vk_graphics_pipeline: Use VK_KHR_push_descriptor when available

~51% faster on Nvidia compared to previous method.

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -16,38 +16,50 @@
 #include "video_core/texture_cache/texture_cache.h"
 #include "video_core/texture_cache/types.h"
 #include "video_core/textures/texture.h"
+#include "video_core/vulkan_common/vulkan_device.h"
 
 namespace Vulkan {
 
 class DescriptorLayoutBuilder {
 public:
-    DescriptorLayoutBuilder(const vk::Device& device_) : device{&device_} {}
+    DescriptorLayoutBuilder(const Device& device_) : device{&device_} {}
 
-    vk::DescriptorSetLayout CreateDescriptorSetLayout() const {
+    bool CanUsePushDescriptor() const noexcept {
+        return device->IsKhrPushDescriptorSupported() &&
+               num_descriptors <= device->MaxPushDescriptors();
+    }
+
+    vk::DescriptorSetLayout CreateDescriptorSetLayout(bool use_push_descriptor) const {
         if (bindings.empty()) {
             return nullptr;
         }
-        return device->CreateDescriptorSetLayout({
+        const VkDescriptorSetLayoutCreateFlags flags =
+            use_push_descriptor ? VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR : 0;
+        return device->GetLogical().CreateDescriptorSetLayout({
             .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
             .pNext = nullptr,
-            .flags = 0,
+            .flags = flags,
             .bindingCount = static_cast<u32>(bindings.size()),
             .pBindings = bindings.data(),
         });
     }
 
     vk::DescriptorUpdateTemplateKHR CreateTemplate(VkDescriptorSetLayout descriptor_set_layout,
-                                                   VkPipelineLayout pipeline_layout) const {
+                                                   VkPipelineLayout pipeline_layout,
+                                                   bool use_push_descriptor) const {
         if (entries.empty()) {
             return nullptr;
         }
-        return device->CreateDescriptorUpdateTemplateKHR({
+        const VkDescriptorUpdateTemplateType type =
+            use_push_descriptor ? VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR
+                                : VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET_KHR;
+        return device->GetLogical().CreateDescriptorUpdateTemplateKHR({
             .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_CREATE_INFO_KHR,
             .pNext = nullptr,
             .flags = 0,
             .descriptorUpdateEntryCount = static_cast<u32>(entries.size()),
             .pDescriptorUpdateEntries = entries.data(),
-            .templateType = VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET_KHR,
+            .templateType = type,
             .descriptorSetLayout = descriptor_set_layout,
             .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
             .pipelineLayout = pipeline_layout,
@@ -56,7 +68,7 @@ public:
     }
 
     vk::PipelineLayout CreatePipelineLayout(VkDescriptorSetLayout descriptor_set_layout) const {
-        return device->CreatePipelineLayout({
+        return device->GetLogical().CreatePipelineLayout({
             .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
             .pNext = nullptr,
             .flags = 0,
@@ -97,14 +109,16 @@ private:
                 .stride = sizeof(DescriptorUpdateEntry),
             });
             ++binding;
+            num_descriptors += descriptors[i].count;
             offset += sizeof(DescriptorUpdateEntry);
         }
     }
 
-    const vk::Device* device{};
+    const Device* device{};
     boost::container::small_vector<VkDescriptorSetLayoutBinding, 32> bindings;
     boost::container::small_vector<VkDescriptorUpdateTemplateEntryKHR, 32> entries;
     u32 binding{};
+    u32 num_descriptors{};
     size_t offset{};
 };
 

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -37,15 +37,14 @@ ComputePipeline::ComputePipeline(const Device& device_, DescriptorPool& descript
                 uniform_buffer_sizes.begin());
 
     auto func{[this, &descriptor_pool, shader_notify] {
-        DescriptorLayoutBuilder builder{device.GetLogical()};
+        DescriptorLayoutBuilder builder{device};
         builder.Add(info, VK_SHADER_STAGE_COMPUTE_BIT);
 
-        descriptor_set_layout = builder.CreateDescriptorSetLayout();
+        descriptor_set_layout = builder.CreateDescriptorSetLayout(false);
         pipeline_layout = builder.CreatePipelineLayout(*descriptor_set_layout);
         descriptor_update_template =
-            builder.CreateTemplate(*descriptor_set_layout, *pipeline_layout);
+            builder.CreateTemplate(*descriptor_set_layout, *pipeline_layout, false);
         descriptor_allocator = descriptor_pool.Allocator(*descriptor_set_layout, info);
-
         const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT subgroup_size_ci{
             .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT,
             .pNext = nullptr,
@@ -186,7 +185,6 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
     const void* const descriptor_data{update_descriptor_queue.UpdateData()};
     scheduler.Record([this, descriptor_data](vk::CommandBuffer cmdbuf) {
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
-
         if (!descriptor_set_layout) {
             return;
         }

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -40,7 +40,7 @@ constexpr size_t NUM_STAGES = Maxwell::MaxShaderStage;
 constexpr size_t MAX_IMAGE_ELEMENTS = 64;
 
 DescriptorLayoutBuilder MakeBuilder(const Device& device, std::span<const Shader::Info> infos) {
-    DescriptorLayoutBuilder builder{device.GetLogical()};
+    DescriptorLayoutBuilder builder{device};
     for (size_t index = 0; index < infos.size(); ++index) {
         static constexpr std::array stages{
             VK_SHADER_STAGE_VERTEX_BIT,
@@ -229,12 +229,15 @@ GraphicsPipeline::GraphicsPipeline(
     }
     auto func{[this, shader_notify, &render_pass_cache, &descriptor_pool] {
         DescriptorLayoutBuilder builder{MakeBuilder(device, stage_infos)};
-        descriptor_set_layout = builder.CreateDescriptorSetLayout();
+        uses_push_descriptor = builder.CanUsePushDescriptor();
+        descriptor_set_layout = builder.CreateDescriptorSetLayout(uses_push_descriptor);
+        if (!uses_push_descriptor) {
             descriptor_allocator = descriptor_pool.Allocator(*descriptor_set_layout, stage_infos);
-
+        }
         const VkDescriptorSetLayout set_layout{*descriptor_set_layout};
         pipeline_layout = builder.CreatePipelineLayout(set_layout);
-        descriptor_update_template = builder.CreateTemplate(set_layout, *pipeline_layout);
+        descriptor_update_template =
+            builder.CreateTemplate(set_layout, *pipeline_layout, uses_push_descriptor);
 
         const VkRenderPass render_pass{render_pass_cache.Get(MakeRenderPassKey(key.state))};
         Validate();
@@ -462,11 +465,16 @@ void GraphicsPipeline::ConfigureDraw() {
         if (!descriptor_set_layout) {
             return;
         }
+        if (uses_push_descriptor) {
+            cmdbuf.PushDescriptorSetWithTemplateKHR(*descriptor_update_template, *pipeline_layout,
+                                                    0, descriptor_data);
+        } else {
             const VkDescriptorSet descriptor_set{descriptor_allocator.Commit()};
             const vk::Device& dev{device.GetLogical()};
             dev.UpdateDescriptorSet(descriptor_set, *descriptor_update_template, descriptor_data);
             cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline_layout, 0,
                                       descriptor_set, nullptr);
+        }
     });
 }
 

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -148,6 +148,7 @@ private:
     std::condition_variable build_condvar;
     std::mutex build_mutex;
     std::atomic_bool is_built{false};
+    bool uses_push_descriptor{false};
 };
 
 } // namespace Vulkan

src/video_core/vulkan_common/vulkan_device.cpp

@@ -767,6 +767,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
         test(khr_uniform_buffer_standard_layout,
              VK_KHR_UNIFORM_BUFFER_STANDARD_LAYOUT_EXTENSION_NAME, true);
         test(khr_spirv_1_4, VK_KHR_SPIRV_1_4_EXTENSION_NAME, true);
+        test(khr_push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, true);
         test(has_khr_shader_float16_int8, VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, false);
         test(ext_depth_range_unrestricted, VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME, true);
         test(ext_index_type_uint8, VK_EXT_INDEX_TYPE_UINT8_EXTENSION_NAME, true);
@@ -932,6 +933,16 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
             khr_workgroup_memory_explicit_layout = true;
         }
     }
+    if (khr_push_descriptor) {
+        VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor;
+        push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
+        push_descriptor.pNext = nullptr;
+
+        physical_properties.pNext = &push_descriptor;
+        physical.GetProperties2KHR(physical_properties);
+
+        max_push_descriptors = push_descriptor.maxPushDescriptors;
+    }
     return extensions;
 }
 

src/video_core/vulkan_common/vulkan_device.h

@@ -154,6 +154,11 @@ public:
         return guest_warp_stages & stage;
     }
 
+    /// Returns the maximum number of push descriptors.
+    u32 MaxPushDescriptors() const {
+        return max_push_descriptors;
+    }
+
     /// Returns true if formatless image load is supported.
     bool IsFormatlessImageLoadSupported() const {
         return is_formatless_image_load_supported;
@@ -194,6 +199,11 @@ public:
         return khr_spirv_1_4;
     }
 
+    /// Returns true if the device supports VK_KHR_push_descriptor.
+    bool IsKhrPushDescriptorSupported() const {
+        return khr_push_descriptor;
+    }
+
     /// Returns true if the device supports VK_KHR_workgroup_memory_explicit_layout.
     bool IsKhrWorkgroupMemoryExplicitLayoutSupported() const {
         return khr_workgroup_memory_explicit_layout;
@@ -330,6 +340,7 @@ private:
     VkDriverIdKHR driver_id{};                  ///< Driver ID.
     VkShaderStageFlags guest_warp_stages{};     ///< Stages where the guest warp size can be forced.
     u64 device_access_memory{};                 ///< Total size of device local memory in bytes.
+    u32 max_push_descriptors{};                 ///< Maximum number of push descriptors
     bool is_optimal_astc_supported{};           ///< Support for native ASTC.
     bool is_float16_supported{};                ///< Support for float16 arithmetics.
     bool is_warp_potentially_bigger{};          ///< Host warp size can be bigger than guest.
@@ -345,6 +356,7 @@ private:
     bool khr_uniform_buffer_standard_layout{};  ///< Support for scalar uniform buffer layouts.
     bool khr_spirv_1_4{};                       ///< Support for VK_KHR_spirv_1_4.
     bool khr_workgroup_memory_explicit_layout{}; ///< Support for explicit workgroup layouts.
+    bool khr_push_descriptor{};                  ///< Support for VK_KHR_push_descritor.
     bool ext_index_type_uint8{};                 ///< Support for VK_EXT_index_type_uint8.
     bool ext_sampler_filter_minmax{};            ///< Support for VK_EXT_sampler_filter_minmax.
     bool ext_depth_range_unrestricted{};         ///< Support for VK_EXT_depth_range_unrestricted.

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -103,6 +103,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdFillBuffer);
     X(vkCmdPipelineBarrier);
     X(vkCmdPushConstants);
+    X(vkCmdPushDescriptorSetWithTemplateKHR);
     X(vkCmdSetBlendConstants);
     X(vkCmdSetDepthBias);
     X(vkCmdSetDepthBounds);

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -193,15 +193,16 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkBeginCommandBuffer vkBeginCommandBuffer{};
     PFN_vkBindBufferMemory vkBindBufferMemory{};
     PFN_vkBindImageMemory vkBindImageMemory{};
+    PFN_vkCmdBeginDebugUtilsLabelEXT vkCmdBeginDebugUtilsLabelEXT{};
     PFN_vkCmdBeginQuery vkCmdBeginQuery{};
     PFN_vkCmdBeginRenderPass vkCmdBeginRenderPass{};
     PFN_vkCmdBeginTransformFeedbackEXT vkCmdBeginTransformFeedbackEXT{};
-    PFN_vkCmdBeginDebugUtilsLabelEXT vkCmdBeginDebugUtilsLabelEXT{};
     PFN_vkCmdBindDescriptorSets vkCmdBindDescriptorSets{};
     PFN_vkCmdBindIndexBuffer vkCmdBindIndexBuffer{};
     PFN_vkCmdBindPipeline vkCmdBindPipeline{};
     PFN_vkCmdBindTransformFeedbackBuffersEXT vkCmdBindTransformFeedbackBuffersEXT{};
     PFN_vkCmdBindVertexBuffers vkCmdBindVertexBuffers{};
+    PFN_vkCmdBindVertexBuffers2EXT vkCmdBindVertexBuffers2EXT{};
     PFN_vkCmdBlitImage vkCmdBlitImage{};
     PFN_vkCmdClearAttachments vkCmdClearAttachments{};
     PFN_vkCmdCopyBuffer vkCmdCopyBuffer{};
@@ -211,35 +212,35 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdDispatch vkCmdDispatch{};
     PFN_vkCmdDraw vkCmdDraw{};
     PFN_vkCmdDrawIndexed vkCmdDrawIndexed{};
+    PFN_vkCmdEndDebugUtilsLabelEXT vkCmdEndDebugUtilsLabelEXT{};
     PFN_vkCmdEndQuery vkCmdEndQuery{};
     PFN_vkCmdEndRenderPass vkCmdEndRenderPass{};
     PFN_vkCmdEndTransformFeedbackEXT vkCmdEndTransformFeedbackEXT{};
-    PFN_vkCmdEndDebugUtilsLabelEXT vkCmdEndDebugUtilsLabelEXT{};
     PFN_vkCmdFillBuffer vkCmdFillBuffer{};
     PFN_vkCmdPipelineBarrier vkCmdPipelineBarrier{};
     PFN_vkCmdPushConstants vkCmdPushConstants{};
+    PFN_vkCmdPushDescriptorSetWithTemplateKHR vkCmdPushDescriptorSetWithTemplateKHR{};
+    PFN_vkCmdResolveImage vkCmdResolveImage{};
     PFN_vkCmdSetBlendConstants vkCmdSetBlendConstants{};
+    PFN_vkCmdSetCullModeEXT vkCmdSetCullModeEXT{};
     PFN_vkCmdSetDepthBias vkCmdSetDepthBias{};
     PFN_vkCmdSetDepthBounds vkCmdSetDepthBounds{};
-    PFN_vkCmdSetEvent vkCmdSetEvent{};
-    PFN_vkCmdSetScissor vkCmdSetScissor{};
-    PFN_vkCmdSetStencilCompareMask vkCmdSetStencilCompareMask{};
-    PFN_vkCmdSetStencilReference vkCmdSetStencilReference{};
-    PFN_vkCmdSetStencilWriteMask vkCmdSetStencilWriteMask{};
-    PFN_vkCmdSetViewport vkCmdSetViewport{};
-    PFN_vkCmdWaitEvents vkCmdWaitEvents{};
-    PFN_vkCmdBindVertexBuffers2EXT vkCmdBindVertexBuffers2EXT{};
-    PFN_vkCmdSetCullModeEXT vkCmdSetCullModeEXT{};
     PFN_vkCmdSetDepthBoundsTestEnableEXT vkCmdSetDepthBoundsTestEnableEXT{};
     PFN_vkCmdSetDepthCompareOpEXT vkCmdSetDepthCompareOpEXT{};
     PFN_vkCmdSetDepthTestEnableEXT vkCmdSetDepthTestEnableEXT{};
     PFN_vkCmdSetDepthWriteEnableEXT vkCmdSetDepthWriteEnableEXT{};
+    PFN_vkCmdSetEvent vkCmdSetEvent{};
     PFN_vkCmdSetFrontFaceEXT vkCmdSetFrontFaceEXT{};
     PFN_vkCmdSetPrimitiveTopologyEXT vkCmdSetPrimitiveTopologyEXT{};
+    PFN_vkCmdSetScissor vkCmdSetScissor{};
+    PFN_vkCmdSetStencilCompareMask vkCmdSetStencilCompareMask{};
     PFN_vkCmdSetStencilOpEXT vkCmdSetStencilOpEXT{};
+    PFN_vkCmdSetStencilReference vkCmdSetStencilReference{};
     PFN_vkCmdSetStencilTestEnableEXT vkCmdSetStencilTestEnableEXT{};
+    PFN_vkCmdSetStencilWriteMask vkCmdSetStencilWriteMask{};
     PFN_vkCmdSetVertexInputEXT vkCmdSetVertexInputEXT{};
-    PFN_vkCmdResolveImage vkCmdResolveImage{};
+    PFN_vkCmdSetViewport vkCmdSetViewport{};
+    PFN_vkCmdWaitEvents vkCmdWaitEvents{};
     PFN_vkCreateBuffer vkCreateBuffer{};
     PFN_vkCreateBufferView vkCreateBufferView{};
     PFN_vkCreateCommandPool vkCreateCommandPool{};
@@ -990,6 +991,12 @@ public:
                                      dynamic_offsets.size(), dynamic_offsets.data());
     }
 
+    void PushDescriptorSetWithTemplateKHR(VkDescriptorUpdateTemplateKHR update_template,
+                                          VkPipelineLayout layout, u32 set,
+                                          const void* data) const noexcept {
+        dld->vkCmdPushDescriptorSetWithTemplateKHR(handle, update_template, layout, set, data);
+    }
+
     void BindPipeline(VkPipelineBindPoint bind_point, VkPipeline pipeline) const noexcept {
         dld->vkCmdBindPipeline(handle, bind_point, pipeline);
     }