Merge pull request #16196 from hrydgard/improved-render-stats

Improved stats in the Vulkan GPU profiler

CMakeLists.txt

@@ -650,6 +650,8 @@ add_library(Common STATIC
 	Common/Math/lin/vec3.h
 	Common/Math/math_util.cpp
 	Common/Math/math_util.h
+	Common/Math/Statistics.h
+	Common/Math/Statistics.cpp
 	Common/Net/HTTPClient.cpp
 	Common/Net/HTTPClient.h
 	Common/Net/HTTPHeaders.cpp

Common/Common.vcxproj

@@ -466,6 +466,7 @@
     <ClInclude Include="Math\lin\matrix4x4.h" />
     <ClInclude Include="Math\lin\vec3.h" />
     <ClInclude Include="Math\math_util.h" />
+    <ClInclude Include="Math\Statistics.h" />
     <ClInclude Include="Net\NetBuffer.h" />
     <ClInclude Include="Net\HTTPClient.h" />
     <ClInclude Include="Net\HTTPHeaders.h" />
@@ -886,6 +887,7 @@
     <ClCompile Include="Math\lin\matrix4x4.cpp" />
     <ClCompile Include="Math\lin\vec3.cpp" />
     <ClCompile Include="Math\math_util.cpp" />
+    <ClCompile Include="Math\Statistics.cpp" />
     <ClCompile Include="Net\NetBuffer.cpp" />
     <ClCompile Include="Net\HTTPClient.cpp" />
     <ClCompile Include="Net\HTTPHeaders.cpp" />

Common/Common.vcxproj.filters

@@ -422,6 +422,9 @@
     <ClInclude Include="GPU\Vulkan\VulkanFrameData.h">
       <Filter>GPU\Vulkan</Filter>
     </ClInclude>
+    <ClInclude Include="Math\Statistics.h">
+      <Filter>Math</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="ABI.cpp" />
@@ -797,6 +800,9 @@
     <ClCompile Include="GPU\Vulkan\VulkanFrameData.cpp">
       <Filter>GPU\Vulkan</Filter>
     </ClCompile>
+    <ClCompile Include="Math\Statistics.cpp">
+      <Filter>Math</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Crypto">

Common/GPU/Vulkan/VulkanQueueRunner.cpp

@@ -515,6 +515,8 @@ void VulkanQueueRunner::PreprocessSteps(std::vector<VKRStep *> &steps) {
 					}
 					MergeRenderAreaRectInto(&steps[i]->render.renderArea, steps[j]->render.renderArea);
 					steps[i]->render.renderPassType = MergeRPTypes(steps[i]->render.renderPassType, steps[j]->render.renderPassType);
+					steps[i]->render.numDraws += steps[j]->render.numDraws;
+					steps[i]->render.numReads += steps[j]->render.numReads;
 					// Cheaply skip the first step.
 					steps[j]->stepType = VKRStepType::RENDER_SKIP;
 					break;
@@ -936,6 +938,8 @@ void VulkanQueueRunner::ApplyRenderPassMerge(std::vector<VKRStep *> &steps) {
 		// So we don't consider it for other things, maybe doesn't matter.
 		src->dependencies.clear();
 		src->stepType = VKRStepType::RENDER_SKIP;
+		dst->render.numDraws += src->render.numDraws;
+		dst->render.numReads += src->render.numReads;
 		dst->render.pipelineFlags |= src->render.pipelineFlags;
 		dst->render.renderPassType = MergeRPTypes(dst->render.renderPassType, src->render.renderPassType);
 	};

Common/GPU/Vulkan/VulkanRenderManager.cpp

@@ -327,7 +327,12 @@ void CreateImage(VulkanContext *vulkan, VkCommandBuffer cmd, VKRImage &img, int
 	img.tag = tag ? tag : "N/A";
 }
 
-VulkanRenderManager::VulkanRenderManager(VulkanContext *vulkan) : vulkan_(vulkan), queueRunner_(vulkan) {
+VulkanRenderManager::VulkanRenderManager(VulkanContext *vulkan)
+	: vulkan_(vulkan), queueRunner_(vulkan),
+	initTimeMs_("initTimeMs"),
+	totalGPUTimeMs_("totalGPUTimeMs"),
+	renderCPUTimeMs_("renderCPUTimeMs")
+{
 	inflightFramesAtStart_ = vulkan_->GetInflightFrames();
 
 	frameDataShared_.Init(vulkan);
@@ -579,14 +584,25 @@ void VulkanRenderManager::BeginFrame(bool enableProfiling, bool enableLogProfile
 				std::stringstream str;
 
 				char line[256];
-				snprintf(line, sizeof(line), "Total GPU time: %0.3f ms\n", ((double)((queryResults[numQueries - 1] - queryResults[0]) & timestampDiffMask) * timestampConversionFactor));
+				totalGPUTimeMs_.Update(((double)((queryResults[numQueries - 1] - queryResults[0]) & timestampDiffMask) * timestampConversionFactor));
+				totalGPUTimeMs_.Format(line, sizeof(line));
 				str << line;
-				snprintf(line, sizeof(line), "Render CPU time: %0.3f ms\n", (frameData.profile.cpuEndTime - frameData.profile.cpuStartTime) * 1000.0);
+				renderCPUTimeMs_.Update((frameData.profile.cpuEndTime - frameData.profile.cpuStartTime) * 1000.0);
+				renderCPUTimeMs_.Format(line, sizeof(line));
 				str << line;
 				for (int i = 0; i < numQueries - 1; i++) {
 					uint64_t diff = (queryResults[i + 1] - queryResults[i]) & timestampDiffMask;
 					double milliseconds = (double)diff * timestampConversionFactor;
-					snprintf(line, sizeof(line), "%s: %0.3f ms\n", frameData.profile.timestampDescriptions[i + 1].c_str(), milliseconds);
+
+					// Can't use SimpleStat for these very easily since these are dynamic per frame.
+					// Only the first one is static, the initCmd.
+					// Could try some hashtable tracking for the rest, later.
+					if (i == 0) {
+						initTimeMs_.Update(milliseconds);
+						initTimeMs_.Format(line, sizeof(line));
+					} else {
+						snprintf(line, sizeof(line), "%s: %0.3f ms\n", frameData.profile.timestampDescriptions[i + 1].c_str(), milliseconds);
+					}
 					str << line;
 				}
 				frameData.profile.profileSummary = str.str();
@@ -1344,3 +1360,9 @@ void VulkanRenderManager::FlushSync() {
 		frameData.syncDone = false;
 	}
 }
+
+void VulkanRenderManager::ResetStats() {
+	initTimeMs_.Reset();
+	totalGPUTimeMs_.Reset();
+	renderCPUTimeMs_.Reset();
+}

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -12,6 +12,7 @@
 #include <thread>
 #include <queue>
 
+#include "Common/Math/Statistics.h"
 #include "Common/Thread/Promise.h"
 #include "Common/System/Display.h"
 #include "Common/GPU/Vulkan/VulkanContext.h"
@@ -465,6 +466,8 @@ class VulkanRenderManager {
 		return outOfDateFrames_ > VulkanContext::MAX_INFLIGHT_FRAMES;
 	}
 
+	void ResetStats();
+
 private:
 	void EndCurRenderStep();
 
@@ -535,4 +538,9 @@ class VulkanRenderManager {
 
 	// pipelines to check and possibly create at the end of the current render pass.
 	std::vector<VKRGraphicsPipeline *> pipelinesToCheck_;
+
+	// For nicer output in the little internal GPU profiler.
+	SimpleStat initTimeMs_;
+	SimpleStat totalGPUTimeMs_;
+	SimpleStat renderCPUTimeMs_;
 };

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -455,6 +455,10 @@ class VKContext : public DrawContext {
 
 	void FlushState() override {}
 
+	void ResetStats() override {
+		renderManager_.ResetStats();
+	}
+
 	std::string GetInfoString(InfoField info) const override {
 		// TODO: Make these actually query the right information
 		switch (info) {

Common/GPU/thin3d.h

@@ -735,6 +735,9 @@ class DrawContext {
 	// Flush state like scissors etc so the caller can do its own custom drawing.
 	virtual void FlushState() {}
 
+	// This is called when we launch a new game, so any collected internal stats in the backends don't carry over.
+	virtual void ResetStats() {}
+
 	virtual int GetCurrentStepId() const = 0;
 
 protected:

Common/Math/Statistics.cpp

@@ -0,0 +1,11 @@
+#include <cstdio>
+
+#include "Common/Math/Statistics.h"
+
+void SimpleStat::Format(char *buffer, size_t sz) {
+	if (min_ == INFINITY) {
+		snprintf(buffer, sz, "%s: N/A\n", name_);
+	} else {
+		snprintf(buffer, sz, "%s: %0.2f (%0.2f..%0.2f, avg %0.2f)\n", name_, value_, min_, max_, smoothed_);
+	}
+}

Common/Math/Statistics.h

@@ -0,0 +1,42 @@
+#pragma once
+
+#include <cmath>
+
+// Very simple stat for convenience. Keeps track of min, max, smoothed.
+struct SimpleStat {
+	SimpleStat(const char *name) : name_(name) { Reset(); }
+
+	void Update(double value) {
+		value_ = value;
+		if (min_ == INFINITY) {
+			smoothed_ = value;
+		} else {
+			// TODO: Make factor adjustable?
+			smoothed_ = 0.99 * smoothed_ + 0.01 * value;
+		}
+		if (value < min_) {
+			min_ = value;
+		}
+		if (value > max_) {
+			max_ = value;
+		}
+	}
+
+	void Reset() {
+		value_ = 0.0;
+		smoothed_ = 0.0;  // doens't really need init
+		min_ = INFINITY;
+		max_ = -INFINITY;
+	}
+
+	void Format(char *buffer, size_t sz);
+
+private:
+	const char *name_;
+
+	// These are initialized in Reset().
+	double value_;
+	double min_;
+	double max_;
+	double smoothed_;
+};

Common/Math/math_util.cpp

@@ -1,44 +1,5 @@
 #include "Common/Math/math_util.h"
-#include <stdlib.h>
 
-// QNX can only use RunFast mode and it is already the default.
-#if defined(__ARM_ARCH_7A__)
-// Enables 'RunFast' VFP mode.
-void EnableFZ() {
-	int x;
-	asm(
-		"fmrx %[result],FPSCR \r\n"
-		"orr %[result],%[result],#16777216 \r\n"
-		"fmxr FPSCR,%[result]"
-		:[result] "=r" (x) : :
-	);
-	//printf("ARM FPSCR: %08x\n",x);
-}
+#include <cstdlib>
 
-// New fastmode code from: http://pandorawiki.org/Floating_Point_Optimization
-// These settings turbocharge the slow VFP unit on Cortex-A8 based chips by setting
-// restrictions that permit running VFP instructions on the NEON unit.
-// Denormal flush-to-zero, for example.
-void FPU_SetFastMode() {
-	static const unsigned int x = 0x04086060;
-	static const unsigned int y = 0x03000000;
-	int r;
-	asm volatile (
-		"fmrx	%0, fpscr			\n\t"	//r0 = FPSCR
-		"and	%0, %0, %1			\n\t"	//r0 = r0 & 0x04086060
-		"orr	%0, %0, %2			\n\t"	//r0 = r0 | 0x03000000
-		"fmxr	fpscr, %0			\n\t"	//FPSCR = r0
-		: "=r"(r)
-		: "r"(x), "r"(y)
-		);
-}
-
-#else
-
-void EnableFZ() {
-	// TODO
-}
-
-void FPU_SetFastMode() {}
-
-#endif
+// Could delete this file, but might find use again.

Common/Math/math_util.h

@@ -189,12 +189,3 @@ inline uint16_t ShrinkToHalf(float full) {
 	FP16 fp = float_to_half_fast3(fp32);
 	return fp.u;
 }
-
-// FPU control.
-void EnableFZ();
-
-// Enable both FZ and Default-NaN. Is documented to flip some ARM implementation into a "run-fast" mode
-// where they can schedule VFP instructions on the NEON unit (these implementations have
-// very slow VFP units).
-// http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0274h/Babffifj.html
-void FPU_SetFastMode();

GPU/Common/FramebufferManagerCommon.cpp

@@ -1215,7 +1215,7 @@ void FramebufferManagerCommon::CopyFramebufferForColorTexture(VirtualFramebuffer
 	}
 
 	if (x < src->drawnWidth && y < src->drawnHeight && w > 0 && h > 0) {
-		BlitFramebuffer(dst, x, y, src, x, y, w, h, 0, RASTER_COLOR, "Blit_CopyFramebufferForColorTexture");
+		BlitFramebuffer(dst, x, y, src, x, y, w, h, 0, RASTER_COLOR, "CopyFBForColorTexture");
 	}
 }
 

Qt/QtMain.cpp

@@ -316,7 +316,6 @@ static int mainInternal(QApplication &a) {
 	emugl->resize(pixel_xres, pixel_yres);
 	emugl->showFullScreen();
 #endif
-	EnableFZ();
 	// Disable screensaver
 #if defined(QT_HAS_SYSTEMINFO)
 	QScreenSaver ssObject(emugl);

SDL/SDLMain.cpp

@@ -807,7 +807,6 @@ int main(int argc, char *argv[]) {
 	} else {
 		joystick = nullptr;
 	}
-	EnableFZ();
 
 	int framecount = 0;
 	bool mouseDown = false;

UI/EmuScreen.cpp

@@ -344,6 +344,8 @@ void EmuScreen::bootGame(const Path &filename) {
 
 	loadingViewColor_->Divert(0xFFFFFFFF, 0.75f);
 	loadingViewVisible_->Divert(UI::V_VISIBLE, 0.75f);
+
+	screenManager()->getDrawContext()->ResetStats();
 }
 
 void EmuScreen::bootComplete() {

UWP/CommonUWP/CommonUWP.vcxproj

@@ -387,6 +387,7 @@
     <ClInclude Include="..\..\Common\BitSet.h" />
     <ClInclude Include="..\..\Common\Buffer.h" />
     <ClInclude Include="..\..\Common\File\AndroidStorage.h" />
+    <ClInclude Include="..\..\Common\Math\Statistics.h" />
     <ClInclude Include="..\..\Common\Net\NetBuffer.h" />
     <ClInclude Include="..\..\Common\Data\Collections\ConstMap.h" />
     <ClInclude Include="..\..\Common\Data\Collections\FixedSizeQueue.h" />
@@ -519,6 +520,7 @@
     <ClCompile Include="..\..\Common\ArmEmitter.cpp" />
     <ClCompile Include="..\..\Common\Buffer.cpp" />
     <ClCompile Include="..\..\Common\File\AndroidStorage.cpp" />
+    <ClCompile Include="..\..\Common\Math\Statistics.cpp" />
     <ClCompile Include="..\..\Common\Net\NetBuffer.cpp" />
     <ClCompile Include="..\..\Common\Data\Color\RGBAUtil.cpp" />
     <ClCompile Include="..\..\Common\Data\Convert\SmallDataConvert.cpp" />
@@ -639,4 +641,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

UWP/CommonUWP/CommonUWP.vcxproj.filters

@@ -372,6 +372,9 @@
     <ClCompile Include="..\..\Common\File\AndroidStorage.cpp">
       <Filter>File</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\Common\Math\Statistics.cpp">
+      <Filter>Math</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="targetver.h" />
@@ -688,11 +691,9 @@
     <ClInclude Include="..\..\Common\File\AndroidStorage.h">
       <Filter>File</Filter>
     </ClInclude>
-  </ItemGroup>
-  <ItemGroup>
-    <Text Include="..\..\ext\libpng17\CMakeLists.txt">
-      <Filter>ext\libpng17</Filter>
-    </Text>
+    <ClInclude Include="..\..\Common\Math\Statistics.h">
+      <Filter>Math</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="..\..\Common\Math\fast\fast_matrix_neon.S">
@@ -701,5 +702,6 @@
     <None Include="..\..\Common\Math\lin\matrix_neon.s">
       <Filter>Math\lin</Filter>
     </None>
+    <None Include="..\..\ext\libpng17\CMakeLists.txt" />
   </ItemGroup>
-</Project>
+</Project>

android/jni/Android.mk

@@ -164,6 +164,7 @@ EXEC_AND_LIB_FILES := \
   $(SRC)/Common/Input/InputState.cpp \
   $(SRC)/Common/Math/fast/fast_matrix.c \
   $(SRC)/Common/Math/math_util.cpp \
+  $(SRC)/Common/Math/Statistics.cpp \
   $(SRC)/Common/Math/curves.cpp \
   $(SRC)/Common/Math/expression_parser.cpp \
   $(SRC)/Common/Math/lin/vec3.cpp.arm \

ios/ViewController.mm

@@ -807,6 +807,3 @@ void bindDefaultFBO()
 {
 	[sharedViewController bindDefaultFBO];
 }
-
-void EnableFZ(){};
-void DisableFZ(){};

libretro/Makefile.common

@@ -271,6 +271,7 @@ SOURCES_CXX += \
 	$(COMMONDIR)/Math/curves.cpp \
 	$(COMMONDIR)/Math/expression_parser.cpp \
 	$(COMMONDIR)/Math/math_util.cpp \
+	$(COMMONDIR)/Math/Statistics.cpp \
 	$(COMMONDIR)/Math/lin/vec3.cpp \
 	$(COMMONDIR)/Math/lin/matrix4x4.cpp \
 	$(COMMONDIR)/Net/HTTPClient.cpp \