Merge pull request #9552 from liamwhite/turbo

vulkan: implement 'turbo mode' clock booster

src/common/settings.cpp

@@ -185,6 +185,7 @@ void RestoreGlobalState(bool is_powered_on) {
     // Renderer
     values.fsr_sharpening_slider.SetGlobal(true);
     values.renderer_backend.SetGlobal(true);
+    values.renderer_force_max_clock.SetGlobal(true);
     values.vulkan_device.SetGlobal(true);
     values.aspect_ratio.SetGlobal(true);
     values.max_anisotropy.SetGlobal(true);

src/common/settings.h

@@ -415,6 +415,7 @@ struct Values {
     // Renderer
     SwitchableSetting<RendererBackend, true> renderer_backend{
         RendererBackend::Vulkan, RendererBackend::OpenGL, RendererBackend::Null, "backend"};
+    SwitchableSetting<bool> renderer_force_max_clock{true, "force_max_clock"};
     Setting<bool> renderer_debug{false, "debug"};
     Setting<bool> renderer_shader_feedback{false, "shader_feedback"};
     Setting<bool> enable_nsight_aftermath{false, "nsight_aftermath"};

src/video_core/CMakeLists.txt

@@ -191,6 +191,8 @@ add_library(video_core STATIC
     renderer_vulkan/vk_texture_cache.cpp
     renderer_vulkan/vk_texture_cache.h
     renderer_vulkan/vk_texture_cache_base.cpp
+    renderer_vulkan/vk_turbo_mode.cpp
+    renderer_vulkan/vk_turbo_mode.h
     renderer_vulkan/vk_update_descriptor.cpp
     renderer_vulkan/vk_update_descriptor.h
     shader_cache.cpp

src/video_core/host_shaders/CMakeLists.txt

@@ -47,6 +47,7 @@ set(SHADER_FILES
     vulkan_present_scaleforce_fp16.frag
     vulkan_present_scaleforce_fp32.frag
     vulkan_quad_indexed.comp
+    vulkan_turbo_mode.comp
     vulkan_uint8.comp
 )
 

src/video_core/host_shaders/vulkan_turbo_mode.comp

@@ -0,0 +1,29 @@
+// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 460 core
+
+layout (local_size_x = 16, local_size_y = 8, local_size_z = 1) in;
+
+layout (binding = 0) buffer ThreadData {
+    uint data[];
+};
+
+uint xorshift32(uint x) {
+    x ^= x << 13;
+    x ^= x >> 17;
+    x ^= x << 5;
+    return x;
+}
+
+uint getGlobalIndex() {
+    return gl_GlobalInvocationID.x + gl_GlobalInvocationID.y * gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+}
+
+void main() {
+    uint myIndex = xorshift32(getGlobalIndex());
+    uint otherIndex = xorshift32(myIndex);
+
+    uint otherValue = atomicAdd(data[otherIndex % data.length()], 0) + 1;
+    atomicAdd(data[myIndex % data.length()], otherValue);
+}

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -78,6 +78,8 @@ std::string BuildCommaSeparatedExtensions(std::vector<std::string> available_ext
     return separated_extensions;
 }
 
+} // Anonymous namespace
+
 Device CreateDevice(const vk::Instance& instance, const vk::InstanceDispatch& dld,
                     VkSurfaceKHR surface) {
     const std::vector<VkPhysicalDevice> devices = instance.EnumeratePhysicalDevices();
@@ -89,7 +91,6 @@ Device CreateDevice(const vk::Instance& instance, const vk::InstanceDispatch& dl
     const vk::PhysicalDevice physical_device(devices[device_index], dld);
     return Device(*instance, physical_device, surface, dld);
 }
-} // Anonymous namespace
 
 RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
                                Core::Frontend::EmuWindow& emu_window,
@@ -109,6 +110,9 @@ RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
                   screen_info),
       rasterizer(render_window, gpu, cpu_memory, screen_info, device, memory_allocator,
                  state_tracker, scheduler) {
+    if (Settings::values.renderer_force_max_clock.GetValue()) {
+        turbo_mode.emplace(instance, dld);
+    }
     Report();
 } catch (const vk::Exception& exception) {
     LOG_ERROR(Render_Vulkan, "Vulkan initialization failed with error: {}", exception.what());

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -13,6 +13,7 @@
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_state_tracker.h"
 #include "video_core/renderer_vulkan/vk_swapchain.h"
+#include "video_core/renderer_vulkan/vk_turbo_mode.h"
 #include "video_core/vulkan_common/vulkan_device.h"
 #include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -31,6 +32,9 @@ class GPU;
 
 namespace Vulkan {
 
+Device CreateDevice(const vk::Instance& instance, const vk::InstanceDispatch& dld,
+                    VkSurfaceKHR surface);
+
 class RendererVulkan final : public VideoCore::RendererBase {
 public:
     explicit RendererVulkan(Core::TelemetrySession& telemtry_session,
@@ -74,6 +78,7 @@ private:
     Swapchain swapchain;
     BlitScreen blit_screen;
     RasterizerVulkan rasterizer;
+    std::optional<TurboMode> turbo_mode;
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_turbo_mode.cpp

@@ -0,0 +1,205 @@
+// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "common/literals.h"
+#include "video_core/host_shaders/vulkan_turbo_mode_comp_spv.h"
+#include "video_core/renderer_vulkan/renderer_vulkan.h"
+#include "video_core/renderer_vulkan/vk_shader_util.h"
+#include "video_core/renderer_vulkan/vk_turbo_mode.h"
+#include "video_core/vulkan_common/vulkan_device.h"
+
+namespace Vulkan {
+
+using namespace Common::Literals;
+
+TurboMode::TurboMode(const vk::Instance& instance, const vk::InstanceDispatch& dld)
+    : m_device{CreateDevice(instance, dld, VK_NULL_HANDLE)}, m_allocator{m_device, false} {
+    m_thread = std::jthread([&](auto stop_token) { Run(stop_token); });
+}
+
+TurboMode::~TurboMode() = default;
+
+void TurboMode::Run(std::stop_token stop_token) {
+    auto& dld = m_device.GetLogical();
+
+    // Allocate buffer. 2MiB should be sufficient.
+    auto buffer = dld.CreateBuffer(VkBufferCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .size = 2_MiB,
+        .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    });
+
+    // Commit some device local memory for the buffer.
+    auto commit = m_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
+
+    // Create the descriptor pool to contain our descriptor.
+    constexpr VkDescriptorPoolSize pool_size{
+        .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .descriptorCount = 1,
+    };
+
+    auto descriptor_pool = dld.CreateDescriptorPool(VkDescriptorPoolCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
+        .maxSets = 1,
+        .poolSizeCount = 1,
+        .pPoolSizes = &pool_size,
+    });
+
+    // Create the descriptor set layout from the pool.
+    constexpr VkDescriptorSetLayoutBinding layout_binding{
+        .binding = 0,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    };
+
+    auto descriptor_set_layout = dld.CreateDescriptorSetLayout(VkDescriptorSetLayoutCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .bindingCount = 1,
+        .pBindings = &layout_binding,
+    });
+
+    // Actually create the descriptor set.
+    auto descriptor_set = descriptor_pool.Allocate(VkDescriptorSetAllocateInfo{
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
+        .pNext = nullptr,
+        .descriptorPool = *descriptor_pool,
+        .descriptorSetCount = 1,
+        .pSetLayouts = descriptor_set_layout.address(),
+    });
+
+    // Create the shader.
+    auto shader = BuildShader(m_device, VULKAN_TURBO_MODE_COMP_SPV);
+
+    // Create the pipeline layout.
+    auto pipeline_layout = dld.CreatePipelineLayout(VkPipelineLayoutCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .setLayoutCount = 1,
+        .pSetLayouts = descriptor_set_layout.address(),
+        .pushConstantRangeCount = 0,
+        .pPushConstantRanges = nullptr,
+    });
+
+    // Actually create the pipeline.
+    const VkPipelineShaderStageCreateInfo shader_stage{
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+        .module = *shader,
+        .pName = "main",
+        .pSpecializationInfo = nullptr,
+    };
+
+    auto pipeline = dld.CreateComputePipeline(VkComputePipelineCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .stage = shader_stage,
+        .layout = *pipeline_layout,
+        .basePipelineHandle = VK_NULL_HANDLE,
+        .basePipelineIndex = 0,
+    });
+
+    // Create a fence to wait on.
+    auto fence = dld.CreateFence(VkFenceCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+    });
+
+    // Create a command pool to allocate a command buffer from.
+    auto command_pool = dld.CreateCommandPool(VkCommandPoolCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
+        .pNext = nullptr,
+        .flags =
+            VK_COMMAND_POOL_CREATE_TRANSIENT_BIT | VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
+        .queueFamilyIndex = m_device.GetGraphicsFamily(),
+    });
+
+    // Create a single command buffer.
+    auto cmdbufs = command_pool.Allocate(1, VK_COMMAND_BUFFER_LEVEL_PRIMARY);
+    auto cmdbuf = vk::CommandBuffer{cmdbufs[0], m_device.GetDispatchLoader()};
+
+    while (!stop_token.stop_requested()) {
+        // Reset the fence.
+        fence.Reset();
+
+        // Update descriptor set.
+        const VkDescriptorBufferInfo buffer_info{
+            .buffer = *buffer,
+            .offset = 0,
+            .range = VK_WHOLE_SIZE,
+        };
+
+        const VkWriteDescriptorSet buffer_write{
+            .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+            .pNext = nullptr,
+            .dstSet = descriptor_set[0],
+            .dstBinding = 0,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+            .pImageInfo = nullptr,
+            .pBufferInfo = &buffer_info,
+            .pTexelBufferView = nullptr,
+        };
+
+        dld.UpdateDescriptorSets(std::array{buffer_write}, {});
+
+        // Set up the command buffer.
+        cmdbuf.Begin(VkCommandBufferBeginInfo{
+            .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+            .pNext = nullptr,
+            .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+            .pInheritanceInfo = nullptr,
+        });
+
+        // Clear the buffer.
+        cmdbuf.FillBuffer(*buffer, 0, VK_WHOLE_SIZE, 0);
+
+        // Bind descriptor set.
+        cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline_layout, 0,
+                                  descriptor_set, {});
+
+        // Bind the pipeline.
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
+
+        // Dispatch.
+        cmdbuf.Dispatch(64, 64, 1);
+
+        // Finish.
+        cmdbuf.End();
+
+        const VkSubmitInfo submit_info{
+            .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+            .pNext = nullptr,
+            .waitSemaphoreCount = 0,
+            .pWaitSemaphores = nullptr,
+            .pWaitDstStageMask = nullptr,
+            .commandBufferCount = 1,
+            .pCommandBuffers = cmdbuf.address(),
+            .signalSemaphoreCount = 0,
+            .pSignalSemaphores = nullptr,
+        };
+
+        m_device.GetGraphicsQueue().Submit(std::array{submit_info}, *fence);
+
+        // Wait for completion.
+        fence.Wait();
+    }
+}
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_turbo_mode.h

@@ -0,0 +1,26 @@
+// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "common/polyfill_thread.h"
+#include "video_core/vulkan_common/vulkan_device.h"
+#include "video_core/vulkan_common/vulkan_memory_allocator.h"
+#include "video_core/vulkan_common/vulkan_wrapper.h"
+
+namespace Vulkan {
+
+class TurboMode {
+public:
+    explicit TurboMode(const vk::Instance& instance, const vk::InstanceDispatch& dld);
+    ~TurboMode();
+
+private:
+    void Run(std::stop_token stop_token);
+
+    Device m_device;
+    MemoryAllocator m_allocator;
+    std::jthread m_thread;
+};
+
+} // namespace Vulkan

src/video_core/vulkan_common/vulkan_device.cpp

@@ -1472,7 +1472,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
         is_patch_list_restart_supported =
             primitive_topology_list_restart.primitiveTopologyPatchListRestart;
     }
-    if (has_khr_image_format_list && has_khr_swapchain_mutable_format) {
+    if (requires_surface && has_khr_image_format_list && has_khr_swapchain_mutable_format) {
         extensions.push_back(VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
         extensions.push_back(VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME);
         khr_swapchain_mutable_format = true;

src/yuzu/configuration/config.cpp

@@ -690,6 +690,7 @@ void Config::ReadRendererValues() {
     qt_config->beginGroup(QStringLiteral("Renderer"));
 
     ReadGlobalSetting(Settings::values.renderer_backend);
+    ReadGlobalSetting(Settings::values.renderer_force_max_clock);
     ReadGlobalSetting(Settings::values.vulkan_device);
     ReadGlobalSetting(Settings::values.fullscreen_mode);
     ReadGlobalSetting(Settings::values.aspect_ratio);
@@ -1306,6 +1307,9 @@ void Config::SaveRendererValues() {
                  static_cast<u32>(Settings::values.renderer_backend.GetValue(global)),
                  static_cast<u32>(Settings::values.renderer_backend.GetDefault()),
                  Settings::values.renderer_backend.UsingGlobal());
+    WriteSetting(QString::fromStdString(Settings::values.renderer_force_max_clock.GetLabel()),
+                 static_cast<u32>(Settings::values.renderer_force_max_clock.GetValue(global)),
+                 static_cast<u32>(Settings::values.renderer_force_max_clock.GetDefault()));
     WriteGlobalSetting(Settings::values.vulkan_device);
     WriteSetting(QString::fromStdString(Settings::values.fullscreen_mode.GetLabel()),
                  static_cast<u32>(Settings::values.fullscreen_mode.GetValue(global)),

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -25,6 +25,7 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->use_asynchronous_shaders->setEnabled(runtime_lock);
     ui->anisotropic_filtering_combobox->setEnabled(runtime_lock);
 
+    ui->renderer_force_max_clock->setChecked(Settings::values.renderer_force_max_clock.GetValue());
     ui->use_vsync->setChecked(Settings::values.use_vsync.GetValue());
     ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue());
     ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time.GetValue());
@@ -39,6 +40,8 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
             Settings::values.max_anisotropy.GetValue());
     } else {
         ConfigurationShared::SetPerGameSetting(ui->gpu_accuracy, &Settings::values.gpu_accuracy);
+        ConfigurationShared::SetPerGameSetting(ui->renderer_force_max_clock,
+                                               &Settings::values.renderer_force_max_clock);
         ConfigurationShared::SetPerGameSetting(ui->anisotropic_filtering_combobox,
                                                &Settings::values.max_anisotropy);
         ConfigurationShared::SetHighlight(ui->label_gpu_accuracy,
@@ -50,6 +53,9 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
 
 void ConfigureGraphicsAdvanced::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.gpu_accuracy, ui->gpu_accuracy);
+    ConfigurationShared::ApplyPerGameSetting(&Settings::values.renderer_force_max_clock,
+                                             ui->renderer_force_max_clock,
+                                             renderer_force_max_clock);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.max_anisotropy,
                                              ui->anisotropic_filtering_combobox);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_vsync, ui->use_vsync, use_vsync);
@@ -81,6 +87,8 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
     // Disable if not global (only happens during game)
     if (Settings::IsConfiguringGlobal()) {
         ui->gpu_accuracy->setEnabled(Settings::values.gpu_accuracy.UsingGlobal());
+        ui->renderer_force_max_clock->setEnabled(
+            Settings::values.renderer_force_max_clock.UsingGlobal());
         ui->use_vsync->setEnabled(Settings::values.use_vsync.UsingGlobal());
         ui->use_asynchronous_shaders->setEnabled(
             Settings::values.use_asynchronous_shaders.UsingGlobal());
@@ -95,6 +103,9 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
         return;
     }
 
+    ConfigurationShared::SetColoredTristate(ui->renderer_force_max_clock,
+                                            Settings::values.renderer_force_max_clock,
+                                            renderer_force_max_clock);
     ConfigurationShared::SetColoredTristate(ui->use_vsync, Settings::values.use_vsync, use_vsync);
     ConfigurationShared::SetColoredTristate(ui->use_asynchronous_shaders,
                                             Settings::values.use_asynchronous_shaders,

src/yuzu/configuration/configure_graphics_advanced.h

@@ -36,6 +36,7 @@ private:
 
     std::unique_ptr<Ui::ConfigureGraphicsAdvanced> ui;
 
+    ConfigurationShared::CheckState renderer_force_max_clock;
     ConfigurationShared::CheckState use_vsync;
     ConfigurationShared::CheckState use_asynchronous_shaders;
     ConfigurationShared::CheckState use_fast_gpu_time;

src/yuzu/configuration/configure_graphics_advanced.ui

@@ -69,6 +69,16 @@
           </layout>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="renderer_force_max_clock">
+          <property name="toolTip">
+           <string>Runs work in the background while waiting for graphics commands to keep the GPU from lowering its clock speed.</string>
+          </property>
+          <property name="text">
+           <string>Force maximum clocks (Vulkan only)</string>
+          </property>
+         </widget>
+        </item>
         <item>
          <widget class="QCheckBox" name="use_vsync">
           <property name="toolTip">

src/yuzu_cmd/config.cpp

@@ -296,6 +296,7 @@ void Config::ReadValues() {
 
     // Renderer
     ReadSetting("Renderer", Settings::values.renderer_backend);
+    ReadSetting("Renderer", Settings::values.renderer_force_max_clock);
     ReadSetting("Renderer", Settings::values.renderer_debug);
     ReadSetting("Renderer", Settings::values.renderer_shader_feedback);
     ReadSetting("Renderer", Settings::values.enable_nsight_aftermath);