl_uav_structured(struct vkd3d_shader_instruction * shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_UAV, &ins->declaration.structured_resource.reg); ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT; ins->declaration.structured_resource.byte_stride = *tokens; + tokens++; if (ins->declaration.structured_resource.byte_stride % 4) FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride); shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.structured_resource.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[1].offset; + else + ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[0].offset; }
static void shader_sm5_read_dcl_tgsm_raw(struct vkd3d_shader_instruction *ins, @@ -909,9 +931,14 @@ static void shader_sm5_read_dcl_resource_structured(struct vkd3d_shader_instruct
shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_RESOURCE, &ins->declaration.structured_resource.reg); ins->declaration.structured_resource.byte_stride = *tokens; + tokens++; if (ins->declaration.structured_resource.byte_stride % 4) FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride); shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.structured_resource.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[1].offset; + else + ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[0].offset; }
static void shader_sm5_read_dcl_resource_raw(struct vkd3d_shader_instruction *ins, @@ -922,6 +949,10 @@ static void shader_sm5_read_dcl_resource_raw(struct vkd3d_shader_instruction *in
shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_RESOURCE, &ins->declaration.dst); shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.raw_resource.register_space); + if (shader_is_sm_5_1(priv)) + ins->declaration.raw_resource.register_index = ins->declaration.dst.reg.idx[1].offset; + else + ins->declaration.raw_resource.register_index = ins->declaration.dst.reg.idx[0].offset; }
static void shader_sm5_read_sync(struct vkd3d_shader_instruction *ins, diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index a949e4a..aaae1b6 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -2052,6 +2052,7 @@ struct vkd3d_hull_shader_variables
struct vkd3d_dxbc_compiler { + struct vkd3d_shader_version shader_version; struct vkd3d_spirv_builder spirv_builder;
uint32_t options; @@ -2107,6 +2108,11 @@ struct vkd3d_dxbc_compiler size_t spec_constants_size; };
+static bool shader_is_sm_5_1(const struct vkd3d_dxbc_compiler *compiler) +{ + return (compiler->shader_version.major * 100 + compiler->shader_version.minor) >= 501; +} + static bool is_control_point_phase(const struct vkd3d_shader_phase *phase) { return phase && phase->type == VKD3DSIH_HS_CONTROL_POINT_PHASE; @@ -2131,6 +2137,8 @@ struct vkd3d_dxbc_compiler *vkd3d_dxbc_compiler_create(const struct vkd3d_shader
memset(compiler, 0, sizeof(*compiler));
+ compiler->shader_version = *shader_version; + max_element_count = max(output_signature->element_count, patch_constant_signature->element_count); if (!(compiler->output_info = vkd3d_calloc(max_element_count, sizeof(*compiler->output_info)))) { @@ -2227,9 +2235,10 @@ static bool vkd3d_dxbc_compiler_check_shader_visibility(const struct vkd3d_dxbc_ }
static struct vkd3d_push_constant_buffer_binding *vkd3d_dxbc_compiler_find_push_constant_buffer( - const struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_register *reg) + const struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_constant_buffer *cb) { - unsigned int reg_idx = reg->idx[0].offset; + unsigned int reg_idx = cb->register_index; + unsigned int reg_space = cb->register_space; unsigned int i;
for (i = 0; i < compiler->shader_interface.push_constant_buffer_count; ++i) @@ -2239,7 +2248,7 @@ static struct vkd3d_push_constant_buffer_binding *vkd3d_dxbc_compiler_find_push_ if (!vkd3d_dxbc_compiler_check_shader_visibility(compiler, current->pc.shader_visibility)) continue;
- if (current->pc.register_index == reg_idx) + if (current->pc.register_index == reg_idx && current->pc.register_space == reg_space) return current; }
@@ -2276,13 +2285,13 @@ static bool vkd3d_dxbc_compiler_has_combined_sampler(const struct vkd3d_dxbc_com
static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor_binding( struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_register *reg, + unsigned int reg_idx, unsigned int reg_space, enum vkd3d_shader_resource_type resource_type, bool is_uav_counter) { const struct vkd3d_shader_interface_info *shader_interface = &compiler->shader_interface; enum vkd3d_shader_descriptor_type descriptor_type; enum vkd3d_shader_binding_flag resource_type_flag; struct vkd3d_shader_descriptor_binding binding; - unsigned int reg_idx = reg->idx[0].offset; unsigned int i;
descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_UNKNOWN; @@ -2313,7 +2322,7 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor if (current->offset) FIXME("Atomic counter offsets are not supported yet.\n");
- if (current->register_index == reg_idx) + if (current->register_index == reg_idx && current->register_space == reg_space) return current->binding; } if (shader_interface->uav_counter_count) @@ -2331,7 +2340,7 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor if (!vkd3d_dxbc_compiler_check_shader_visibility(compiler, current->shader_visibility)) continue;
- if (current->type == descriptor_type && current->register_index == reg_idx) + if (current->type == descriptor_type && current->register_index == reg_idx && current->register_space == reg_space) return current->binding; } if (shader_interface->binding_count) @@ -2354,12 +2363,12 @@ static void vkd3d_dxbc_compiler_emit_descriptor_binding(struct vkd3d_dxbc_compil }
static void vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(struct vkd3d_dxbc_compiler *compiler, - uint32_t variable_id, const struct vkd3d_shader_register *reg, + uint32_t variable_id, const struct vkd3d_shader_register *reg, unsigned int reg_idx, unsigned int reg_space, enum vkd3d_shader_resource_type resource_type, bool is_uav_counter) { struct vkd3d_shader_descriptor_binding binding;
- binding = vkd3d_dxbc_compiler_get_descriptor_binding(compiler, reg, resource_type, is_uav_counter); + binding = vkd3d_dxbc_compiler_get_descriptor_binding(compiler, reg, reg_idx, reg_space, resource_type, is_uav_counter); vkd3d_dxbc_compiler_emit_descriptor_binding(compiler, variable_id, &binding); }
@@ -2825,7 +2834,8 @@ static void vkd3d_dxbc_compiler_emit_dereference_register(struct vkd3d_dxbc_comp { assert(!reg->idx[0].rel_addr); indexes[index_count++] = vkd3d_dxbc_compiler_get_constant_uint(compiler, register_info->member_idx); - indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler, ®->idx[1]); + indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler, + ®->idx[shader_is_sm_5_1(compiler) ? 2 : 1]); } else if (reg->type == VKD3DSPR_IMMCONSTBUFFER) { @@ -2835,6 +2845,11 @@ static void vkd3d_dxbc_compiler_emit_dereference_register(struct vkd3d_dxbc_comp { indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler, ®->idx[1]); } + else if (reg->type == VKD3DSPR_SAMPLER) + { + /* SM 5.1 will have an index here referring to something which we throw away. */ + index_count = 0; + } else if (register_info->is_aggregate) { struct vkd3d_shader_register_index reg_idx = reg->idx[0]; @@ -4957,10 +4972,7 @@ static void vkd3d_dxbc_compiler_emit_dcl_constant_buffer(struct vkd3d_dxbc_compi
assert(!(instruction->flags & ~VKD3DSI_INDEXED_DYNAMIC));
- if (cb->register_space) - FIXME("Unhandled register space %u.\n", cb->register_space); - - if ((push_cb = vkd3d_dxbc_compiler_find_push_constant_buffer(compiler, reg))) + if ((push_cb = vkd3d_dxbc_compiler_find_push_constant_buffer(compiler, cb))) { /* Push constant buffers are handled in * vkd3d_dxbc_compiler_emit_push_constant_buffers(). @@ -4990,7 +5002,7 @@ static void vkd3d_dxbc_compiler_emit_dcl_constant_buffer(struct vkd3d_dxbc_compi pointer_type_id, storage_class, 0);
vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(compiler, - var_id, reg, VKD3D_SHADER_RESOURCE_BUFFER, false); + var_id, reg, cb->register_index, cb->register_space, VKD3D_SHADER_RESOURCE_BUFFER, false);
vkd3d_dxbc_compiler_emit_register_debug_name(builder, var_id, reg);
@@ -5042,9 +5054,6 @@ static void vkd3d_dxbc_compiler_emit_dcl_sampler(struct vkd3d_dxbc_compiler *com uint32_t type_id, ptr_type_id, var_id; struct vkd3d_symbol reg_symbol;
- if (instruction->declaration.sampler.register_space) - FIXME("Unhandled register space %u.\n", instruction->declaration.sampler.register_space); - if (vkd3d_dxbc_compiler_has_combined_sampler(compiler, NULL, reg)) return;
@@ -5054,7 +5063,9 @@ static void vkd3d_dxbc_compiler_emit_dcl_sampler(struct vkd3d_dxbc_compiler *com ptr_type_id, storage_class, 0);
vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(compiler, - var_id, reg, VKD3D_SHADER_RESOURCE_NONE, false); + var_id, reg, + instruction->declaration.sampler.register_index, instruction->declaration.sampler.register_space, + VKD3D_SHADER_RESOURCE_NONE, false);
vkd3d_dxbc_compiler_emit_register_debug_name(builder, var_id, reg);
@@ -5185,7 +5196,8 @@ static void vkd3d_dxbc_compiler_emit_combined_sampler_declarations(struct vkd3d_ }
static void vkd3d_dxbc_compiler_emit_resource_declaration(struct vkd3d_dxbc_compiler *compiler, - const struct vkd3d_shader_register *reg, enum vkd3d_shader_resource_type resource_type, + const struct vkd3d_shader_register *reg, unsigned int reg_idx, unsigned reg_space, + enum vkd3d_shader_resource_type resource_type, enum vkd3d_data_type resource_data_type, unsigned int structure_stride, bool raw) { uint32_t counter_type_id, type_id, ptr_type_id, var_id, counter_var_id = 0; @@ -5220,7 +5232,7 @@ static void vkd3d_dxbc_compiler_emit_resource_declaration(struct vkd3d_dxbc_comp var_id = vkd3d_spirv_build_op_variable(builder, &builder->global_stream, ptr_type_id, storage_class, 0);
- vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(compiler, var_id, reg, resource_type, false); + vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(compiler, var_id, reg, reg_idx, reg_space, resource_type, false);
vkd3d_dxbc_compiler_emit_register_debug_name(builder, var_id, reg);
@@ -5243,7 +5255,7 @@ static void vkd3d_dxbc_compiler_emit_resource_declaration(struct vkd3d_dxbc_comp ptr_type_id, storage_class, 0);
vkd3d_dxbc_compiler_emit_descriptor_binding_for_reg(compiler, - counter_var_id, reg, resource_type, true); + counter_var_id, reg, reg_idx, reg_space, resource_type, true);
vkd3d_spirv_build_op_name(builder, counter_var_id, "u%u_counter", reg->idx[0].offset); } @@ -5264,12 +5276,11 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource(struct vkd3d_dxbc_compiler *co { const struct vkd3d_shader_semantic *semantic = &instruction->declaration.semantic;
- if (semantic->register_space) - FIXME("Unhandled register space %u.\n", semantic->register_space); if (instruction->flags) FIXME("Unhandled UAV flags %#x.\n", instruction->flags);
vkd3d_dxbc_compiler_emit_resource_declaration(compiler, &semantic->reg.reg, + semantic->register_index, semantic->register_space, semantic->resource_type, semantic->resource_data_type, 0, false); }
@@ -5278,12 +5289,11 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource_raw(struct vkd3d_dxbc_compiler { const struct vkd3d_shader_raw_resource *resource = &instruction->declaration.raw_resource;
- if (resource->register_space) - FIXME("Unhandled register space %u.\n", resource->register_space); if (instruction->flags) FIXME("Unhandled UAV flags %#x.\n", instruction->flags);
vkd3d_dxbc_compiler_emit_resource_declaration(compiler, &resource->dst.reg, + resource->register_index, resource->register_space, VKD3D_SHADER_RESOURCE_BUFFER, VKD3D_DATA_UINT, 0, true); }
@@ -5294,12 +5304,11 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource_structured(struct vkd3d_dxbc_c const struct vkd3d_shader_register *reg = &resource->reg.reg; unsigned int stride = resource->byte_stride;
- if (resource->register_space) - FIXME("Unhandled register space %u.\n", resource->register_space); if (instruction->flags) FIXME("Unhandled UAV flags %#x.\n", instruction->flags);
vkd3d_dxbc_compiler_emit_resource_declaration(compiler, reg, + resource->register_index, resource->register_space, VKD3D_SHADER_RESOURCE_BUFFER, VKD3D_DATA_UINT, stride / 4, false); }
diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h index 940cb76..1c052a0 100644 --- a/libs/vkd3d-shader/vkd3d_shader_private.h +++ b/libs/vkd3d-shader/vkd3d_shader_private.h @@ -615,6 +615,7 @@ struct vkd3d_shader_semantic enum vkd3d_data_type resource_data_type; struct vkd3d_shader_dst_param reg; unsigned int register_space; + unsigned int register_index; };
enum vkd3d_shader_input_sysval_semantic @@ -662,6 +663,7 @@ struct vkd3d_shader_register_semantic struct vkd3d_shader_sampler { struct vkd3d_shader_src_param src; + unsigned int register_index; unsigned int register_space; };
@@ -669,6 +671,7 @@ struct vkd3d_shader_constant_buffer { struct vkd3d_shader_src_param src; unsigned int size; + unsigned int register_index; unsigned int register_space; };
@@ -676,12 +679,14 @@ struct vkd3d_shader_structured_resource { struct vkd3d_shader_dst_param reg; unsigned int byte_stride; + unsigned int register_index; unsigned int register_space; };
struct vkd3d_shader_raw_resource { struct vkd3d_shader_dst_param dst; + unsigned int register_index; unsigned int register_space; };
diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index 50cff1f..ed20de7 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -308,12 +308,6 @@ static bool vk_binding_from_d3d12_descriptor_range(struct VkDescriptorSetLayoutB = vk_descriptor_type_from_d3d12_range_type(descriptor_range->RangeType, is_buffer); binding_desc->descriptorCount = 1;
- if (descriptor_range->RegisterSpace) - { - FIXME("Unhandled register space %u.\n", descriptor_range->RegisterSpace); - return false; - } - binding_desc->stageFlags = stage_flags_from_visibility(shader_visibility); binding_desc->pImmutableSamplers = NULL;
@@ -509,6 +503,7 @@ static HRESULT d3d12_root_signature_init_push_constants(struct d3d12_root_signat ? push_constants[0].stageFlags : stage_flags_from_visibility(p->ShaderVisibility); root_constant->offset = offset;
+ root_signature->root_constants[j].register_space = p->u.Constants.RegisterSpace; root_signature->root_constants[j].register_index = p->u.Constants.ShaderRegister; root_signature->root_constants[j].shader_visibility = vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility); @@ -532,7 +527,7 @@ struct vkd3d_descriptor_set_context };
static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature *root_signature, - enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_idx, + enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_space, unsigned int register_idx, bool buffer_descriptor, enum vkd3d_shader_visibility shader_visibility, struct vkd3d_descriptor_set_context *context) { @@ -540,6 +535,7 @@ static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature * = &root_signature->descriptor_mapping[context->descriptor_index++];
mapping->type = descriptor_type; + mapping->register_space = register_space; mapping->register_index = register_idx; mapping->shader_visibility = shader_visibility; mapping->flags = buffer_descriptor ? VKD3D_SHADER_BINDING_FLAG_BUFFER : VKD3D_SHADER_BINDING_FLAG_IMAGE; @@ -548,7 +544,7 @@ static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature * }
static uint32_t d3d12_root_signature_assign_vk_bindings(struct d3d12_root_signature *root_signature, - enum vkd3d_shader_descriptor_type descriptor_type, unsigned int base_register_idx, + enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_space, unsigned int base_register_idx, unsigned int binding_count, bool is_buffer_descriptor, bool duplicate_descriptors, enum vkd3d_shader_visibility shader_visibility, struct vkd3d_descriptor_set_context *context) { @@ -565,10 +561,10 @@ static uint32_t d3d12_root_signature_assign_vk_bindings(struct d3d12_root_signat { if (duplicate_descriptors) d3d12_root_signature_append_vk_binding(root_signature, descriptor_type, - base_register_idx + i, true, shader_visibility, context); + register_space, base_register_idx + i, true, shader_visibility, context);
d3d12_root_signature_append_vk_binding(root_signature, descriptor_type, - base_register_idx + i, is_buffer_descriptor, shader_visibility, context); + register_space, base_register_idx + i, is_buffer_descriptor, shader_visibility, context); } return first_binding; } @@ -624,7 +620,7 @@ static HRESULT d3d12_root_signature_init_root_descriptor_tables(struct d3d12_roo
vk_binding = d3d12_root_signature_assign_vk_bindings(root_signature, vkd3d_descriptor_type_from_d3d12_range_type(range->RangeType), - range->BaseShaderRegister, range->NumDescriptors, false, true, + range->RegisterSpace, range->BaseShaderRegister, range->NumDescriptors, false, true, vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility), context);
/* Unroll descriptor range. */ @@ -657,6 +653,7 @@ static HRESULT d3d12_root_signature_init_root_descriptor_tables(struct d3d12_roo table->ranges[j].binding = vk_binding; table->ranges[j].descriptor_magic = vkd3d_descriptor_magic_from_d3d12(range->RangeType); table->ranges[j].base_register_idx = range->BaseShaderRegister; + table->ranges[j].register_space = range->RegisterSpace; } }
@@ -690,7 +687,7 @@ static HRESULT d3d12_root_signature_init_root_descriptors(struct d3d12_root_sign
cur_binding->binding = d3d12_root_signature_assign_vk_bindings(root_signature, vkd3d_descriptor_type_from_d3d12_root_parameter_type(p->ParameterType), - p->u.Descriptor.ShaderRegister, 1, true, false, + p->u.Descriptor.RegisterSpace, p->u.Descriptor.ShaderRegister, 1, true, false, vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility), context); cur_binding->descriptorType = vk_descriptor_type_from_d3d12_root_parameter(p->ParameterType); cur_binding->descriptorCount = 1; @@ -727,7 +724,7 @@ static HRESULT d3d12_root_signature_init_static_samplers(struct d3d12_root_signa return hr;
cur_binding->binding = d3d12_root_signature_assign_vk_bindings(root_signature, - VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER, s->ShaderRegister, 1, false, false, + VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER, s->RegisterSpace, s->ShaderRegister, 1, false, false, vkd3d_shader_visibility_from_d3d12(s->ShaderVisibility), context); cur_binding->descriptorType = VK_DESCRIPTOR_TYPE_SAMPLER; cur_binding->descriptorCount = 1; @@ -1441,6 +1438,7 @@ static HRESULT d3d12_pipeline_state_init_compute_uav_counters(struct d3d12_pipel if (!(shader_info->uav_counter_mask & (1u << i))) continue;
+ state->uav_counters[j].register_space = 0; state->uav_counters[j].register_index = i; state->uav_counters[j].shader_visibility = VKD3D_SHADER_VISIBILITY_COMPUTE; state->uav_counters[j].binding.set = context.set_index; diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 58e9aa9..bd54c67 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -634,6 +634,7 @@ struct d3d12_root_descriptor_table_range
uint32_t descriptor_magic; unsigned int base_register_idx; + unsigned int register_space; };
struct d3d12_root_descriptor_table
On Fri, 1 Nov 2019 at 17:08, Hans-Kristian Arntzen post@arntzen-software.no wrote:
The resource index is not found in idx[0], but idx[1] when using SM 5.1, and register space information depends on resource type.
Pass down space/index information directly to resource declaration functions instead rather than relying on this information being in reg.idx[0].
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
include/vkd3d_shader.h | 3 ++ libs/vkd3d-shader/dxbc.c | 31 ++++++++++++ libs/vkd3d-shader/spirv.c | 63 ++++++++++++++---------- libs/vkd3d-shader/vkd3d_shader_private.h | 5 ++ libs/vkd3d/state.c | 24 +++++---- libs/vkd3d/vkd3d_private.h | 1 + 6 files changed, 87 insertions(+), 40 deletions(-)
Hi Hans-Kristian,
This seems to make sense at first sight, but do you also have some tests to go along with these patches?
As an aside, could you please set the git format-patch subject prefix (format.subjectprefix) for vkd3d patches to "PATCH vkd3d", so that these can be properly distinguished from Wine patches?
Henri
Currently, descriptor pools are allocated with the intention of being used as a suballocator, but for a bindless style, root signatures can contain many thousand descriptors.
Fall back to single allocations for these cases and do not attempt to recycle these descriptor sets.
This path is triggered by Control DX12 which creates a RS with > ~200k SRVs in register space 1.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no --- libs/vkd3d/command.c | 67 ++++++++++++++++++++++++++++++-------- libs/vkd3d/state.c | 19 ++--------- libs/vkd3d/vkd3d_private.h | 23 +++++++++++++ 3 files changed, 78 insertions(+), 31 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 3f78d17..6f448d5 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1264,13 +1264,24 @@ static bool d3d12_command_allocator_add_framebuffer(struct d3d12_command_allocat }
static bool d3d12_command_allocator_add_descriptor_pool(struct d3d12_command_allocator *allocator, - VkDescriptorPool pool) + VkDescriptorPool pool, bool recycle) { - if (!vkd3d_array_reserve((void **)&allocator->descriptor_pools, &allocator->descriptor_pools_size, + if (recycle) + { + if (!vkd3d_array_reserve((void **)&allocator->descriptor_pools, &allocator->descriptor_pools_size, allocator->descriptor_pool_count + 1, sizeof(*allocator->descriptor_pools))) - return false; + return false;
- allocator->descriptor_pools[allocator->descriptor_pool_count++] = pool; + allocator->descriptor_pools[allocator->descriptor_pool_count++] = pool; + } + else + { + if (!vkd3d_array_reserve((void **)&allocator->descriptor_pools_huge, &allocator->descriptor_pools_huge_size, + allocator->descriptor_pool_huge_count + 1, sizeof(*allocator->descriptor_pools_huge))) + return false; + + allocator->descriptor_pools_huge[allocator->descriptor_pool_huge_count++] = pool; + }
return true; } @@ -1313,9 +1324,10 @@ static bool d3d12_command_allocator_add_transfer_buffer(struct d3d12_command_all }
static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( - struct d3d12_command_allocator *allocator) + struct d3d12_command_allocator *allocator, const struct d3d12_root_signature_info *root_signature) { - static const VkDescriptorPoolSize pool_sizes[] = + bool huge_root_signature = false; + VkDescriptorPoolSize pool_sizes[] = { {VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, 1024}, {VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, 1024}, @@ -1324,6 +1336,22 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( {VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, 1024}, {VK_DESCRIPTOR_TYPE_SAMPLER, 1024}, }; + + if (root_signature && ( + root_signature->cbv_count > 1024 || + root_signature->srv_count > 1024 || + root_signature->uav_count + root_signature->buffer_uav_count > 1024 || + root_signature->sampler_count > 1024)) + { + huge_root_signature = true; + pool_sizes[0].descriptorCount = root_signature->cbv_count; + pool_sizes[1].descriptorCount = root_signature->srv_count; + pool_sizes[2].descriptorCount = root_signature->srv_count; + pool_sizes[3].descriptorCount = root_signature->uav_count + root_signature->buffer_uav_count; + pool_sizes[4].descriptorCount = root_signature->uav_count; + pool_sizes[5].descriptorCount = root_signature->sampler_count; + } + struct d3d12_device *device = allocator->device; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; struct VkDescriptorPoolCreateInfo pool_desc; @@ -1331,7 +1359,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( VkDescriptorPool vk_pool; VkResult vr;
- if (allocator->free_descriptor_pool_count > 0) + if (!huge_root_signature && allocator->free_descriptor_pool_count > 0) { vk_pool = allocator->free_descriptor_pools[allocator->free_descriptor_pool_count - 1]; allocator->free_descriptor_pools[allocator->free_descriptor_pool_count - 1] = VK_NULL_HANDLE; @@ -1342,7 +1370,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL; pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; - pool_desc.maxSets = 512; + pool_desc.maxSets = huge_root_signature ? 1 : 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes; if ((vr = VK_CALL(vkCreateDescriptorPool(vk_device, &pool_desc, NULL, &vk_pool))) < 0) @@ -1352,7 +1380,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( } }
- if (!(d3d12_command_allocator_add_descriptor_pool(allocator, vk_pool))) + if (!(d3d12_command_allocator_add_descriptor_pool(allocator, vk_pool, !huge_root_signature))) { ERR("Failed to add descriptor pool.\n"); VK_CALL(vkDestroyDescriptorPool(vk_device, vk_pool, NULL)); @@ -1363,7 +1391,8 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( }
static VkDescriptorSet d3d12_command_allocator_allocate_descriptor_set( - struct d3d12_command_allocator *allocator, VkDescriptorSetLayout vk_set_layout) + struct d3d12_command_allocator *allocator, VkDescriptorSetLayout vk_set_layout, + const struct d3d12_root_signature_info *root_signature) { struct d3d12_device *device = allocator->device; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; @@ -1373,7 +1402,7 @@ static VkDescriptorSet d3d12_command_allocator_allocate_descriptor_set( VkResult vr;
if (!allocator->vk_descriptor_pool) - allocator->vk_descriptor_pool = d3d12_command_allocator_allocate_descriptor_pool(allocator); + allocator->vk_descriptor_pool = d3d12_command_allocator_allocate_descriptor_pool(allocator, root_signature); if (!allocator->vk_descriptor_pool) return VK_NULL_HANDLE;
@@ -1387,7 +1416,7 @@ static VkDescriptorSet d3d12_command_allocator_allocate_descriptor_set(
allocator->vk_descriptor_pool = VK_NULL_HANDLE; if (vr == VK_ERROR_FRAGMENTED_POOL || vr == VK_ERROR_OUT_OF_POOL_MEMORY_KHR) - allocator->vk_descriptor_pool = d3d12_command_allocator_allocate_descriptor_pool(allocator); + allocator->vk_descriptor_pool = d3d12_command_allocator_allocate_descriptor_pool(allocator, root_signature); if (!allocator->vk_descriptor_pool) { ERR("Failed to allocate descriptor set, vr %d.\n", vr); @@ -1478,6 +1507,12 @@ static void d3d12_command_allocator_free_resources(struct d3d12_command_allocato } allocator->descriptor_pool_count = 0;
+ for (i = 0; i < allocator->descriptor_pool_huge_count; ++i) + { + VK_CALL(vkDestroyDescriptorPool(device->vk_device, allocator->descriptor_pools_huge[i], NULL)); + } + allocator->descriptor_pool_huge_count = 0; + for (i = 0; i < allocator->framebuffer_count; ++i) { VK_CALL(vkDestroyFramebuffer(device->vk_device, allocator->framebuffers[i], NULL)); @@ -1755,6 +1790,10 @@ static HRESULT d3d12_command_allocator_init(struct d3d12_command_allocator *allo allocator->descriptor_pools_size = 0; allocator->descriptor_pool_count = 0;
+ allocator->descriptor_pools_huge = NULL; + allocator->descriptor_pools_huge_size = 0; + allocator->descriptor_pool_huge_count = 0; + allocator->views = NULL; allocator->views_size = 0; allocator->view_count = 0; @@ -2547,7 +2586,7 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, - root_signature->vk_set_layout); + root_signature->vk_set_layout, &root_signature->info);
bindings->in_use = false;
@@ -2582,7 +2621,7 @@ static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_comm if (list->state->uav_counter_mask) { bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, - list->state->vk_set_layout); + list->state->vk_set_layout, NULL); } else bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index ed20de7..eb87350 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -314,23 +314,6 @@ static bool vk_binding_from_d3d12_descriptor_range(struct VkDescriptorSetLayoutB return true; }
-struct d3d12_root_signature_info -{ - size_t cbv_count; - size_t buffer_uav_count; - size_t uav_count; - size_t buffer_srv_count; - size_t srv_count; - size_t sampler_count; - - size_t descriptor_count; - - size_t root_constant_count; - size_t root_descriptor_count; - - size_t cost; -}; - static HRESULT d3d12_root_signature_info_count_descriptors(struct d3d12_root_signature_info *info, const D3D12_DESCRIPTOR_RANGE *range) { @@ -919,6 +902,8 @@ static HRESULT d3d12_root_signature_init(struct d3d12_root_signature *root_signa if (FAILED(hr = vkd3d_private_store_init(&root_signature->private_store))) goto fail;
+ root_signature->info = info; + d3d12_device_add_ref(root_signature->device = device);
return S_OK; diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index bd54c67..3a263ef 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -666,6 +666,23 @@ struct d3d12_root_parameter };
/* ID3D12RootSignature */ +struct d3d12_root_signature_info +{ + size_t cbv_count; + size_t buffer_uav_count; + size_t uav_count; + size_t buffer_srv_count; + size_t srv_count; + size_t sampler_count; + + size_t descriptor_count; + + size_t root_constant_count; + size_t root_descriptor_count; + + size_t cost; +}; + struct d3d12_root_signature { ID3D12RootSignature ID3D12RootSignature_iface; @@ -699,6 +716,8 @@ struct d3d12_root_signature unsigned int static_sampler_count; VkSampler *static_samplers;
+ struct d3d12_root_signature_info info; + struct d3d12_device *device;
struct vkd3d_private_store private_store; @@ -846,6 +865,10 @@ struct d3d12_command_allocator size_t descriptor_pools_size; size_t descriptor_pool_count;
+ VkDescriptorPool *descriptor_pools_huge; + size_t descriptor_pools_huge_size; + size_t descriptor_pool_huge_count; + struct vkd3d_view **views; size_t views_size; size_t view_count;
I am not sure what tree you are working on, but applying this patch to GIT@4576236199769eacadc6de49e03d05df999f9181 leads to a lot of:
libs/vkd3d/command.c: In function ‘d3d12_command_list_resolve_descriptor_table_normal’: libs/vkd3d/command.c:2815:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_table’ 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ libs/vkd3d/command.c:2815:1: warning: ISO C90 forbids mixed declarations and code [-Wdeclaration-after-statement] 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~ libs/vkd3d/command.c:2824:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_tables’ 2824 | static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and subsequent errors.
Does this patch depend on some other patch(set) to compile?
From recent commits to vkd3d, it seems as some patches are picked, and some not depending on some criteria. Working on a massively changed git tree submitting patches that does not include its dependencies is kinda hard for me to test :)
If you have a git repo with a working tree i am willing to test stuff.
Sorry if you feel i am stepping on some toes here tho...
Sveinar
On 01.11.2019 14:36, Hans-Kristian Arntzen wrote:
Use EXT_descriptor_indexing's UPDATE_AFTER_BIND feature to support semantics required by RS 1.0 VOLATILE descriptors. We implement this by deferring all updates of desciptor sets until Submit time.
This is fine, as command buffers cannot be executed simultaneously on D3D12, so at Submit time, we know that the command buffer is not being executed on the GPU, and updating descriptors for multiple submissions is correct.
If EXT_descriptor_indexing is not available, the fallback is the older method, which matches RS 1.1 STATIC descriptor model.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
libs/vkd3d/command.c | 358 +++++++++++++++++++++++++++++-------- libs/vkd3d/device.c | 20 ++- libs/vkd3d/state.c | 22 +++ libs/vkd3d/vkd3d_private.h | 19 +- tests/d3d12.c | 2 +- 5 files changed, 338 insertions(+), 83 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 0532ec0..3f78d17 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1341,7 +1341,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( { pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL;
pool_desc.flags = 0;
pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; pool_desc.maxSets = 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes;@@ -1865,6 +1865,10 @@ static void d3d12_command_list_invalidate_bindings(struct d3d12_command_list *li if (!state) return;
- /* Each pipeline state has its own set layout for UAV counters
* based on their implicit usage in the shader.* Binding a different pipeline state means having to re-remit* UAV counters in a new descriptor set (and layout). */ if (state->uav_counter_mask) { struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[state->vk_bind_point];@@ -2196,6 +2200,7 @@ static ULONG STDMETHODCALLTYPE d3d12_command_list_Release(ID3D12GraphicsCommandL if (list->allocator) d3d12_command_allocator_free_command_buffer(list->allocator, list);
vkd3d_free(list->descriptor_updates); vkd3d_free(list); d3d12_device_release(device);@@ -2334,6 +2339,9 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list,
list->state = NULL;
- /* Recycle deferred descriptor update memory if possible. */
- list->descriptor_updates_count = 0;
memset(list->so_counter_buffers, 0, sizeof(list->so_counter_buffers)); memset(list->so_counter_buffer_offsets, 0, sizeof(list->so_counter_buffer_offsets));@@ -2534,15 +2542,54 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * time in between. Thus, the contents must not be altered (overwritten * by an update command, or freed) between when the command is recorded * and when the command completes executing on the queue."
** Even if we have descriptor indexing and UPDATE_AFTER_BIND,* we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, root_signature->vk_set_layout);bindings->in_use = false; bindings->descriptor_table_dirty_mask |= bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask |= bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask;}
+static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point)+{
- struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
- if (bindings->uav_counter_descriptor_set && !bindings->uav_counter_in_use)
return;- /* We cannot modify bound descriptor sets. We need a new descriptor set if
* we are about to update resource bindings.** The Vulkan spec says:** "The descriptor set contents bound by a call to* vkCmdBindDescriptorSets may be consumed during host execution of the* command, or during shader execution of the resulting draws, or any* time in between. Thus, the contents must not be altered (overwritten* by an update command, or freed) between when the command is recorded* and when the command completes executing on the queue."** Even if we have descriptor indexing and UPDATE_AFTER_BIND,* we need at the very least a new descriptor set.*/- if (list->state->uav_counter_mask)
- {
bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator,list->state->vk_set_layout);- }
- else
bindings->uav_counter_descriptor_set = VK_NULL_HANDLE;- bindings->uav_counter_in_use = false;
+}
- static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_descriptor_write, VkDescriptorImageInfo *vk_image_info, const struct d3d12_desc *descriptor, uint32_t descriptor_range_magic, VkDescriptorSet vk_descriptor_set,
@@ -2611,26 +2658,61 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des return true; }
-static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor)+static void d3d12_command_list_defer_update_descriptor_table(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point, uint64_t table_mask, bool uav)- const struct d3d12_desc *base_descriptor;
- unsigned i;
- struct d3d12_deferred_descriptor_set_update *update; struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
- struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write;
- const struct d3d12_root_signature *root_signature = bindings->root_signature;
- for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i)
- {
if (table_mask & ((uint64_t)1 << i)){if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))){vkd3d_array_reserve((void **)&list->descriptor_updates, &list->descriptor_updates_size,list->descriptor_updates_count + 1, sizeof(*list->descriptor_updates));update = &list->descriptor_updates[list->descriptor_updates_count];update->base_descriptor = base_descriptor;update->index = i;update->root_signature = bindings->root_signature;update->descriptor_set = uav ? bindings->uav_counter_descriptor_set : bindings->descriptor_set;update->uav = uav;list->descriptor_updates_count++;}elseWARN("Descriptor table %u is not set.\n", i);}- }
+}
+static void d3d12_command_list_resolve_descriptor_table_uav(struct d3d12_command_list *list,
const struct d3d12_deferred_descriptor_set_update *update)+{
- const struct d3d12_root_signature *root_signature = update->root_signature; const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs;
- struct VkDescriptorImageInfo image_infos[24], *current_image_info; const struct d3d12_root_descriptor_table *descriptor_table; const struct d3d12_root_descriptor_table_range *range; VkDevice vk_device = list->device->vk_device;
- unsigned int i, j, descriptor_count;
- struct d3d12_desc *descriptor;
- descriptor_table = root_signature_get_descriptor_table(root_signature, index);
unsigned int i, j;
unsigned int uav_counter_count;
const struct d3d12_desc *descriptor;
VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS];
VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS];
VkDescriptorSet vk_descriptor_set;
const struct d3d12_desc *base_descriptor = update->base_descriptor;
descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); descriptor = base_descriptor;
- descriptor_count = 0;
- current_descriptor_write = descriptor_writes;
- current_image_info = image_infos;
- vk_descriptor_set = update->descriptor_set;
- if (!vk_descriptor_set)
return;- /* FIXME: There should be a smarter way than scanning through all the descriptor table ranges for this. */ for (i = 0; i < descriptor_table->range_count; ++i) { range = &descriptor_table->ranges[i];
@@ -2644,20 +2726,74 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list { unsigned int register_idx = range->base_register_idx + j;
/* Track UAV counters. */
/* Fish out UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV
&& register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views))
&& register_idx < ARRAY_SIZE(vk_uav_counter_views)) { VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV
? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE;if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view)bindings->uav_counter_dirty_mask |= 1u << register_idx;bindings->vk_uav_counter_views[register_idx] = vk_counter_view;
? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE;vk_uav_counter_views[register_idx] = vk_counter_view; }}- }
- uav_counter_count = vkd3d_popcount(list->state->uav_counter_mask);
- assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
- for (i = 0; i < uav_counter_count; ++i)
- {
const struct vkd3d_shader_uav_counter_binding *uav_counter = &list->state->uav_counters[i];assert(vk_uav_counter_views[uav_counter->register_index]);vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;vk_descriptor_writes[i].pNext = NULL;vk_descriptor_writes[i].dstSet = vk_descriptor_set;vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding;vk_descriptor_writes[i].dstArrayElement = 0;vk_descriptor_writes[i].descriptorCount = 1;vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;vk_descriptor_writes[i].pImageInfo = NULL;vk_descriptor_writes[i].pBufferInfo = NULL;vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index];- }
- VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));
+}
+static void d3d12_command_list_resolve_descriptor_table_normal(struct d3d12_command_list *list,
const struct d3d12_deferred_descriptor_set_update *update)+{
const struct d3d12_root_descriptor_table *descriptor_table;
struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write;
struct VkDescriptorImageInfo image_infos[24], *current_image_info;
const struct d3d12_root_signature *root_signature = update->root_signature;
const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs;
const struct d3d12_root_descriptor_table_range *range;
VkDevice vk_device = list->device->vk_device;
unsigned int i, j, descriptor_count;
const struct d3d12_desc *descriptor;
const struct d3d12_desc *base_descriptor = update->base_descriptor;
descriptor_table = root_signature_get_descriptor_table(root_signature, update->index);
descriptor = update->base_descriptor;
descriptor_count = 0;
current_descriptor_write = descriptor_writes;
current_image_info = image_infos;
for (i = 0; i < descriptor_table->range_count; ++i)
{
range = &descriptor_table->ranges[i];if (range->offset != D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND){descriptor = base_descriptor + range->offset;}for (j = 0; j < range->descriptor_count; ++j, ++descriptor){ if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write,
current_image_info, descriptor, range->descriptor_magic,bindings->descriptor_set, range->binding, j))
current_image_info, descriptor, range->descriptor_magic,update->descriptor_set, range->binding, j)) continue; ++descriptor_count;@@ -2677,6 +2813,52 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list VK_CALL(vkUpdateDescriptorSets(vk_device, descriptor_count, descriptor_writes, 0, NULL)); }
+static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list,
const struct d3d12_deferred_descriptor_set_update *update)+{
- if (update->uav)
d3d12_command_list_resolve_descriptor_table_uav(list, update);- else
d3d12_command_list_resolve_descriptor_table_normal(list, update);+}
+static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) +{
- unsigned i;
- for (i = 0; i < list->descriptor_updates_count; i++)
d3d12_command_list_resolve_descriptor_table(list, &list->descriptor_updates[i]);+}
+static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor)+{
- struct d3d12_deferred_descriptor_set_update update;
- struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
- const struct d3d12_root_signature *root_signature = bindings->root_signature;
- update.descriptor_set = bindings->descriptor_set;
- update.index = index;
- update.root_signature = root_signature;
- update.base_descriptor = base_descriptor;
- update.uav = false;
- d3d12_command_list_resolve_descriptor_table_normal(list, &update);
+}
+static void d3d12_command_list_update_uav_counter_descriptor_table(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor)+{
- struct d3d12_deferred_descriptor_set_update update;
- struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
- const struct d3d12_root_signature *root_signature = bindings->root_signature;
- update.descriptor_set = bindings->uav_counter_descriptor_set;
- update.index = index;
- update.root_signature = root_signature;
- update.base_descriptor = base_descriptor;
- update.uav = true;
- d3d12_command_list_resolve_descriptor_table_uav(list, &update);
+}
- static bool vk_write_descriptor_set_from_root_descriptor(VkWriteDescriptorSet *vk_descriptor_write, const struct d3d12_root_parameter *root_parameter, VkDescriptorSet vk_descriptor_set, VkBufferView *vk_buffer_view, const VkDescriptorBufferInfo *vk_buffer_info)
@@ -2781,55 +2963,6 @@ done: vkd3d_free(buffer_infos); }
-static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_command_list *list,
VkPipelineBindPoint bind_point)-{
- VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS];
- struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
- const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs;
- const struct d3d12_pipeline_state *state = list->state;
- VkDevice vk_device = list->device->vk_device;
- VkDescriptorSet vk_descriptor_set;
- unsigned int uav_counter_count;
- unsigned int i;
- if (!state || !(state->uav_counter_mask & bindings->uav_counter_dirty_mask))
return;- uav_counter_count = vkd3d_popcount(state->uav_counter_mask);
- assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
- vk_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, state->vk_set_layout);
- if (!vk_descriptor_set)
return;- for (i = 0; i < uav_counter_count; ++i)
- {
const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i];const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views;assert(vk_uav_counter_views[uav_counter->register_index]);vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;vk_descriptor_writes[i].pNext = NULL;vk_descriptor_writes[i].dstSet = vk_descriptor_set;vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding;vk_descriptor_writes[i].dstArrayElement = 0;vk_descriptor_writes[i].descriptorCount = 1;vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;vk_descriptor_writes[i].pImageInfo = NULL;vk_descriptor_writes[i].pBufferInfo = NULL;vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index];- }
- VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
state->vk_pipeline_layout, state->set_index, 1, &vk_descriptor_set, 0, NULL));- bindings->uav_counter_dirty_mask = 0;
-}
- static void d3d12_command_list_update_descriptors(struct d3d12_command_list *list, VkPipelineBindPoint bind_point) {
@@ -2842,31 +2975,86 @@ static void d3d12_command_list_update_descriptors(struct d3d12_command_list *lis if (!rs || !rs->vk_set_layout) return;
- if ((bindings->descriptor_table_active_mask | bindings->push_descriptor_dirty_mask |
(list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)) == 0)- {
/* Nothing is dirty, so just return early. */return;- }
if (bindings->descriptor_table_dirty_mask || bindings->push_descriptor_dirty_mask) d3d12_command_list_prepare_descriptors(list, bind_point);- if (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)
d3d12_command_list_prepare_uav_counter_descriptors(list, bind_point);
- for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i)
- if (list->device->vk_info.EXT_descriptor_indexing)
- {
d3d12_command_list_defer_update_descriptor_table(list, bind_point, bindings->descriptor_table_dirty_mask,false);- }
- else {
if (bindings->descriptor_table_dirty_mask & ((uint64_t)1 << i))
/* FIXME: FOR_EACH_BIT */for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) {
if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))d3d12_command_list_update_descriptor_table(list, bind_point, i, base_descriptor);elseWARN("Descriptor table %u is not set.\n", i);
if (bindings->descriptor_table_dirty_mask & ((uint64_t) 1 << i)){if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))d3d12_command_list_update_descriptor_table(list, bind_point, i, base_descriptor);elseWARN("Descriptor table %u is not set.\n", i);} } } bindings->descriptor_table_dirty_mask = 0;/* Need to go through all descriptor tables here in the root signature,
* not just descriptor_table_dirty_mask. Binding a different shader may not invalidate descriptor tables,* but it may invalidate the UAV counter set. */if (bindings->uav_counter_dirty_mask)
{
if (list->device->vk_info.EXT_descriptor_indexing){d3d12_command_list_defer_update_descriptor_table(list, bind_point, bindings->descriptor_table_active_mask,true);}else{/* FIXME: FOR_EACH_BIT */for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); i++){if (bindings->descriptor_table_active_mask & ((uint64_t) 1 << i)){if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))d3d12_command_list_update_uav_counter_descriptor_table(list, bind_point, i, base_descriptor);elseWARN("Descriptor table %u is not set.\n", i);}}}}
bindings->uav_counter_dirty_mask = 0;
d3d12_command_list_update_push_descriptors(list, bind_point);
- if (bindings->descriptor_set)
- /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */
- if (bindings->descriptor_set && bindings->descriptor_set != bindings->descriptor_set_bound) { VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, rs->vk_pipeline_layout, rs->main_set, 1, &bindings->descriptor_set, 0, NULL)); bindings->in_use = true;
bindings->descriptor_set_bound = bindings->descriptor_set; }
- d3d12_command_list_update_uav_counter_descriptors(list, bind_point);
/* Don't rebind the same descriptor set as long as we're in same pipeline layout. */
if (bindings->uav_counter_descriptor_set &&
bindings->uav_counter_descriptor_set != bindings->uav_counter_descriptor_set_bound){
VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,list->state->vk_pipeline_layout, list->state->set_index, 1, &bindings->uav_counter_descriptor_set, 0, NULL));bindings->uav_counter_in_use = true;bindings->uav_counter_descriptor_set_bound = bindings->uav_counter_descriptor_set;} }
static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list)
@@ -4032,6 +4220,10 @@ static void d3d12_command_list_set_root_signature(struct d3d12_command_list *lis bindings->descriptor_set = VK_NULL_HANDLE; bindings->descriptor_table_dirty_mask = bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask = bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask;
bindings->uav_counter_descriptor_set = VK_NULL_HANDLE;
bindings->descriptor_set_bound = VK_NULL_HANDLE;
bindings->uav_counter_descriptor_set_bound = VK_NULL_HANDLE; }
static void STDMETHODCALLTYPE d3d12_command_list_SetComputeRootSignature(ID3D12GraphicsCommandList1 *iface,
@@ -5405,6 +5597,9 @@ static HRESULT d3d12_command_list_init(struct d3d12_command_list *list, struct d d3d12_device_add_ref(list->device = device);
list->allocator = allocator;
list->descriptor_updates = NULL;
list->descriptor_updates_count = 0;
list->descriptor_updates_size = 0;
if (SUCCEEDED(hr = d3d12_command_allocator_allocate_command_buffer(allocator, list))) {
@@ -5642,6 +5837,13 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm return; }
/* Descriptors in root signature in 1.0 are VOLATILE by default, so* the descriptor heap only need to be valid right before we submit them to the GPU.* If we have EXT_descriptor_indexing enabled with UpdateAfterBind, we update* descriptor sets here rather than while we're recording the command buffer.* For each submission of the command buffer, we can modify the descriptor heap as we please. */d3d12_command_list_resolve_descriptor_tables(cmd_list);buffers[i] = cmd_list->vk_command_buffer; }diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index 13ebc70..e159c1f 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -1436,11 +1436,25 @@ static HRESULT vkd3d_init_device_caps(struct d3d12_device *device, }
if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing
&& (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind|| descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind
&& descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind&& descriptor_indexing->descriptorBindingSampledImageUpdateAfterBind&& descriptor_indexing->descriptorBindingStorageImageUpdateAfterBind&& descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind)- {
TRACE("Enabling VK_EXT_descriptor_indexing for volatile descriptor updates.\n");- }
- else
- {
WARN("VK_EXT_descriptor indexing not supported in sufficient capacity. Volatile descriptor updates will not work.\n");vulkan_info->EXT_descriptor_indexing = false;- }
- if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing
&& (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind || descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind|| descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind || descriptor_indexing->descriptorBindingStorageTexelBufferUpdateAfterBind)
&& !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind)
&& !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) { WARN("Disabling robust buffer access for the update after bind feature.\n"); features->robustBufferAccess = VK_FALSE;diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index a321fa4..50cff1f 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -745,21 +745,43 @@ static HRESULT vkd3d_create_descriptor_set_layout(struct d3d12_device *device, VkDescriptorSetLayoutCreateFlags flags, unsigned int binding_count, const VkDescriptorSetLayoutBinding *bindings, VkDescriptorSetLayout *set_layout) {
unsigned int i; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkDescriptorSetLayoutCreateInfo set_desc; VkResult vr;
VkDescriptorSetLayoutBindingFlagsCreateInfoEXT flags_info;
VkDescriptorBindingFlagsEXT *binding_flags = NULL;
set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO; set_desc.pNext = NULL; set_desc.flags = flags; set_desc.bindingCount = binding_count; set_desc.pBindings = bindings;
if (!(flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR) && device->vk_info.EXT_descriptor_indexing)
{
set_desc.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT;flags_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT;flags_info.pNext = NULL;flags_info.bindingCount = binding_count;binding_flags = vkd3d_malloc(sizeof(*binding_flags) * binding_count);for (i = 0; i < binding_count; i++){binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT |VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT;}flags_info.pBindingFlags = binding_flags;set_desc.pNext = &flags_info;}
if ((vr = VK_CALL(vkCreateDescriptorSetLayout(device->vk_device, &set_desc, NULL, set_layout))) < 0) { WARN("Failed to create Vulkan descriptor set layout, vr %d.\n", vr);vkd3d_free(binding_flags); return hresult_from_vk_result(vr); }vkd3d_free(binding_flags); return S_OK; }
diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2d62fda..58e9aa9 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -888,13 +888,16 @@ struct vkd3d_pipeline_bindings const struct d3d12_root_signature *root_signature;
VkDescriptorSet descriptor_set;
VkDescriptorSet uav_counter_descriptor_set;
VkDescriptorSet descriptor_set_bound;
VkDescriptorSet uav_counter_descriptor_set_bound; bool in_use;
bool uav_counter_in_use;
D3D12_GPU_DESCRIPTOR_HANDLE descriptor_tables[D3D12_MAX_ROOT_COST]; uint64_t descriptor_table_dirty_mask; uint64_t descriptor_table_active_mask;
VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; uint8_t uav_counter_dirty_mask;
/* Needed when VK_KHR_push_descriptor is not available. */
@@ -903,6 +906,16 @@ struct vkd3d_pipeline_bindings uint32_t push_descriptor_active_mask; };
+struct d3d12_deferred_descriptor_set_update +{
- const struct d3d12_desc *base_descriptor;
- unsigned int index;
- const struct d3d12_root_signature *root_signature;
- VkDescriptorSet descriptor_set;
- bool uav;
+};
- /* ID3D12CommandList */ struct d3d12_command_list {
@@ -946,6 +959,10 @@ struct d3d12_command_list VkBuffer so_counter_buffers[D3D12_SO_BUFFER_SLOT_COUNT]; VkDeviceSize so_counter_buffer_offsets[D3D12_SO_BUFFER_SLOT_COUNT];
- struct d3d12_deferred_descriptor_set_update *descriptor_updates;
- size_t descriptor_updates_size;
- size_t descriptor_updates_count;
struct vkd3d_private_store private_store;diff --git a/tests/d3d12.c b/tests/d3d12.c index 858cf99..16a048c 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -15799,7 +15799,7 @@ static void test_update_descriptor_heap_after_closing_command_list(void) D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE); get_texture_readback_with_command_list(context.render_target, 0, &rb, queue, command_list); value = get_readback_uint(&rb, 0, 0, 0);
- todo ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value);
ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); release_resource_readback(&rb);
ID3D12DescriptorHeap_Release(cpu_heap);
The patch was made based on that parent commit and compiles fine for me.
Cheers, Hans-Kristian
On 11/1/19 9:27 PM, Sveinar Søpler wrote:
I am not sure what tree you are working on, but applying this patch to GIT@4576236199769eacadc6de49e03d05df999f9181 leads to a lot of:
libs/vkd3d/command.c: In function ‘d3d12_command_list_resolve_descriptor_table_normal’: libs/vkd3d/command.c:2815:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_table’ 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ libs/vkd3d/command.c:2815:1: warning: ISO C90 forbids mixed declarations and code [-Wdeclaration-after-statement] 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~ libs/vkd3d/command.c:2824:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_tables’ 2824 | static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and subsequent errors.
Does this patch depend on some other patch(set) to compile?
From recent commits to vkd3d, it seems as some patches are picked, and some not depending on some criteria. Working on a massively changed git tree submitting patches that does not include its dependencies is kinda hard for me to test :)
If you have a git repo with a working tree i am willing to test stuff.
Sorry if you feel i am stepping on some toes here tho...
Sveinar
On 01.11.2019 14:36, Hans-Kristian Arntzen wrote:
Use EXT_descriptor_indexing's UPDATE_AFTER_BIND feature to support semantics required by RS 1.0 VOLATILE descriptors. We implement this by deferring all updates of desciptor sets until Submit time.
This is fine, as command buffers cannot be executed simultaneously on D3D12, so at Submit time, we know that the command buffer is not being executed on the GPU, and updating descriptors for multiple submissions is correct.
If EXT_descriptor_indexing is not available, the fallback is the older method, which matches RS 1.1 STATIC descriptor model.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
libs/vkd3d/command.c | 358 +++++++++++++++++++++++++++++-------- libs/vkd3d/device.c | 20 ++- libs/vkd3d/state.c | 22 +++ libs/vkd3d/vkd3d_private.h | 19 +- tests/d3d12.c | 2 +- 5 files changed, 338 insertions(+), 83 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 0532ec0..3f78d17 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1341,7 +1341,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( { pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL; - pool_desc.flags = 0; + pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; pool_desc.maxSets = 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes; @@ -1865,6 +1865,10 @@ static void d3d12_command_list_invalidate_bindings(struct d3d12_command_list *li if (!state) return; + /* Each pipeline state has its own set layout for UAV counters + * based on their implicit usage in the shader. + * Binding a different pipeline state means having to re-remit + * UAV counters in a new descriptor set (and layout). */ if (state->uav_counter_mask) { struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[state->vk_bind_point]; @@ -2196,6 +2200,7 @@ static ULONG STDMETHODCALLTYPE d3d12_command_list_Release(ID3D12GraphicsCommandL if (list->allocator) d3d12_command_allocator_free_command_buffer(list->allocator, list); + vkd3d_free(list->descriptor_updates); vkd3d_free(list); d3d12_device_release(device); @@ -2334,6 +2339,9 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list, list->state = NULL; + /* Recycle deferred descriptor update memory if possible. */ + list->descriptor_updates_count = 0;
memset(list->so_counter_buffers, 0, sizeof(list->so_counter_buffers)); memset(list->so_counter_buffer_offsets, 0, sizeof(list->so_counter_buffer_offsets)); @@ -2534,15 +2542,54 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * time in between. Thus, the contents must not be altered (overwritten * by an update command, or freed) between when the command is recorded * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, root_signature->vk_set_layout);
bindings->in_use = false; bindings->descriptor_table_dirty_mask |= bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask |= bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask; } +static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point) +{ + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
+ if (bindings->uav_counter_descriptor_set && !bindings->uav_counter_in_use) + return;
+ /* We cannot modify bound descriptor sets. We need a new descriptor set if + * we are about to update resource bindings. + * + * The Vulkan spec says: + * + * "The descriptor set contents bound by a call to + * vkCmdBindDescriptorSets may be consumed during host execution of the + * command, or during shader execution of the resulting draws, or any + * time in between. Thus, the contents must not be altered (overwritten + * by an update command, or freed) between when the command is recorded + * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. + */
+ if (list->state->uav_counter_mask) + { + bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, + list->state->vk_set_layout); + } + else + bindings->uav_counter_descriptor_set = VK_NULL_HANDLE;
+ bindings->uav_counter_in_use = false; +}
static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_descriptor_write, VkDescriptorImageInfo *vk_image_info, const struct d3d12_desc *descriptor, uint32_t descriptor_range_magic, VkDescriptorSet vk_descriptor_set, @@ -2611,26 +2658,61 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des return true; } -static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +static void d3d12_command_list_defer_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, uint64_t table_mask, bool uav) { + const struct d3d12_desc *base_descriptor; + unsigned i; + struct d3d12_deferred_descriptor_set_update *update; struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; - const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + { + if (table_mask & ((uint64_t)1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + { + vkd3d_array_reserve((void **)&list->descriptor_updates, &list->descriptor_updates_size,
- list->descriptor_updates_count + 1,
sizeof(*list->descriptor_updates)); + update = &list->descriptor_updates[list->descriptor_updates_count];
+ update->base_descriptor = base_descriptor; + update->index = i; + update->root_signature = bindings->root_signature; + update->descriptor_set = uav ? bindings->uav_counter_descriptor_set : bindings->descriptor_set; + update->uav = uav; + list->descriptor_updates_count++; + } + else + WARN("Descriptor table %u is not set.\n", i); + } + } +}
+static void d3d12_command_list_resolve_descriptor_table_uav(struct d3d12_command_list *list,
- const struct d3d12_deferred_descriptor_set_update *update)
+{ + const struct d3d12_root_signature *root_signature = update->root_signature; const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - struct VkDescriptorImageInfo image_infos[24], *current_image_info; const struct d3d12_root_descriptor_table *descriptor_table; const struct d3d12_root_descriptor_table_range *range; VkDevice vk_device = list->device->vk_device; - unsigned int i, j, descriptor_count; - struct d3d12_desc *descriptor;
- descriptor_table = root_signature_get_descriptor_table(root_signature, index); + unsigned int i, j; + unsigned int uav_counter_count; + const struct d3d12_desc *descriptor; + VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkDescriptorSet vk_descriptor_set; + const struct d3d12_desc *base_descriptor = update->base_descriptor; + descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); descriptor = base_descriptor; - descriptor_count = 0; - current_descriptor_write = descriptor_writes; - current_image_info = image_infos;
+ vk_descriptor_set = update->descriptor_set; + if (!vk_descriptor_set) + return;
+ /* FIXME: There should be a smarter way than scanning through all the descriptor table ranges for this. */ for (i = 0; i < descriptor_table->range_count; ++i) { range = &descriptor_table->ranges[i]; @@ -2644,20 +2726,74 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list { unsigned int register_idx = range->base_register_idx + j; - /* Track UAV counters. */ + /* Fish out UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV - && register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views)) + && register_idx < ARRAY_SIZE(vk_uav_counter_views)) { VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV - ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; - if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view) - bindings->uav_counter_dirty_mask |= 1u << register_idx; - bindings->vk_uav_counter_views[register_idx] = vk_counter_view; + ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; + vk_uav_counter_views[register_idx] = vk_counter_view; } + } + }
+ uav_counter_count = vkd3d_popcount(list->state->uav_counter_mask); + assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
+ for (i = 0; i < uav_counter_count; ++i) + { + const struct vkd3d_shader_uav_counter_binding *uav_counter = &list->state->uav_counters[i];
- assert(vk_uav_counter_views[uav_counter->register_index]);
+ vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; + vk_descriptor_writes[i].pNext = NULL; + vk_descriptor_writes[i].dstSet = vk_descriptor_set; + vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; + vk_descriptor_writes[i].dstArrayElement = 0; + vk_descriptor_writes[i].descriptorCount = 1; + vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; + vk_descriptor_writes[i].pImageInfo = NULL; + vk_descriptor_writes[i].pBufferInfo = NULL; + vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; + }
+ VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL)); +}
+static void d3d12_command_list_resolve_descriptor_table_normal(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + const struct d3d12_root_descriptor_table *descriptor_table; + struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; + struct VkDescriptorImageInfo image_infos[24], *current_image_info; + const struct d3d12_root_signature *root_signature = update->root_signature; + const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; + const struct d3d12_root_descriptor_table_range *range; + VkDevice vk_device = list->device->vk_device; + unsigned int i, j, descriptor_count; + const struct d3d12_desc *descriptor; + const struct d3d12_desc *base_descriptor = update->base_descriptor;
+ descriptor_table = root_signature_get_descriptor_table(root_signature, update->index);
+ descriptor = update->base_descriptor; + descriptor_count = 0; + current_descriptor_write = descriptor_writes; + current_image_info = image_infos; + for (i = 0; i < descriptor_table->range_count; ++i) + { + range = &descriptor_table->ranges[i]; + if (range->offset != D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND) + { + descriptor = base_descriptor + range->offset; + }
+ for (j = 0; j < range->descriptor_count; ++j, ++descriptor) + { if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write, - current_image_info, descriptor, range->descriptor_magic, - bindings->descriptor_set, range->binding, j))
- current_image_info, descriptor, range->descriptor_magic,
- update->descriptor_set, range->binding, j))
continue; ++descriptor_count; @@ -2677,6 +2813,52 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list VK_CALL(vkUpdateDescriptorSets(vk_device, descriptor_count, descriptor_writes, 0, NULL)); } +static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + if (update->uav) + d3d12_command_list_resolve_descriptor_table_uav(list, update); + else
- d3d12_command_list_resolve_descriptor_table_normal(list, update);
+}
+static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) +{ + unsigned i; + for (i = 0; i < list->descriptor_updates_count; i++) + d3d12_command_list_resolve_descriptor_table(list, &list->descriptor_updates[i]); +}
+static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = false; + d3d12_command_list_resolve_descriptor_table_normal(list, &update); +}
+static void d3d12_command_list_update_uav_counter_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->uav_counter_descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = true; + d3d12_command_list_resolve_descriptor_table_uav(list, &update); +}
static bool vk_write_descriptor_set_from_root_descriptor(VkWriteDescriptorSet *vk_descriptor_write, const struct d3d12_root_parameter *root_parameter, VkDescriptorSet vk_descriptor_set, VkBufferView *vk_buffer_view, const VkDescriptorBufferInfo *vk_buffer_info) @@ -2781,55 +2963,6 @@ done: vkd3d_free(buffer_infos); } -static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point) -{ - VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; - struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - const struct d3d12_pipeline_state *state = list->state; - VkDevice vk_device = list->device->vk_device; - VkDescriptorSet vk_descriptor_set; - unsigned int uav_counter_count; - unsigned int i;
- if (!state || !(state->uav_counter_mask & bindings->uav_counter_dirty_mask)) - return;
- uav_counter_count = vkd3d_popcount(state->uav_counter_mask); - assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
- vk_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, state->vk_set_layout); - if (!vk_descriptor_set) - return;
- for (i = 0; i < uav_counter_count; ++i) - { - const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i]; - const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views;
- assert(vk_uav_counter_views[uav_counter->register_index]);
- vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; - vk_descriptor_writes[i].pNext = NULL; - vk_descriptor_writes[i].dstSet = vk_descriptor_set; - vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; - vk_descriptor_writes[i].dstArrayElement = 0; - vk_descriptor_writes[i].descriptorCount = 1; - vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - vk_descriptor_writes[i].pImageInfo = NULL; - vk_descriptor_writes[i].pBufferInfo = NULL; - vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; - }
- VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, - state->vk_pipeline_layout, state->set_index, 1, &vk_descriptor_set, 0, NULL));
- bindings->uav_counter_dirty_mask = 0; -}
static void d3d12_command_list_update_descriptors(struct d3d12_command_list *list, VkPipelineBindPoint bind_point) { @@ -2842,31 +2975,86 @@ static void d3d12_command_list_update_descriptors(struct d3d12_command_list *lis if (!rs || !rs->vk_set_layout) return; + if ((bindings->descriptor_table_active_mask | bindings->push_descriptor_dirty_mask | + (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)) == 0) + { + /* Nothing is dirty, so just return early. */ + return; + }
if (bindings->descriptor_table_dirty_mask || bindings->push_descriptor_dirty_mask) d3d12_command_list_prepare_descriptors(list, bind_point); + if (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)
- d3d12_command_list_prepare_uav_counter_descriptors(list, bind_point);
- for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + if (list->device->vk_info.EXT_descriptor_indexing) + { + d3d12_command_list_defer_update_descriptor_table(list, bind_point, bindings->descriptor_table_dirty_mask,
- false);
+ } + else { - if (bindings->descriptor_table_dirty_mask & ((uint64_t)1 << i)) + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) { - if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); - else - WARN("Descriptor table %u is not set.\n", i); + if (bindings->descriptor_table_dirty_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } } } bindings->descriptor_table_dirty_mask = 0; + /* Need to go through all descriptor tables here in the root signature, + * not just descriptor_table_dirty_mask. Binding a different shader may not invalidate descriptor tables, + * but it may invalidate the UAV counter set. */ + if (bindings->uav_counter_dirty_mask) + { + if (list->device->vk_info.EXT_descriptor_indexing) + {
- d3d12_command_list_defer_update_descriptor_table(list, bind_point,
bindings->descriptor_table_active_mask,
- true);
+ } + else + { + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); i++) + { + if (bindings->descriptor_table_active_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_uav_counter_descriptor_table(list,
bind_point, i, base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } + } + } + } + bindings->uav_counter_dirty_mask = 0;
d3d12_command_list_update_push_descriptors(list, bind_point); - if (bindings->descriptor_set) + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->descriptor_set && bindings->descriptor_set != bindings->descriptor_set_bound) { VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, rs->vk_pipeline_layout, rs->main_set, 1, &bindings->descriptor_set, 0, NULL)); bindings->in_use = true; + bindings->descriptor_set_bound = bindings->descriptor_set; } - d3d12_command_list_update_uav_counter_descriptors(list, bind_point); + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->uav_counter_descriptor_set && + bindings->uav_counter_descriptor_set != bindings->uav_counter_descriptor_set_bound) + {
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
+ list->state->vk_pipeline_layout, list->state->set_index, 1, &bindings->uav_counter_descriptor_set, 0, NULL)); + bindings->uav_counter_in_use = true; + bindings->uav_counter_descriptor_set_bound = bindings->uav_counter_descriptor_set; + } } static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list) @@ -4032,6 +4220,10 @@ static void d3d12_command_list_set_root_signature(struct d3d12_command_list *lis bindings->descriptor_set = VK_NULL_HANDLE; bindings->descriptor_table_dirty_mask = bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask = bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask;
+ bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; + bindings->descriptor_set_bound = VK_NULL_HANDLE; + bindings->uav_counter_descriptor_set_bound = VK_NULL_HANDLE; } static void STDMETHODCALLTYPE d3d12_command_list_SetComputeRootSignature(ID3D12GraphicsCommandList1 *iface, @@ -5405,6 +5597,9 @@ static HRESULT d3d12_command_list_init(struct d3d12_command_list *list, struct d d3d12_device_add_ref(list->device = device); list->allocator = allocator; + list->descriptor_updates = NULL; + list->descriptor_updates_count = 0; + list->descriptor_updates_size = 0; if (SUCCEEDED(hr = d3d12_command_allocator_allocate_command_buffer(allocator, list))) { @@ -5642,6 +5837,13 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm return; } + /* Descriptors in root signature in 1.0 are VOLATILE by default, so + * the descriptor heap only need to be valid right before we submit them to the GPU. + * If we have EXT_descriptor_indexing enabled with UpdateAfterBind, we update + * descriptor sets here rather than while we're recording the command buffer. + * For each submission of the command buffer, we can modify the descriptor heap as we please. */ + d3d12_command_list_resolve_descriptor_tables(cmd_list);
buffers[i] = cmd_list->vk_command_buffer; } diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index 13ebc70..e159c1f 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -1436,11 +1436,25 @@ static HRESULT vkd3d_init_device_caps(struct d3d12_device *device, } if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing - && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind - || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingSampledImageUpdateAfterBind + && descriptor_indexing->descriptorBindingStorageImageUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind) + { + TRACE("Enabling VK_EXT_descriptor_indexing for volatile descriptor updates.\n"); + } + else + { + WARN("VK_EXT_descriptor indexing not supported in sufficient capacity. Volatile descriptor updates will not work.\n"); + vulkan_info->EXT_descriptor_indexing = false; + }
+ if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing + && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind || descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind + || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind || descriptor_indexing->descriptorBindingStorageTexelBufferUpdateAfterBind) - && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) + && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) { WARN("Disabling robust buffer access for the update after bind feature.\n"); features->robustBufferAccess = VK_FALSE; diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index a321fa4..50cff1f 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -745,21 +745,43 @@ static HRESULT vkd3d_create_descriptor_set_layout(struct d3d12_device *device, VkDescriptorSetLayoutCreateFlags flags, unsigned int binding_count, const VkDescriptorSetLayoutBinding *bindings, VkDescriptorSetLayout *set_layout) { + unsigned int i; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkDescriptorSetLayoutCreateInfo set_desc; VkResult vr; + VkDescriptorSetLayoutBindingFlagsCreateInfoEXT flags_info; + VkDescriptorBindingFlagsEXT *binding_flags = NULL; set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO; set_desc.pNext = NULL; set_desc.flags = flags; set_desc.bindingCount = binding_count; set_desc.pBindings = bindings;
+ if (!(flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR) && device->vk_info.EXT_descriptor_indexing) + { + set_desc.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT; + flags_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT; + flags_info.pNext = NULL; + flags_info.bindingCount = binding_count; + binding_flags = vkd3d_malloc(sizeof(*binding_flags) * binding_count); + for (i = 0; i < binding_count; i++) + { + binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT |
- VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT;
+ } + flags_info.pBindingFlags = binding_flags; + set_desc.pNext = &flags_info; + }
if ((vr = VK_CALL(vkCreateDescriptorSetLayout(device->vk_device, &set_desc, NULL, set_layout))) < 0) { WARN("Failed to create Vulkan descriptor set layout, vr %d.\n", vr); + vkd3d_free(binding_flags); return hresult_from_vk_result(vr); } + vkd3d_free(binding_flags); return S_OK; } diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2d62fda..58e9aa9 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -888,13 +888,16 @@ struct vkd3d_pipeline_bindings const struct d3d12_root_signature *root_signature; VkDescriptorSet descriptor_set; + VkDescriptorSet uav_counter_descriptor_set; + VkDescriptorSet descriptor_set_bound; + VkDescriptorSet uav_counter_descriptor_set_bound; bool in_use; + bool uav_counter_in_use; D3D12_GPU_DESCRIPTOR_HANDLE descriptor_tables[D3D12_MAX_ROOT_COST]; uint64_t descriptor_table_dirty_mask; uint64_t descriptor_table_active_mask; - VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; uint8_t uav_counter_dirty_mask; /* Needed when VK_KHR_push_descriptor is not available. */ @@ -903,6 +906,16 @@ struct vkd3d_pipeline_bindings uint32_t push_descriptor_active_mask; }; +struct d3d12_deferred_descriptor_set_update +{ + const struct d3d12_desc *base_descriptor; + unsigned int index;
+ const struct d3d12_root_signature *root_signature; + VkDescriptorSet descriptor_set; + bool uav; +};
/* ID3D12CommandList */ struct d3d12_command_list { @@ -946,6 +959,10 @@ struct d3d12_command_list VkBuffer so_counter_buffers[D3D12_SO_BUFFER_SLOT_COUNT]; VkDeviceSize so_counter_buffer_offsets[D3D12_SO_BUFFER_SLOT_COUNT]; + struct d3d12_deferred_descriptor_set_update *descriptor_updates; + size_t descriptor_updates_size; + size_t descriptor_updates_count;
struct vkd3d_private_store private_store; }; diff --git a/tests/d3d12.c b/tests/d3d12.c index 858cf99..16a048c 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -15799,7 +15799,7 @@ static void test_update_descriptor_heap_after_closing_command_list(void) D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE); get_texture_readback_with_command_list(context.render_target, 0, &rb, queue, command_list); value = get_readback_uint(&rb, 0, 0, 0); - todo ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); + ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); release_resource_readback(&rb); ID3D12DescriptorHeap_Release(cpu_heap);
You are absolutely right. My bad! I had some dirt in MY tree after fiddling with the patch from this bug https://bugs.winehq.org/show_bug.cgi?id=46410 , and by doing a fresh clone it checks out AND compiles. Sorry for the noise :)
I added the 3 patches you posted:
https://source.winehq.org/patches/data/172721
https://source.winehq.org/patches/data/172677
https://source.winehq.org/patches/data/172679
And doing this, without adding the patch posted on this bug, WoW seems to work without flashing textures. Will do some more testing once this is committed to GIT. Thanks :)
Sveinar
On 01.11.2019 22:03, Hans-Kristian Arntzen wrote:
The patch was made based on that parent commit and compiles fine for me.
Cheers, Hans-Kristian
On 11/1/19 9:27 PM, Sveinar Søpler wrote:
I am not sure what tree you are working on, but applying this patch to GIT@4576236199769eacadc6de49e03d05df999f9181 leads to a lot of:
libs/vkd3d/command.c: In function ‘d3d12_command_list_resolve_descriptor_table_normal’: libs/vkd3d/command.c:2815:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_table’ 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ libs/vkd3d/command.c:2815:1: warning: ISO C90 forbids mixed declarations and code [-Wdeclaration-after-statement] 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~ libs/vkd3d/command.c:2824:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_tables’ 2824 | static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and subsequent errors.
Does this patch depend on some other patch(set) to compile?
From recent commits to vkd3d, it seems as some patches are picked, and some not depending on some criteria. Working on a massively changed git tree submitting patches that does not include its dependencies is kinda hard for me to test :)
If you have a git repo with a working tree i am willing to test stuff.
Sorry if you feel i am stepping on some toes here tho...
Sveinar
On 01.11.2019 14:36, Hans-Kristian Arntzen wrote:
Use EXT_descriptor_indexing's UPDATE_AFTER_BIND feature to support semantics required by RS 1.0 VOLATILE descriptors. We implement this by deferring all updates of desciptor sets until Submit time.
This is fine, as command buffers cannot be executed simultaneously on D3D12, so at Submit time, we know that the command buffer is not being executed on the GPU, and updating descriptors for multiple submissions is correct.
If EXT_descriptor_indexing is not available, the fallback is the older method, which matches RS 1.1 STATIC descriptor model.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
libs/vkd3d/command.c | 358 +++++++++++++++++++++++++++++-------- libs/vkd3d/device.c | 20 ++- libs/vkd3d/state.c | 22 +++ libs/vkd3d/vkd3d_private.h | 19 +- tests/d3d12.c | 2 +- 5 files changed, 338 insertions(+), 83 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 0532ec0..3f78d17 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1341,7 +1341,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( { pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL; - pool_desc.flags = 0; + pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; pool_desc.maxSets = 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes; @@ -1865,6 +1865,10 @@ static void d3d12_command_list_invalidate_bindings(struct d3d12_command_list *li if (!state) return; + /* Each pipeline state has its own set layout for UAV counters + * based on their implicit usage in the shader. + * Binding a different pipeline state means having to re-remit + * UAV counters in a new descriptor set (and layout). */ if (state->uav_counter_mask) { struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[state->vk_bind_point]; @@ -2196,6 +2200,7 @@ static ULONG STDMETHODCALLTYPE d3d12_command_list_Release(ID3D12GraphicsCommandL if (list->allocator) d3d12_command_allocator_free_command_buffer(list->allocator, list); + vkd3d_free(list->descriptor_updates); vkd3d_free(list); d3d12_device_release(device); @@ -2334,6 +2339,9 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list, list->state = NULL; + /* Recycle deferred descriptor update memory if possible. */ + list->descriptor_updates_count = 0;
memset(list->so_counter_buffers, 0, sizeof(list->so_counter_buffers)); memset(list->so_counter_buffer_offsets, 0, sizeof(list->so_counter_buffer_offsets)); @@ -2534,15 +2542,54 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * time in between. Thus, the contents must not be altered (overwritten * by an update command, or freed) between when the command is recorded * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, root_signature->vk_set_layout);
bindings->in_use = false; bindings->descriptor_table_dirty_mask |= bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask |= bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask; } +static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point) +{ + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
+ if (bindings->uav_counter_descriptor_set && !bindings->uav_counter_in_use) + return;
+ /* We cannot modify bound descriptor sets. We need a new descriptor set if + * we are about to update resource bindings. + * + * The Vulkan spec says: + * + * "The descriptor set contents bound by a call to + * vkCmdBindDescriptorSets may be consumed during host execution of the + * command, or during shader execution of the resulting draws, or any + * time in between. Thus, the contents must not be altered (overwritten + * by an update command, or freed) between when the command is recorded + * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. + */
+ if (list->state->uav_counter_mask) + { + bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, + list->state->vk_set_layout); + } + else + bindings->uav_counter_descriptor_set = VK_NULL_HANDLE;
+ bindings->uav_counter_in_use = false; +}
static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_descriptor_write, VkDescriptorImageInfo *vk_image_info, const struct d3d12_desc *descriptor, uint32_t descriptor_range_magic, VkDescriptorSet vk_descriptor_set, @@ -2611,26 +2658,61 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des return true; } -static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +static void d3d12_command_list_defer_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, uint64_t table_mask, bool uav) { + const struct d3d12_desc *base_descriptor; + unsigned i; + struct d3d12_deferred_descriptor_set_update *update; struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; - const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + { + if (table_mask & ((uint64_t)1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + { + vkd3d_array_reserve((void **)&list->descriptor_updates, &list->descriptor_updates_size,
- list->descriptor_updates_count + 1,
sizeof(*list->descriptor_updates)); + update = &list->descriptor_updates[list->descriptor_updates_count];
+ update->base_descriptor = base_descriptor; + update->index = i; + update->root_signature = bindings->root_signature; + update->descriptor_set = uav ? bindings->uav_counter_descriptor_set : bindings->descriptor_set; + update->uav = uav; + list->descriptor_updates_count++; + } + else + WARN("Descriptor table %u is not set.\n", i); + } + } +}
+static void d3d12_command_list_resolve_descriptor_table_uav(struct d3d12_command_list *list,
- const struct d3d12_deferred_descriptor_set_update *update)
+{ + const struct d3d12_root_signature *root_signature = update->root_signature; const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - struct VkDescriptorImageInfo image_infos[24], *current_image_info; const struct d3d12_root_descriptor_table *descriptor_table; const struct d3d12_root_descriptor_table_range *range; VkDevice vk_device = list->device->vk_device; - unsigned int i, j, descriptor_count; - struct d3d12_desc *descriptor;
- descriptor_table = root_signature_get_descriptor_table(root_signature, index); + unsigned int i, j; + unsigned int uav_counter_count; + const struct d3d12_desc *descriptor; + VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkDescriptorSet vk_descriptor_set; + const struct d3d12_desc *base_descriptor = update->base_descriptor; + descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); descriptor = base_descriptor; - descriptor_count = 0; - current_descriptor_write = descriptor_writes; - current_image_info = image_infos;
+ vk_descriptor_set = update->descriptor_set; + if (!vk_descriptor_set) + return;
+ /* FIXME: There should be a smarter way than scanning through all the descriptor table ranges for this. */ for (i = 0; i < descriptor_table->range_count; ++i) { range = &descriptor_table->ranges[i]; @@ -2644,20 +2726,74 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list { unsigned int register_idx = range->base_register_idx + j; - /* Track UAV counters. */ + /* Fish out UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV - && register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views)) + && register_idx < ARRAY_SIZE(vk_uav_counter_views)) { VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV - ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; - if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view) - bindings->uav_counter_dirty_mask |= 1u << register_idx;
- bindings->vk_uav_counter_views[register_idx] = vk_counter_view;
+ ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; + vk_uav_counter_views[register_idx] = vk_counter_view; } + } + }
+ uav_counter_count = vkd3d_popcount(list->state->uav_counter_mask); + assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
+ for (i = 0; i < uav_counter_count; ++i) + { + const struct vkd3d_shader_uav_counter_binding *uav_counter = &list->state->uav_counters[i];
- assert(vk_uav_counter_views[uav_counter->register_index]);
+ vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; + vk_descriptor_writes[i].pNext = NULL; + vk_descriptor_writes[i].dstSet = vk_descriptor_set; + vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; + vk_descriptor_writes[i].dstArrayElement = 0; + vk_descriptor_writes[i].descriptorCount = 1; + vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; + vk_descriptor_writes[i].pImageInfo = NULL; + vk_descriptor_writes[i].pBufferInfo = NULL; + vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; + }
+ VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL)); +}
+static void d3d12_command_list_resolve_descriptor_table_normal(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + const struct d3d12_root_descriptor_table *descriptor_table; + struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; + struct VkDescriptorImageInfo image_infos[24], *current_image_info; + const struct d3d12_root_signature *root_signature = update->root_signature; + const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; + const struct d3d12_root_descriptor_table_range *range; + VkDevice vk_device = list->device->vk_device; + unsigned int i, j, descriptor_count; + const struct d3d12_desc *descriptor; + const struct d3d12_desc *base_descriptor = update->base_descriptor;
+ descriptor_table = root_signature_get_descriptor_table(root_signature, update->index);
+ descriptor = update->base_descriptor; + descriptor_count = 0; + current_descriptor_write = descriptor_writes; + current_image_info = image_infos; + for (i = 0; i < descriptor_table->range_count; ++i) + { + range = &descriptor_table->ranges[i]; + if (range->offset != D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND) + { + descriptor = base_descriptor + range->offset; + }
+ for (j = 0; j < range->descriptor_count; ++j, ++descriptor) + { if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write, - current_image_info, descriptor, range->descriptor_magic, - bindings->descriptor_set, range->binding, j))
- current_image_info, descriptor, range->descriptor_magic,
- update->descriptor_set, range->binding, j))
continue; ++descriptor_count; @@ -2677,6 +2813,52 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list VK_CALL(vkUpdateDescriptorSets(vk_device, descriptor_count, descriptor_writes, 0, NULL)); } +static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + if (update->uav) + d3d12_command_list_resolve_descriptor_table_uav(list, update); + else
- d3d12_command_list_resolve_descriptor_table_normal(list, update);
+}
+static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) +{ + unsigned i; + for (i = 0; i < list->descriptor_updates_count; i++) + d3d12_command_list_resolve_descriptor_table(list, &list->descriptor_updates[i]); +}
+static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = false; + d3d12_command_list_resolve_descriptor_table_normal(list, &update); +}
+static void d3d12_command_list_update_uav_counter_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->uav_counter_descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = true; + d3d12_command_list_resolve_descriptor_table_uav(list, &update); +}
static bool vk_write_descriptor_set_from_root_descriptor(VkWriteDescriptorSet *vk_descriptor_write, const struct d3d12_root_parameter *root_parameter, VkDescriptorSet vk_descriptor_set, VkBufferView *vk_buffer_view, const VkDescriptorBufferInfo *vk_buffer_info) @@ -2781,55 +2963,6 @@ done: vkd3d_free(buffer_infos); } -static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point) -{ - VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; - struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - const struct d3d12_pipeline_state *state = list->state; - VkDevice vk_device = list->device->vk_device; - VkDescriptorSet vk_descriptor_set; - unsigned int uav_counter_count; - unsigned int i;
- if (!state || !(state->uav_counter_mask & bindings->uav_counter_dirty_mask)) - return;
- uav_counter_count = vkd3d_popcount(state->uav_counter_mask); - assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
- vk_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, state->vk_set_layout); - if (!vk_descriptor_set) - return;
- for (i = 0; i < uav_counter_count; ++i) - { - const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i]; - const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views;
- assert(vk_uav_counter_views[uav_counter->register_index]);
- vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; - vk_descriptor_writes[i].pNext = NULL; - vk_descriptor_writes[i].dstSet = vk_descriptor_set; - vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; - vk_descriptor_writes[i].dstArrayElement = 0; - vk_descriptor_writes[i].descriptorCount = 1; - vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - vk_descriptor_writes[i].pImageInfo = NULL; - vk_descriptor_writes[i].pBufferInfo = NULL; - vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; - }
- VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
- state->vk_pipeline_layout, state->set_index, 1, &vk_descriptor_set, 0, NULL));
- bindings->uav_counter_dirty_mask = 0; -}
static void d3d12_command_list_update_descriptors(struct d3d12_command_list *list, VkPipelineBindPoint bind_point) { @@ -2842,31 +2975,86 @@ static void d3d12_command_list_update_descriptors(struct d3d12_command_list *lis if (!rs || !rs->vk_set_layout) return; + if ((bindings->descriptor_table_active_mask | bindings->push_descriptor_dirty_mask | + (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)) == 0) + { + /* Nothing is dirty, so just return early. */ + return; + }
if (bindings->descriptor_table_dirty_mask || bindings->push_descriptor_dirty_mask) d3d12_command_list_prepare_descriptors(list, bind_point); + if (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)
- d3d12_command_list_prepare_uav_counter_descriptors(list, bind_point);
- for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + if (list->device->vk_info.EXT_descriptor_indexing) + {
- d3d12_command_list_defer_update_descriptor_table(list, bind_point,
bindings->descriptor_table_dirty_mask,
- false);
+ } + else { - if (bindings->descriptor_table_dirty_mask & ((uint64_t)1 << i)) + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) { - if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); - else - WARN("Descriptor table %u is not set.\n", i); + if (bindings->descriptor_table_dirty_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } } } bindings->descriptor_table_dirty_mask = 0; + /* Need to go through all descriptor tables here in the root signature, + * not just descriptor_table_dirty_mask. Binding a different shader may not invalidate descriptor tables, + * but it may invalidate the UAV counter set. */ + if (bindings->uav_counter_dirty_mask) + { + if (list->device->vk_info.EXT_descriptor_indexing) + {
- d3d12_command_list_defer_update_descriptor_table(list, bind_point,
bindings->descriptor_table_active_mask,
- true);
+ } + else + { + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); i++) + { + if (bindings->descriptor_table_active_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_uav_counter_descriptor_table(list,
bind_point, i, base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } + } + } + } + bindings->uav_counter_dirty_mask = 0;
d3d12_command_list_update_push_descriptors(list, bind_point); - if (bindings->descriptor_set) + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->descriptor_set && bindings->descriptor_set != bindings->descriptor_set_bound) { VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, rs->vk_pipeline_layout, rs->main_set, 1, &bindings->descriptor_set, 0, NULL)); bindings->in_use = true; + bindings->descriptor_set_bound = bindings->descriptor_set; } - d3d12_command_list_update_uav_counter_descriptors(list, bind_point); + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->uav_counter_descriptor_set && + bindings->uav_counter_descriptor_set != bindings->uav_counter_descriptor_set_bound) + {
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
+ list->state->vk_pipeline_layout, list->state->set_index, 1, &bindings->uav_counter_descriptor_set, 0, NULL)); + bindings->uav_counter_in_use = true; + bindings->uav_counter_descriptor_set_bound = bindings->uav_counter_descriptor_set; + } } static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list) @@ -4032,6 +4220,10 @@ static void d3d12_command_list_set_root_signature(struct d3d12_command_list *lis bindings->descriptor_set = VK_NULL_HANDLE; bindings->descriptor_table_dirty_mask = bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask = bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask;
+ bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; + bindings->descriptor_set_bound = VK_NULL_HANDLE; + bindings->uav_counter_descriptor_set_bound = VK_NULL_HANDLE; } static void STDMETHODCALLTYPE d3d12_command_list_SetComputeRootSignature(ID3D12GraphicsCommandList1 *iface, @@ -5405,6 +5597,9 @@ static HRESULT d3d12_command_list_init(struct d3d12_command_list *list, struct d d3d12_device_add_ref(list->device = device); list->allocator = allocator; + list->descriptor_updates = NULL; + list->descriptor_updates_count = 0; + list->descriptor_updates_size = 0; if (SUCCEEDED(hr = d3d12_command_allocator_allocate_command_buffer(allocator, list))) { @@ -5642,6 +5837,13 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm return; } + /* Descriptors in root signature in 1.0 are VOLATILE by default, so + * the descriptor heap only need to be valid right before we submit them to the GPU. + * If we have EXT_descriptor_indexing enabled with UpdateAfterBind, we update + * descriptor sets here rather than while we're recording the command buffer. + * For each submission of the command buffer, we can modify the descriptor heap as we please. */
- d3d12_command_list_resolve_descriptor_tables(cmd_list);
buffers[i] = cmd_list->vk_command_buffer; } diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index 13ebc70..e159c1f 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -1436,11 +1436,25 @@ static HRESULT vkd3d_init_device_caps(struct d3d12_device *device, } if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing - && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind - || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingSampledImageUpdateAfterBind + && descriptor_indexing->descriptorBindingStorageImageUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind) + { + TRACE("Enabling VK_EXT_descriptor_indexing for volatile descriptor updates.\n"); + } + else + { + WARN("VK_EXT_descriptor indexing not supported in sufficient capacity. Volatile descriptor updates will not work.\n"); + vulkan_info->EXT_descriptor_indexing = false; + }
+ if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing + && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind || descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind + || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind || descriptor_indexing->descriptorBindingStorageTexelBufferUpdateAfterBind) - && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) + && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) { WARN("Disabling robust buffer access for the update after bind feature.\n"); features->robustBufferAccess = VK_FALSE; diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index a321fa4..50cff1f 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -745,21 +745,43 @@ static HRESULT vkd3d_create_descriptor_set_layout(struct d3d12_device *device, VkDescriptorSetLayoutCreateFlags flags, unsigned int binding_count, const VkDescriptorSetLayoutBinding *bindings, VkDescriptorSetLayout *set_layout) { + unsigned int i; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkDescriptorSetLayoutCreateInfo set_desc; VkResult vr; + VkDescriptorSetLayoutBindingFlagsCreateInfoEXT flags_info; + VkDescriptorBindingFlagsEXT *binding_flags = NULL; set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO; set_desc.pNext = NULL; set_desc.flags = flags; set_desc.bindingCount = binding_count; set_desc.pBindings = bindings;
+ if (!(flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR) && device->vk_info.EXT_descriptor_indexing) + { + set_desc.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT; + flags_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT; + flags_info.pNext = NULL; + flags_info.bindingCount = binding_count; + binding_flags = vkd3d_malloc(sizeof(*binding_flags) * binding_count); + for (i = 0; i < binding_count; i++) + { + binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT |
- VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT;
+ } + flags_info.pBindingFlags = binding_flags; + set_desc.pNext = &flags_info; + }
if ((vr = VK_CALL(vkCreateDescriptorSetLayout(device->vk_device, &set_desc, NULL, set_layout))) < 0) { WARN("Failed to create Vulkan descriptor set layout, vr %d.\n", vr); + vkd3d_free(binding_flags); return hresult_from_vk_result(vr); } + vkd3d_free(binding_flags); return S_OK; } diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2d62fda..58e9aa9 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -888,13 +888,16 @@ struct vkd3d_pipeline_bindings const struct d3d12_root_signature *root_signature; VkDescriptorSet descriptor_set; + VkDescriptorSet uav_counter_descriptor_set; + VkDescriptorSet descriptor_set_bound; + VkDescriptorSet uav_counter_descriptor_set_bound; bool in_use; + bool uav_counter_in_use; D3D12_GPU_DESCRIPTOR_HANDLE descriptor_tables[D3D12_MAX_ROOT_COST]; uint64_t descriptor_table_dirty_mask; uint64_t descriptor_table_active_mask; - VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; uint8_t uav_counter_dirty_mask; /* Needed when VK_KHR_push_descriptor is not available. */ @@ -903,6 +906,16 @@ struct vkd3d_pipeline_bindings uint32_t push_descriptor_active_mask; }; +struct d3d12_deferred_descriptor_set_update +{ + const struct d3d12_desc *base_descriptor; + unsigned int index;
+ const struct d3d12_root_signature *root_signature; + VkDescriptorSet descriptor_set; + bool uav; +};
/* ID3D12CommandList */ struct d3d12_command_list { @@ -946,6 +959,10 @@ struct d3d12_command_list VkBuffer so_counter_buffers[D3D12_SO_BUFFER_SLOT_COUNT]; VkDeviceSize so_counter_buffer_offsets[D3D12_SO_BUFFER_SLOT_COUNT]; + struct d3d12_deferred_descriptor_set_update *descriptor_updates; + size_t descriptor_updates_size; + size_t descriptor_updates_count;
struct vkd3d_private_store private_store; }; diff --git a/tests/d3d12.c b/tests/d3d12.c index 858cf99..16a048c 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -15799,7 +15799,7 @@ static void test_update_descriptor_heap_after_closing_command_list(void) D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE); get_texture_readback_with_command_list(context.render_target, 0, &rb, queue, command_list); value = get_readback_uint(&rb, 0, 0, 0); - todo ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); + ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); release_resource_readback(&rb); ID3D12DescriptorHeap_Release(cpu_heap);
Thanks for testing. I had a hunch WoW would be fixed with this patch as it had stricter invalidation of UAV counters, but I didn't have retail client installed to test.
Cheers, Hans-Kristian
On 11/2/19 6:44 PM, Sveinar Søpler wrote:
You are absolutely right. My bad! I had some dirt in MY tree after fiddling with the patch from this bug https://bugs.winehq.org/show_bug.cgi?id=46410 , and by doing a fresh clone it checks out AND compiles. Sorry for the noise :)
I added the 3 patches you posted:
https://source.winehq.org/patches/data/172721
https://source.winehq.org/patches/data/172677
https://source.winehq.org/patches/data/172679
And doing this, without adding the patch posted on this bug, WoW seems to work without flashing textures. Will do some more testing once this is committed to GIT. Thanks :)
Sveinar
On 01.11.2019 22:03, Hans-Kristian Arntzen wrote:
The patch was made based on that parent commit and compiles fine for me.
Cheers, Hans-Kristian
On 11/1/19 9:27 PM, Sveinar Søpler wrote:
I am not sure what tree you are working on, but applying this patch to GIT@4576236199769eacadc6de49e03d05df999f9181 leads to a lot of:
libs/vkd3d/command.c: In function ‘d3d12_command_list_resolve_descriptor_table_normal’: libs/vkd3d/command.c:2815:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_table’ 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ libs/vkd3d/command.c:2815:1: warning: ISO C90 forbids mixed declarations and code [-Wdeclaration-after-statement] 2815 | static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list, | ^~~~~~ libs/vkd3d/command.c:2824:13: error: invalid storage class for function ‘d3d12_command_list_resolve_descriptor_tables’ 2824 | static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and subsequent errors.
Does this patch depend on some other patch(set) to compile?
From recent commits to vkd3d, it seems as some patches are picked, and some not depending on some criteria. Working on a massively changed git tree submitting patches that does not include its dependencies is kinda hard for me to test :)
If you have a git repo with a working tree i am willing to test stuff.
Sorry if you feel i am stepping on some toes here tho...
Sveinar
On 01.11.2019 14:36, Hans-Kristian Arntzen wrote:
Use EXT_descriptor_indexing's UPDATE_AFTER_BIND feature to support semantics required by RS 1.0 VOLATILE descriptors. We implement this by deferring all updates of desciptor sets until Submit time.
This is fine, as command buffers cannot be executed simultaneously on D3D12, so at Submit time, we know that the command buffer is not being executed on the GPU, and updating descriptors for multiple submissions is correct.
If EXT_descriptor_indexing is not available, the fallback is the older method, which matches RS 1.1 STATIC descriptor model.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
libs/vkd3d/command.c | 358 +++++++++++++++++++++++++++++-------- libs/vkd3d/device.c | 20 ++- libs/vkd3d/state.c | 22 +++ libs/vkd3d/vkd3d_private.h | 19 +- tests/d3d12.c | 2 +- 5 files changed, 338 insertions(+), 83 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 0532ec0..3f78d17 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -1341,7 +1341,7 @@ static VkDescriptorPool d3d12_command_allocator_allocate_descriptor_pool( { pool_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO; pool_desc.pNext = NULL; - pool_desc.flags = 0; + pool_desc.flags = device->vk_info.EXT_descriptor_indexing ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT : 0; pool_desc.maxSets = 512; pool_desc.poolSizeCount = ARRAY_SIZE(pool_sizes); pool_desc.pPoolSizes = pool_sizes; @@ -1865,6 +1865,10 @@ static void d3d12_command_list_invalidate_bindings(struct d3d12_command_list *li if (!state) return; + /* Each pipeline state has its own set layout for UAV counters + * based on their implicit usage in the shader. + * Binding a different pipeline state means having to re-remit + * UAV counters in a new descriptor set (and layout). */ if (state->uav_counter_mask) { struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[state->vk_bind_point]; @@ -2196,6 +2200,7 @@ static ULONG STDMETHODCALLTYPE d3d12_command_list_Release(ID3D12GraphicsCommandL if (list->allocator) d3d12_command_allocator_free_command_buffer(list->allocator, list); + vkd3d_free(list->descriptor_updates); vkd3d_free(list); d3d12_device_release(device); @@ -2334,6 +2339,9 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list, list->state = NULL; + /* Recycle deferred descriptor update memory if possible. */ + list->descriptor_updates_count = 0;
memset(list->so_counter_buffers, 0, sizeof(list->so_counter_buffers)); memset(list->so_counter_buffer_offsets, 0, sizeof(list->so_counter_buffer_offsets)); @@ -2534,15 +2542,54 @@ static void d3d12_command_list_prepare_descriptors(struct d3d12_command_list *li * time in between. Thus, the contents must not be altered (overwritten * by an update command, or freed) between when the command is recorded * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. */ bindings->descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, root_signature->vk_set_layout);
bindings->in_use = false; bindings->descriptor_table_dirty_mask |= bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask |= bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask; } +static void d3d12_command_list_prepare_uav_counter_descriptors(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point) +{ + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point];
+ if (bindings->uav_counter_descriptor_set && !bindings->uav_counter_in_use) + return;
+ /* We cannot modify bound descriptor sets. We need a new descriptor set if + * we are about to update resource bindings. + * + * The Vulkan spec says: + * + * "The descriptor set contents bound by a call to + * vkCmdBindDescriptorSets may be consumed during host execution of the + * command, or during shader execution of the resulting draws, or any + * time in between. Thus, the contents must not be altered (overwritten + * by an update command, or freed) between when the command is recorded + * and when the command completes executing on the queue." + * + * Even if we have descriptor indexing and UPDATE_AFTER_BIND, + * we need at the very least a new descriptor set. + */
+ if (list->state->uav_counter_mask) + { + bindings->uav_counter_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, + list->state->vk_set_layout); + } + else + bindings->uav_counter_descriptor_set = VK_NULL_HANDLE;
+ bindings->uav_counter_in_use = false; +}
static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_descriptor_write, VkDescriptorImageInfo *vk_image_info, const struct d3d12_desc *descriptor, uint32_t descriptor_range_magic, VkDescriptorSet vk_descriptor_set, @@ -2611,26 +2658,61 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des return true; } -static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +static void d3d12_command_list_defer_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, uint64_t table_mask, bool uav) { + const struct d3d12_desc *base_descriptor; + unsigned i; + struct d3d12_deferred_descriptor_set_update *update; struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; - const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + { + if (table_mask & ((uint64_t)1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i]))) + { + vkd3d_array_reserve((void **)&list->descriptor_updates, &list->descriptor_updates_size,
- list->descriptor_updates_count + 1,
sizeof(*list->descriptor_updates)); + update = &list->descriptor_updates[list->descriptor_updates_count];
+ update->base_descriptor = base_descriptor; + update->index = i; + update->root_signature = bindings->root_signature; + update->descriptor_set = uav ? bindings->uav_counter_descriptor_set : bindings->descriptor_set; + update->uav = uav; + list->descriptor_updates_count++; + } + else + WARN("Descriptor table %u is not set.\n", i); + } + } +}
+static void d3d12_command_list_resolve_descriptor_table_uav(struct d3d12_command_list *list,
- const struct d3d12_deferred_descriptor_set_update *update)
+{ + const struct d3d12_root_signature *root_signature = update->root_signature; const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - struct VkDescriptorImageInfo image_infos[24], *current_image_info; const struct d3d12_root_descriptor_table *descriptor_table; const struct d3d12_root_descriptor_table_range *range; VkDevice vk_device = list->device->vk_device; - unsigned int i, j, descriptor_count; - struct d3d12_desc *descriptor;
- descriptor_table = root_signature_get_descriptor_table(root_signature, index); + unsigned int i, j; + unsigned int uav_counter_count; + const struct d3d12_desc *descriptor; + VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; + VkDescriptorSet vk_descriptor_set; + const struct d3d12_desc *base_descriptor = update->base_descriptor; + descriptor_table = root_signature_get_descriptor_table(root_signature, update->index); descriptor = base_descriptor; - descriptor_count = 0; - current_descriptor_write = descriptor_writes; - current_image_info = image_infos;
+ vk_descriptor_set = update->descriptor_set; + if (!vk_descriptor_set) + return;
+ /* FIXME: There should be a smarter way than scanning through all the descriptor table ranges for this. */ for (i = 0; i < descriptor_table->range_count; ++i) { range = &descriptor_table->ranges[i]; @@ -2644,20 +2726,74 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list { unsigned int register_idx = range->base_register_idx
- j;
- /* Track UAV counters. */ + /* Fish out UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV - && register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views)) + && register_idx < ARRAY_SIZE(vk_uav_counter_views)) { VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV - ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; - if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view) - bindings->uav_counter_dirty_mask |= 1u << register_idx;
- bindings->vk_uav_counter_views[register_idx] = vk_counter_view;
+ ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE; + vk_uav_counter_views[register_idx] = vk_counter_view; } + } + }
+ uav_counter_count = vkd3d_popcount(list->state->uav_counter_mask); + assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
+ for (i = 0; i < uav_counter_count; ++i) + { + const struct vkd3d_shader_uav_counter_binding *uav_counter = &list->state->uav_counters[i];
- assert(vk_uav_counter_views[uav_counter->register_index]);
+ vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; + vk_descriptor_writes[i].pNext = NULL; + vk_descriptor_writes[i].dstSet = vk_descriptor_set; + vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; + vk_descriptor_writes[i].dstArrayElement = 0; + vk_descriptor_writes[i].descriptorCount = 1; + vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; + vk_descriptor_writes[i].pImageInfo = NULL; + vk_descriptor_writes[i].pBufferInfo = NULL; + vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; + }
+ VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL)); +}
+static void d3d12_command_list_resolve_descriptor_table_normal(struct d3d12_command_list *list, + const struct d3d12_deferred_descriptor_set_update *update) +{ + const struct d3d12_root_descriptor_table *descriptor_table; + struct VkWriteDescriptorSet descriptor_writes[24], *current_descriptor_write; + struct VkDescriptorImageInfo image_infos[24], *current_image_info; + const struct d3d12_root_signature *root_signature = update->root_signature; + const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; + const struct d3d12_root_descriptor_table_range *range; + VkDevice vk_device = list->device->vk_device; + unsigned int i, j, descriptor_count; + const struct d3d12_desc *descriptor; + const struct d3d12_desc *base_descriptor = update->base_descriptor;
+ descriptor_table = root_signature_get_descriptor_table(root_signature, update->index);
+ descriptor = update->base_descriptor; + descriptor_count = 0; + current_descriptor_write = descriptor_writes; + current_image_info = image_infos; + for (i = 0; i < descriptor_table->range_count; ++i) + { + range = &descriptor_table->ranges[i]; + if (range->offset != D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND) + { + descriptor = base_descriptor + range->offset; + }
+ for (j = 0; j < range->descriptor_count; ++j, ++descriptor) + { if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write, - current_image_info, descriptor, range->descriptor_magic, - bindings->descriptor_set, range->binding, j))
- current_image_info, descriptor, range->descriptor_magic,
- update->descriptor_set, range->binding, j))
continue; ++descriptor_count; @@ -2677,6 +2813,52 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list VK_CALL(vkUpdateDescriptorSets(vk_device, descriptor_count, descriptor_writes, 0, NULL)); } +static void d3d12_command_list_resolve_descriptor_table(struct d3d12_command_list *list,
- const struct d3d12_deferred_descriptor_set_update *update)
+{ + if (update->uav)
- d3d12_command_list_resolve_descriptor_table_uav(list, update);
+ else
- d3d12_command_list_resolve_descriptor_table_normal(list, update);
+}
+static void d3d12_command_list_resolve_descriptor_tables(struct d3d12_command_list *list) +{ + unsigned i; + for (i = 0; i < list->descriptor_updates_count; i++) + d3d12_command_list_resolve_descriptor_table(list, &list->descriptor_updates[i]); +}
+static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = false;
- d3d12_command_list_resolve_descriptor_table_normal(list, &update);
+}
+static void d3d12_command_list_update_uav_counter_descriptor_table(struct d3d12_command_list *list, + VkPipelineBindPoint bind_point, unsigned int index, struct d3d12_desc *base_descriptor) +{ + struct d3d12_deferred_descriptor_set_update update; + struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; + const struct d3d12_root_signature *root_signature = bindings->root_signature;
+ update.descriptor_set = bindings->uav_counter_descriptor_set; + update.index = index; + update.root_signature = root_signature; + update.base_descriptor = base_descriptor; + update.uav = true; + d3d12_command_list_resolve_descriptor_table_uav(list, &update); +}
static bool vk_write_descriptor_set_from_root_descriptor(VkWriteDescriptorSet *vk_descriptor_write, const struct d3d12_root_parameter *root_parameter, VkDescriptorSet vk_descriptor_set, VkBufferView *vk_buffer_view, const VkDescriptorBufferInfo *vk_buffer_info) @@ -2781,55 +2963,6 @@ done: vkd3d_free(buffer_infos); } -static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_command_list *list, - VkPipelineBindPoint bind_point) -{ - VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; - struct vkd3d_pipeline_bindings *bindings = &list->pipeline_bindings[bind_point]; - const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; - const struct d3d12_pipeline_state *state = list->state; - VkDevice vk_device = list->device->vk_device; - VkDescriptorSet vk_descriptor_set; - unsigned int uav_counter_count; - unsigned int i;
- if (!state || !(state->uav_counter_mask & bindings->uav_counter_dirty_mask)) - return;
- uav_counter_count = vkd3d_popcount(state->uav_counter_mask); - assert(uav_counter_count <= ARRAY_SIZE(vk_descriptor_writes));
- vk_descriptor_set = d3d12_command_allocator_allocate_descriptor_set(list->allocator, state->vk_set_layout); - if (!vk_descriptor_set) - return;
- for (i = 0; i < uav_counter_count; ++i) - { - const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i]; - const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views;
- assert(vk_uav_counter_views[uav_counter->register_index]);
- vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; - vk_descriptor_writes[i].pNext = NULL; - vk_descriptor_writes[i].dstSet = vk_descriptor_set; - vk_descriptor_writes[i].dstBinding = uav_counter->binding.binding; - vk_descriptor_writes[i].dstArrayElement = 0; - vk_descriptor_writes[i].descriptorCount = 1; - vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - vk_descriptor_writes[i].pImageInfo = NULL; - vk_descriptor_writes[i].pBufferInfo = NULL; - vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index]; - }
- VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
- state->vk_pipeline_layout, state->set_index, 1, &vk_descriptor_set, 0, NULL));
- bindings->uav_counter_dirty_mask = 0; -}
static void d3d12_command_list_update_descriptors(struct d3d12_command_list *list, VkPipelineBindPoint bind_point) { @@ -2842,31 +2975,86 @@ static void d3d12_command_list_update_descriptors(struct d3d12_command_list *lis if (!rs || !rs->vk_set_layout) return; + if ((bindings->descriptor_table_active_mask | bindings->push_descriptor_dirty_mask | + (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)) == 0) + { + /* Nothing is dirty, so just return early. */ + return; + }
if (bindings->descriptor_table_dirty_mask || bindings->push_descriptor_dirty_mask) d3d12_command_list_prepare_descriptors(list, bind_point); + if (list->state->uav_counter_mask & bindings->uav_counter_dirty_mask)
- d3d12_command_list_prepare_uav_counter_descriptors(list,
bind_point); - for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) + if (list->device->vk_info.EXT_descriptor_indexing) + {
- d3d12_command_list_defer_update_descriptor_table(list,
bind_point, bindings->descriptor_table_dirty_mask,
- false);
+ } + else { - if (bindings->descriptor_table_dirty_mask & ((uint64_t)1 << i)) + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i) { - if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); - else - WARN("Descriptor table %u is not set.\n", i); + if (bindings->descriptor_table_dirty_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_descriptor_table(list, bind_point, i,
base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } } } bindings->descriptor_table_dirty_mask = 0; + /* Need to go through all descriptor tables here in the root signature, + * not just descriptor_table_dirty_mask. Binding a different shader may not invalidate descriptor tables, + * but it may invalidate the UAV counter set. */ + if (bindings->uav_counter_dirty_mask) + { + if (list->device->vk_info.EXT_descriptor_indexing) + {
- d3d12_command_list_defer_update_descriptor_table(list,
bind_point, bindings->descriptor_table_active_mask,
- true);
+ } + else + { + /* FIXME: FOR_EACH_BIT */ + for (i = 0; i < ARRAY_SIZE(bindings->descriptor_tables); i++) + { + if (bindings->descriptor_table_active_mask & ((uint64_t) 1 << i)) + { + if ((base_descriptor = d3d12_desc_from_gpu_handle(bindings->descriptor_tables[i])))
- d3d12_command_list_update_uav_counter_descriptor_table(list,
bind_point, i, base_descriptor); + else + WARN("Descriptor table %u is not set.\n", i); + } + } + } + } + bindings->uav_counter_dirty_mask = 0;
d3d12_command_list_update_push_descriptors(list, bind_point); - if (bindings->descriptor_set) + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->descriptor_set && bindings->descriptor_set != bindings->descriptor_set_bound) { VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point, rs->vk_pipeline_layout, rs->main_set, 1, &bindings->descriptor_set, 0, NULL)); bindings->in_use = true; + bindings->descriptor_set_bound = bindings->descriptor_set; } - d3d12_command_list_update_uav_counter_descriptors(list, bind_point); + /* Don't rebind the same descriptor set as long as we're in same pipeline layout. */ + if (bindings->uav_counter_descriptor_set && + bindings->uav_counter_descriptor_set != bindings->uav_counter_descriptor_set_bound) + {
- VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bind_point,
+ list->state->vk_pipeline_layout, list->state->set_index, 1, &bindings->uav_counter_descriptor_set, 0, NULL)); + bindings->uav_counter_in_use = true; + bindings->uav_counter_descriptor_set_bound = bindings->uav_counter_descriptor_set; + } } static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list) @@ -4032,6 +4220,10 @@ static void d3d12_command_list_set_root_signature(struct d3d12_command_list *lis bindings->descriptor_set = VK_NULL_HANDLE; bindings->descriptor_table_dirty_mask = bindings->descriptor_table_active_mask & root_signature->descriptor_table_mask; bindings->push_descriptor_dirty_mask = bindings->push_descriptor_active_mask & root_signature->push_descriptor_mask;
+ bindings->uav_counter_descriptor_set = VK_NULL_HANDLE; + bindings->descriptor_set_bound = VK_NULL_HANDLE; + bindings->uav_counter_descriptor_set_bound = VK_NULL_HANDLE; } static void STDMETHODCALLTYPE d3d12_command_list_SetComputeRootSignature(ID3D12GraphicsCommandList1 *iface, @@ -5405,6 +5597,9 @@ static HRESULT d3d12_command_list_init(struct d3d12_command_list *list, struct d d3d12_device_add_ref(list->device = device); list->allocator = allocator; + list->descriptor_updates = NULL; + list->descriptor_updates_count = 0; + list->descriptor_updates_size = 0; if (SUCCEEDED(hr = d3d12_command_allocator_allocate_command_buffer(allocator, list))) { @@ -5642,6 +5837,13 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm return; } + /* Descriptors in root signature in 1.0 are VOLATILE by default, so + * the descriptor heap only need to be valid right before we submit them to the GPU. + * If we have EXT_descriptor_indexing enabled with UpdateAfterBind, we update + * descriptor sets here rather than while we're recording the command buffer. + * For each submission of the command buffer, we can modify the descriptor heap as we please. */
- d3d12_command_list_resolve_descriptor_tables(cmd_list);
buffers[i] = cmd_list->vk_command_buffer; } diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index 13ebc70..e159c1f 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -1436,11 +1436,25 @@ static HRESULT vkd3d_init_device_caps(struct d3d12_device *device, } if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing - && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind - || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind + && descriptor_indexing->descriptorBindingSampledImageUpdateAfterBind + && descriptor_indexing->descriptorBindingStorageImageUpdateAfterBind + && descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind) + { + TRACE("Enabling VK_EXT_descriptor_indexing for volatile descriptor updates.\n"); + } + else + { + WARN("VK_EXT_descriptor indexing not supported in sufficient capacity. Volatile descriptor updates will not work.\n"); + vulkan_info->EXT_descriptor_indexing = false; + }
+ if (vulkan_info->EXT_descriptor_indexing && descriptor_indexing + && (descriptor_indexing->descriptorBindingUniformBufferUpdateAfterBind || descriptor_indexing->descriptorBindingUniformTexelBufferUpdateAfterBind + || descriptor_indexing->descriptorBindingStorageBufferUpdateAfterBind || descriptor_indexing->descriptorBindingStorageTexelBufferUpdateAfterBind) - && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) + && !physical_device_info->descriptor_indexing_properties.robustBufferAccessUpdateAfterBind) { WARN("Disabling robust buffer access for the update after bind feature.\n"); features->robustBufferAccess = VK_FALSE; diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index a321fa4..50cff1f 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -745,21 +745,43 @@ static HRESULT vkd3d_create_descriptor_set_layout(struct d3d12_device *device, VkDescriptorSetLayoutCreateFlags flags, unsigned int binding_count, const VkDescriptorSetLayoutBinding *bindings, VkDescriptorSetLayout *set_layout) { + unsigned int i; const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkDescriptorSetLayoutCreateInfo set_desc; VkResult vr; + VkDescriptorSetLayoutBindingFlagsCreateInfoEXT flags_info; + VkDescriptorBindingFlagsEXT *binding_flags = NULL; set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO; set_desc.pNext = NULL; set_desc.flags = flags; set_desc.bindingCount = binding_count; set_desc.pBindings = bindings;
+ if (!(flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR) && device->vk_info.EXT_descriptor_indexing) + { + set_desc.flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT_EXT; + flags_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT; + flags_info.pNext = NULL; + flags_info.bindingCount = binding_count; + binding_flags = vkd3d_malloc(sizeof(*binding_flags) * binding_count); + for (i = 0; i < binding_count; i++) + { + binding_flags[i] = VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT_EXT |
- VK_DESCRIPTOR_BINDING_UPDATE_UNUSED_WHILE_PENDING_BIT_EXT;
+ } + flags_info.pBindingFlags = binding_flags; + set_desc.pNext = &flags_info; + }
if ((vr = VK_CALL(vkCreateDescriptorSetLayout(device->vk_device, &set_desc, NULL, set_layout))) < 0) { WARN("Failed to create Vulkan descriptor set layout, vr %d.\n", vr); + vkd3d_free(binding_flags); return hresult_from_vk_result(vr); } + vkd3d_free(binding_flags); return S_OK; } diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2d62fda..58e9aa9 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -888,13 +888,16 @@ struct vkd3d_pipeline_bindings const struct d3d12_root_signature *root_signature; VkDescriptorSet descriptor_set; + VkDescriptorSet uav_counter_descriptor_set; + VkDescriptorSet descriptor_set_bound; + VkDescriptorSet uav_counter_descriptor_set_bound; bool in_use; + bool uav_counter_in_use; D3D12_GPU_DESCRIPTOR_HANDLE descriptor_tables[D3D12_MAX_ROOT_COST]; uint64_t descriptor_table_dirty_mask; uint64_t descriptor_table_active_mask; - VkBufferView vk_uav_counter_views[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS]; uint8_t uav_counter_dirty_mask; /* Needed when VK_KHR_push_descriptor is not available. */ @@ -903,6 +906,16 @@ struct vkd3d_pipeline_bindings uint32_t push_descriptor_active_mask; }; +struct d3d12_deferred_descriptor_set_update +{ + const struct d3d12_desc *base_descriptor; + unsigned int index;
+ const struct d3d12_root_signature *root_signature; + VkDescriptorSet descriptor_set; + bool uav; +};
/* ID3D12CommandList */ struct d3d12_command_list { @@ -946,6 +959,10 @@ struct d3d12_command_list VkBuffer so_counter_buffers[D3D12_SO_BUFFER_SLOT_COUNT]; VkDeviceSize so_counter_buffer_offsets[D3D12_SO_BUFFER_SLOT_COUNT]; + struct d3d12_deferred_descriptor_set_update *descriptor_updates; + size_t descriptor_updates_size; + size_t descriptor_updates_count;
struct vkd3d_private_store private_store; }; diff --git a/tests/d3d12.c b/tests/d3d12.c index 858cf99..16a048c 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -15799,7 +15799,7 @@ static void test_update_descriptor_heap_after_closing_command_list(void) D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE); get_texture_readback_with_command_list(context.render_target, 0, &rb, queue, command_list); value = get_readback_uint(&rb, 0, 0, 0); - todo ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); + ok(value == 0xff00ff00, "Got unexpected value %#x.\n", value); release_resource_readback(&rb); ID3D12DescriptorHeap_Release(cpu_heap);
-
Hans-Kristian Arntzen -
Henri Verbeet -
Sveinar Søpler