GS:HW: Add option to spin CPU during readbacks

common/D3D12/Context.cpp

@@ -17,6 +17,7 @@
 
 #include "common/D3D12/Context.h"
 #include "common/Assertions.h"
+#include "common/General.h"
 #include "common/ScopedGuard.h"
 #include "common/Console.h"
 #include "D3D12MemAlloc.h"
@@ -382,7 +383,7 @@ void Context::MoveToNextCommandList()
 
 	// We may have to wait if this command list hasn't finished on the GPU.
 	CommandListResources& res = m_command_lists[m_current_command_list];
-	WaitForFence(res.ready_fence_value);
+	WaitForFence(res.ready_fence_value, false);
 	res.ready_fence_value = m_current_fence_value;
 	res.init_command_list_used = false;
 
@@ -445,7 +446,7 @@ ID3D12GraphicsCommandList4* Context::GetInitCommandList()
 	return res.command_lists[0].get();
 }
 
-void Context::ExecuteCommandList(bool wait_for_completion)
+void Context::ExecuteCommandList(WaitType wait_for_completion)
 {
 	CommandListResources& res = m_command_lists[m_current_command_list];
 	HRESULT hr;
@@ -485,8 +486,8 @@ void Context::ExecuteCommandList(bool wait_for_completion)
 	pxAssertRel(SUCCEEDED(hr), "Signal fence");
 
 	MoveToNextCommandList();
-	if (wait_for_completion)
-		WaitForFence(res.ready_fence_value);
+	if (wait_for_completion != WaitType::None)
+		WaitForFence(res.ready_fence_value, wait_for_completion == WaitType::Spin);
 }
 
 void Context::InvalidateSamplerGroups()
@@ -547,7 +548,7 @@ void Context::DestroyPendingResources(CommandListResources& cmdlist)
 
 void Context::DestroyResources()
 {
-	ExecuteCommandList(true);
+	ExecuteCommandList(WaitType::Sleep);
 
 	m_texture_stream_buffer.Destroy(false);
 	m_descriptor_heap_manager.Free(&m_null_srv_descriptor);
@@ -573,11 +574,20 @@ void Context::DestroyResources()
 	m_device.reset();
 }
 
-void Context::WaitForFence(u64 fence)
+void Context::WaitForFence(u64 fence, bool spin)
 {
 	if (m_completed_fence_value >= fence)
 		return;
 
+	if (spin)
+	{
+		u64 value;
+		while ((value = m_fence->GetCompletedValue()) < fence)
+			ShortSpin();
+		m_completed_fence_value = value;
+	}
+	else
+	{
 		// Try non-blocking check.
 		m_completed_fence_value = m_fence->GetCompletedValue();
 		if (m_completed_fence_value < fence)
@@ -588,6 +598,7 @@ void Context::WaitForFence(u64 fence)
 			WaitForSingleObject(m_fence_event, INFINITE);
 			m_completed_fence_value = m_fence->GetCompletedValue();
 		}
+	}
 
 	// Release resources for as many command lists which have completed.
 	u32 index = (m_current_command_list + 1) % NUM_COMMAND_LISTS;
@@ -607,7 +618,7 @@ void Context::WaitForGPUIdle()
 	u32 index = (m_current_command_list + 1) % NUM_COMMAND_LISTS;
 	for (u32 i = 0; i < (NUM_COMMAND_LISTS - 1); i++)
 	{
-		WaitForFence(m_command_lists[index].ready_fence_value);
+		WaitForFence(m_command_lists[index].ready_fence_value, false);
 		index = (index + 1) % NUM_COMMAND_LISTS;
 	}
 }

common/D3D12/Context.h

@@ -122,11 +122,18 @@ namespace D3D12
 		/// Test for support for the specified texture format.
 		bool SupportsTextureFormat(DXGI_FORMAT format);
 
+		enum class WaitType
+		{
+			None,  ///< Don't wait (async)
+			Sleep, ///< Wait normally
+			Spin,  ///< Wait by spinning
+		};
+
 		/// Executes the current command list.
-		void ExecuteCommandList(bool wait_for_completion);
+		void ExecuteCommandList(WaitType wait_for_completion);
 
 		/// Waits for a specific fence.
-		void WaitForFence(u64 fence);
+		void WaitForFence(u64 fence, bool spin);
 
 		/// Waits for any in-flight command buffers to complete.
 		void WaitForGPUIdle();

common/D3D12/StreamBuffer.cpp

@@ -273,7 +273,7 @@ bool StreamBuffer::WaitForClearSpace(u32 num_bytes)
 		return false;
 
 	// Wait until this fence is signaled. This will fire the callback, updating the GPU position.
-	g_d3d12_context->WaitForFence(iter->first);
+	g_d3d12_context->WaitForFence(iter->first, false);
 	m_tracked_fences.erase(m_tracked_fences.begin(), m_current_offset == iter->second ? m_tracked_fences.end() : ++iter);
 	m_current_offset = new_offset;
 	m_current_space = new_space;

common/D3D12/Texture.cpp

@@ -293,7 +293,7 @@ ID3D12GraphicsCommandList* Texture::BeginStreamUpdate(ID3D12GraphicsCommandList*
 	{
 		DevCon.WriteLn("Executing command buffer while waiting for %u bytes (%ux%u) in upload buffer", upload_size, width,
 			height);
-		g_d3d12_context->ExecuteCommandList(false);
+		g_d3d12_context->ExecuteCommandList(Context::WaitType::None);
 		if (!g_d3d12_context->GetTextureStreamBuffer().ReserveMemory(upload_size, D3D12_TEXTURE_DATA_PLACEMENT_ALIGNMENT))
 		{
 			Console.Error("Failed to reserve %u bytes for %ux%u upload", upload_size, width, height);

common/Vulkan/Context.cpp

@@ -17,6 +17,7 @@
 #include "common/Align.h"
 #include "common/Assertions.h"
 #include "common/Console.h"
+#include "common/General.h"
 #include "common/StringUtil.h"
 #include "common/Vulkan/ShaderCompiler.h"
 #include "common/Vulkan/SwapChain.h"
@@ -1437,18 +1438,23 @@ namespace Vulkan
 		vmaSetCurrentFrameIndex(m_allocator, static_cast<u32>(m_next_fence_counter));
 	}
 
-	void Context::ExecuteCommandBuffer(bool wait_for_completion)
+	void Context::ExecuteCommandBuffer(WaitType wait_for_completion)
 	{
 		// If we're waiting for completion, don't bother waking the worker thread.
 		const u32 current_frame = m_current_frame;
 		SubmitCommandBuffer();
 		MoveToNextCommandBuffer();
 
-		if (wait_for_completion)
+		if (wait_for_completion != WaitType::None)
 		{
 			// Calibrate while we wait
 			if (m_wants_new_timestamp_calibration)
 				CalibrateSpinTimestamp();
+			if (wait_for_completion == WaitType::Spin)
+			{
+				while (vkGetFenceStatus(m_device, m_frame_resources[current_frame].fence) == VK_NOT_READY)
+					ShortSpin();
+			}
 			WaitForCommandBufferCompletion(current_frame);
 		}
 	}

common/Vulkan/Context.h

@@ -198,7 +198,14 @@ namespace Vulkan
 			uint32_t present_image_index = 0xFFFFFFFF, bool submit_on_thread = false);
 		void MoveToNextCommandBuffer();
 
-		void ExecuteCommandBuffer(bool wait_for_completion);
+		enum class WaitType
+		{
+			None,
+			Sleep,
+			Spin,
+		};
+
+		void ExecuteCommandBuffer(WaitType wait_for_completion);
 		void WaitForPresentComplete();
 
 		// Was the last present submitted to the queue a failure? If so, we must recreate our swapchain.

pcsx2-qt/Settings/GraphicsSettingsWidget.cpp

@@ -99,6 +99,7 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsDialog* dialog, QWidget*
 	SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.vsync, "EmuCore/GS", "VsyncEnable", 0);
 	SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.enableHWFixes, "EmuCore/GS", "UserHacks", false);
 	SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.spinGPUDuringReadbacks, "EmuCore/GS", "HWSpinGPUForReadbacks", false);
+	SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.spinCPUDuringReadbacks, "EmuCore/GS", "HWSpinCPUForReadbacks", false);
 
 	//////////////////////////////////////////////////////////////////////////
 	// Game Display Settings
@@ -415,6 +416,10 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsDialog* dialog, QWidget*
 			tr("Submits useless work to the GPU during readbacks to prevent it from going into powersave modes. "
 			   "May improve performance but with a significant increase in power usage."));
 
+		dialog->registerWidgetHelp(m_ui.spinGPUDuringReadbacks, tr("Spin CPU During Readbacks"), tr("Unchecked"),
+			tr("Does useless work on the CPU during readbacks to prevent it from going to into powersave modes. "
+			   "May improve performance but with a significant increase in power usage."));
+
 		// Software
 		dialog->registerWidgetHelp(m_ui.extraSWThreads, tr("Extra Rendering Threads"), tr("2 threads"),
 			tr("Number of rendering threads: 0 for single thread, 2 or more for multithread (1 is for debugging). "

pcsx2-qt/Settings/GraphicsSettingsWidget.ui

@@ -650,6 +650,13 @@
            </property>
           </widget>
          </item>
+         <item row="1" column="1">
+          <widget class="QCheckBox" name="spinCPUDuringReadbacks">
+           <property name="text">
+            <string>Spin CPU During Readbacks</string>
+           </property>
+          </widget>
+         </item>
         </layout>
        </item>
       </layout>

pcsx2/Config.h

@@ -497,6 +497,7 @@ struct Pcsx2Config
 
 				bool
 					HWSpinGPUForReadbacks : 1,
+					HWSpinCPUForReadbacks : 1,
 					GPUPaletteConversion : 1,
 					AutoFlushSW : 1,
 					PreloadFrameWithGSData : 1,

pcsx2/Frontend/D3D12HostDisplay.cpp

@@ -333,7 +333,7 @@ bool D3D12HostDisplay::ChangeWindow(const WindowInfo& new_wi)
 void D3D12HostDisplay::DestroySurface()
 {
 	// For some reason if we don't execute the command list here, the swap chain is in use.. not sure where.
-	g_d3d12_context->ExecuteCommandList(true);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::Sleep);
 
 	if (IsFullscreen())
 		SetFullscreen(false, 0, 0, 0.0f);
@@ -438,7 +438,7 @@ void D3D12HostDisplay::ResizeWindow(s32 new_window_width, s32 new_window_height,
 		return;
 
 	// For some reason if we don't execute the command list here, the swap chain is in use.. not sure where.
-	g_d3d12_context->ExecuteCommandList(true);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::Sleep);
 
 	DestroySwapChainRTVs();
 
@@ -509,7 +509,7 @@ bool D3D12HostDisplay::SetFullscreen(bool fullscreen, u32 width, u32 height, flo
 		return true;
 	}
 
-	g_d3d12_context->ExecuteCommandList(true);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::Sleep);
 	DestroySwapChainRTVs();
 	m_swap_chain.reset();
 
@@ -585,7 +585,7 @@ void D3D12HostDisplay::EndPresent()
 	m_current_swap_chain_buffer = ((m_current_swap_chain_buffer + 1) % static_cast<u32>(m_swap_chain_buffers.size()));
 
 	swap_chain_buf.TransitionToState(g_d3d12_context->GetCommandList(), D3D12_RESOURCE_STATE_PRESENT);
-	g_d3d12_context->ExecuteCommandList(false);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::None);
 
 	const bool vsync = static_cast<UINT>(m_vsync_mode != VsyncMode::Off);
 	if (!vsync && m_using_allow_tearing)

pcsx2/Frontend/VulkanHostDisplay.cpp

@@ -86,7 +86,7 @@ bool VulkanHostDisplay::ChangeWindow(const WindowInfo& new_wi)
 
 	if (new_wi.type == WindowInfo::Type::Surfaceless)
 	{
-		g_vulkan_context->ExecuteCommandBuffer(true);
+		g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::Sleep);
 		m_swap_chain.reset();
 		m_window_info = new_wi;
 		return true;
@@ -209,7 +209,7 @@ static bool UploadBufferToTexture(
 	if (!buf.ReserveMemory(upload_size, g_vulkan_context->GetBufferCopyOffsetAlignment()))
 	{
 		Console.WriteLn("Executing command buffer for UploadBufferToTexture()");
-		g_vulkan_context->ExecuteCommandBuffer(false);
+		g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::None);
 		if (!buf.ReserveMemory(upload_size, g_vulkan_context->GetBufferCopyOffsetAlignment()))
 		{
 			Console.WriteLn("Failed to allocate %u bytes in stream buffer for UploadBufferToTexture()", upload_size);
@@ -365,7 +365,7 @@ bool VulkanHostDisplay::BeginPresent(bool frame_skip)
 			if (!m_swap_chain->RecreateSurface(m_window_info))
 			{
 				Console.Error("Failed to recreate surface after loss");
-				g_vulkan_context->ExecuteCommandBuffer(false);
+				g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::None);
 				return false;
 			}
 
@@ -378,7 +378,7 @@ bool VulkanHostDisplay::BeginPresent(bool frame_skip)
 		{
 			// Still submit the command buffer, otherwise we'll end up with several frames waiting.
 			LOG_VULKAN_ERROR(res, "vkAcquireNextImageKHR() failed: ");
-			g_vulkan_context->ExecuteCommandBuffer(false);
+			g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::None);
 			return false;
 		}
 	}

pcsx2/Frontend/imgui_impl_dx12.cpp

@@ -165,7 +165,7 @@ void ImGui_ImplDX12_RenderDrawData(ImDrawData* draw_data)
     if (!bd->VertexStreamBuffer.ReserveMemory(needed_vb, sizeof(ImDrawVert)) ||
       !bd->IndexStreamBuffer.ReserveMemory(needed_ib, sizeof(ImDrawIdx)))
     {
-        g_d3d12_context->ExecuteCommandList(false);
+        g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::None);
         if (!bd->VertexStreamBuffer.ReserveMemory(needed_vb, sizeof(ImDrawVert)) ||
           !bd->IndexStreamBuffer.ReserveMemory(needed_ib, sizeof(ImDrawIdx)))
         {
@@ -230,7 +230,7 @@ void ImGui_ImplDX12_RenderDrawData(ImDrawData* draw_data)
                     if (!g_d3d12_context->GetDescriptorAllocator().Allocate(1, &handle))
                     {
                         // ugh.
-                        g_d3d12_context->ExecuteCommandList(false);
+                        g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::None);
                         ctx = g_d3d12_context->GetCommandList();
                         ImGui_ImplDX12_SetupRenderState(draw_data, ctx);
                         if (!g_d3d12_context->GetDescriptorAllocator().Allocate(1, &handle))

pcsx2/GS/GS.cpp

@@ -1446,6 +1446,7 @@ void GSApp::Init()
 	m_default_configuration["HWDownloadMode"]                             = std::to_string(static_cast<u8>(GSHardwareDownloadMode::Enabled));
 	m_default_configuration["GSDumpCompression"]                          = std::to_string(static_cast<u8>(GSDumpCompressionMethod::LZMA));
 	m_default_configuration["HWSpinGPUForReadbacks"]                      = "0";
+	m_default_configuration["HWSpinCPUForReadbacks"]                      = "0";
 	m_default_configuration["pcrtc_antiblur"]                             = "1";
 	m_default_configuration["disable_interlace_offset"]                   = "0";
 	m_default_configuration["pcrtc_offsets"]                              = "0";

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -1445,7 +1445,7 @@ void GSDevice12::DestroyStagingBuffer()
 
 void GSDevice12::DestroyResources()
 {
-	g_d3d12_context->ExecuteCommandList(true);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::Sleep);
 
 	for (auto& it : m_tfx_pipelines)
 		g_d3d12_context->DeferObjectDestruction(it.second.get());
@@ -1734,10 +1734,20 @@ void GSDevice12::InitializeSamplers()
 		pxFailRel("Failed to initialize samplers");
 }
 
+static D3D12::Context::WaitType GetWaitType(bool wait, bool spin)
+{
+	if (!wait)
+		return D3D12::Context::WaitType::None;
+	if (spin)
+		return D3D12::Context::WaitType::Spin;
+	else
+		return D3D12::Context::WaitType::Sleep;
+}
+
 void GSDevice12::ExecuteCommandList(bool wait_for_completion)
 {
 	EndRenderPass();
-	g_d3d12_context->ExecuteCommandList(wait_for_completion);
+	g_d3d12_context->ExecuteCommandList(GetWaitType(wait_for_completion, GSConfig.HWSpinCPUForReadbacks));
 	InvalidateCachedState();
 }
 
@@ -1758,7 +1768,7 @@ void GSDevice12::ExecuteCommandListAndRestartRenderPass(const char* reason)
 
 	const bool was_in_render_pass = m_in_render_pass;
 	EndRenderPass();
-	g_d3d12_context->ExecuteCommandList(false);
+	g_d3d12_context->ExecuteCommandList(D3D12::Context::WaitType::None);
 	InvalidateCachedState();
 
 	if (was_in_render_pass)

pcsx2/GS/Renderers/Metal/GSDeviceMTL.mm

@@ -26,6 +26,21 @@
 #ifdef __APPLE__
 #include "GSMTLSharedHeader.h"
 
+static constexpr bool IsCommandBufferCompleted(MTLCommandBufferStatus status)
+{
+	switch (status)
+	{
+		case MTLCommandBufferStatusNotEnqueued:
+		case MTLCommandBufferStatusEnqueued:
+		case MTLCommandBufferStatusCommitted:
+		case MTLCommandBufferStatusScheduled:
+			return false;
+		case MTLCommandBufferStatusCompleted:
+		case MTLCommandBufferStatusError:
+			return true;
+	}
+}
+
 GSDevice* MakeGSDeviceMTL()
 {
 	return new GSDeviceMTL();
@@ -1055,7 +1070,15 @@ bool GSDeviceMTL::DownloadTexture(GSTexture* src, const GSVector4i& rect, GSText
 			m_spin_timer = 30;
 		}
 	}
+	if (GSConfig.HWSpinCPUForReadbacks)
+	{
+		while (!IsCommandBufferCompleted([cmdbuf status]))
+			ShortSpin();
+	}
+	else
+	{
 		[cmdbuf waitUntilCompleted];
+	}
 
 	out_map.bits = static_cast<u8*>([m_texture_download_buf contents]);
 	g_perfmon.Put(GSPerfMon::Readbacks, 1);

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp

@@ -1822,7 +1822,7 @@ void GSDeviceVK::DestroyStagingBuffer()
 
 void GSDeviceVK::DestroyResources()
 {
-	g_vulkan_context->ExecuteCommandBuffer(true);
+	g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::Sleep);
 	if (m_tfx_descriptor_sets[0] != VK_NULL_HANDLE)
 		g_vulkan_context->FreeGlobalDescriptorSet(m_tfx_descriptor_sets[0]);
 
@@ -2202,10 +2202,20 @@ bool GSDeviceVK::CreatePersistentDescriptorSets()
 	return true;
 }
 
+static Vulkan::Context::WaitType GetWaitType(bool wait, bool spin)
+{
+	if (!wait)
+		return Vulkan::Context::WaitType::None;
+	if (spin)
+		return Vulkan::Context::WaitType::Spin;
+	else
+		return Vulkan::Context::WaitType::Sleep;
+}
+
 void GSDeviceVK::ExecuteCommandBuffer(bool wait_for_completion)
 {
 	EndRenderPass();
-	g_vulkan_context->ExecuteCommandBuffer(wait_for_completion);
+	g_vulkan_context->ExecuteCommandBuffer(GetWaitType(wait_for_completion, GSConfig.HWSpinCPUForReadbacks));
 	InvalidateCachedState();
 }
 
@@ -2227,7 +2237,7 @@ void GSDeviceVK::ExecuteCommandBufferAndRestartRenderPass(const char* reason)
 	const VkRenderPass render_pass = m_current_render_pass;
 	const GSVector4i render_pass_area(m_current_render_pass_area);
 	EndRenderPass();
-	g_vulkan_context->ExecuteCommandBuffer(false);
+	g_vulkan_context->ExecuteCommandBuffer(Vulkan::Context::WaitType::None);
 	InvalidateCachedState();
 
 	if (render_pass != VK_NULL_HANDLE)

pcsx2/GS/Window/GSSetting.cpp

@@ -167,6 +167,9 @@ const char* dialog_message(int ID, bool* updateText)
 		case IDC_SPIN_GPU:
 			return cvtString("Submits useless work to the GPU during readbacks to prevent it from going into powersave modes.\n"
 				"May improve performance but with a significant increase in power usage.");
+		case IDC_SPIN_CPU:
+			return cvtString("Does useless work on the CPU during readbacks to prevent it from going to into powersave modes.\n"
+				"May improve performance but with a significant increase in power usage.");
 		case IDC_LINEAR_PRESENT:
 			return cvtString("Use bilinear filtering when Upscaling/Downscaling the image to the screen. Disable it if you want a sharper/pixelated output.");
 		// Exclusive for Hardware Renderer

pcsx2/GS/Window/GSSetting.h

@@ -88,6 +88,7 @@ enum
 	// OpenGL Advanced Settings
 	IDC_GEOMETRY_SHADER_OVERRIDE,
 	IDC_SPIN_GPU,
+	IDC_SPIN_CPU,
 	// On-screen Display
 	IDC_OSD_LOG,
 	IDC_OSD_MONITOR,

pcsx2/GS/Window/GSwxDialog.cpp

@@ -285,6 +285,7 @@ RendererTab::RendererTab(wxWindow* parent)
 
 	auto* paltex_prereq = m_ui.addCheckBox(hw_checks_box, "GPU Palette Conversion", "paltex", IDC_PALTEX, hw_prereq);
 	m_ui.addCheckBox(hw_checks_box, "Spin GPU During Readbacks", "HWSpinGPUForReadbacks", IDC_SPIN_GPU);
+	m_ui.addCheckBox(hw_checks_box, "Spin CPU During Readbacks", "HWSpinCPUForReadbacks", IDC_SPIN_CPU);
 	auto aniso_prereq = [this, paltex_prereq]{ return m_is_hardware && paltex_prereq->GetValue() == false; };
 
 	auto* hw_choice_grid = new wxFlexGridSizer(2, space, space);

pcsx2/Pcsx2Config.cpp

@@ -326,6 +326,7 @@ Pcsx2Config::GSOptions::GSOptions()
 
 	HWDownloadMode = GSHardwareDownloadMode::Enabled;
 	HWSpinGPUForReadbacks = false;
+	HWSpinCPUForReadbacks = false;
 	GPUPaletteConversion = false;
 	AutoFlushSW = true;
 	PreloadFrameWithGSData = false;
@@ -550,6 +551,7 @@ void Pcsx2Config::GSOptions::ReloadIniSettings()
 	GSSettingBool(OsdShowInputs);
 
 	GSSettingBool(HWSpinGPUForReadbacks);
+	GSSettingBool(HWSpinCPUForReadbacks);
 	GSSettingBoolEx(GPUPaletteConversion, "paltex");
 	GSSettingBoolEx(AutoFlushSW, "autoflush_sw");
 	GSSettingBoolEx(PreloadFrameWithGSData, "preload_frame_with_gs_data");