Macrobenchmark tweaks & gpu memory dumping
Add support for glob matching
Ensure glob matches are alphabetically sorted
Add feature to dump GPU memory usage after a test pass
Adjust gpu memory dump to be a bit more compact (skip empty outputs)
Test: this
Bug: 187718492
Change-Id: I6dc80b2d3379d8d10001116e1240727d9914bc10
diff --git a/libs/hwui/pipeline/skia/SkiaMemoryTracer.cpp b/libs/hwui/pipeline/skia/SkiaMemoryTracer.cpp
index e48ecf4..0b995bc 100644
--- a/libs/hwui/pipeline/skia/SkiaMemoryTracer.cpp
+++ b/libs/hwui/pipeline/skia/SkiaMemoryTracer.cpp
@@ -126,6 +126,12 @@
mCurrentValues.insert({valueName, {units, value}});
}
+bool SkiaMemoryTracer::hasOutput() {
+ // process any remaining elements
+ processElement();
+ return mResults.size() > 0;
+}
+
void SkiaMemoryTracer::logOutput(String8& log) {
// process any remaining elements
processElement();
diff --git a/libs/hwui/pipeline/skia/SkiaMemoryTracer.h b/libs/hwui/pipeline/skia/SkiaMemoryTracer.h
index e9a7981..b393b075 100644
--- a/libs/hwui/pipeline/skia/SkiaMemoryTracer.h
+++ b/libs/hwui/pipeline/skia/SkiaMemoryTracer.h
@@ -34,6 +34,7 @@
SkiaMemoryTracer(const char* categoryKey, bool itemizeType);
~SkiaMemoryTracer() override {}
+ bool hasOutput();
void logOutput(String8& log);
void logTotals(String8& log);
diff --git a/libs/hwui/renderthread/CacheManager.cpp b/libs/hwui/renderthread/CacheManager.cpp
index d998e50..5047be9 100644
--- a/libs/hwui/renderthread/CacheManager.cpp
+++ b/libs/hwui/renderthread/CacheManager.cpp
@@ -140,7 +140,6 @@
log.appendFormat(" Size: %.2f kB \n", SkGraphics::GetFontCacheUsed() / 1024.0f);
log.appendFormat(" Glyph Count: %d \n", SkGraphics::GetFontCacheCountUsed());
- log.appendFormat("CPU Caches:\n");
std::vector<skiapipeline::ResourcePair> cpuResourceMap = {
{"skia/sk_resource_cache/bitmap_", "Bitmaps"},
{"skia/sk_resource_cache/rrect-blur_", "Masks"},
@@ -149,20 +148,20 @@
};
skiapipeline::SkiaMemoryTracer cpuTracer(cpuResourceMap, false);
SkGraphics::DumpMemoryStatistics(&cpuTracer);
- cpuTracer.logOutput(log);
+ if (cpuTracer.hasOutput()) {
+ log.appendFormat("CPU Caches:\n");
+ cpuTracer.logOutput(log);
+ }
- log.appendFormat("GPU Caches:\n");
skiapipeline::SkiaMemoryTracer gpuTracer("category", true);
mGrContext->dumpMemoryStatistics(&gpuTracer);
- gpuTracer.logOutput(log);
+ if (gpuTracer.hasOutput()) {
+ log.appendFormat("GPU Caches:\n");
+ gpuTracer.logOutput(log);
+ }
- log.appendFormat("Other Caches:\n");
- log.appendFormat(" Current / Maximum\n");
-
- if (renderState) {
- if (renderState->mActiveLayers.size() > 0) {
- log.appendFormat(" Layer Info:\n");
- }
+ if (renderState && renderState->mActiveLayers.size() > 0) {
+ log.appendFormat("Layer Info:\n");
const char* layerType = Properties::getRenderPipelineType() == RenderPipelineType::SkiaGL
? "GlLayer"
diff --git a/libs/hwui/renderthread/RenderProxy.cpp b/libs/hwui/renderthread/RenderProxy.cpp
index 1b4b4b9..ad325cf 100644
--- a/libs/hwui/renderthread/RenderProxy.cpp
+++ b/libs/hwui/renderthread/RenderProxy.cpp
@@ -249,10 +249,10 @@
});
}
-void RenderProxy::dumpGraphicsMemory(int fd) {
+void RenderProxy::dumpGraphicsMemory(int fd, bool includeProfileData) {
if (RenderThread::hasInstance()) {
auto& thread = RenderThread::getInstance();
- thread.queue().runSync([&]() { thread.dumpGraphicsMemory(fd); });
+ thread.queue().runSync([&]() { thread.dumpGraphicsMemory(fd, includeProfileData); });
}
}
diff --git a/libs/hwui/renderthread/RenderProxy.h b/libs/hwui/renderthread/RenderProxy.h
index 288f555..662b445 100644
--- a/libs/hwui/renderthread/RenderProxy.h
+++ b/libs/hwui/renderthread/RenderProxy.h
@@ -109,7 +109,7 @@
// Not exported, only used for testing
void resetProfileInfo();
uint32_t frameTimePercentile(int p);
- static void dumpGraphicsMemory(int fd);
+ static void dumpGraphicsMemory(int fd, bool includeProfileData = true);
static void rotateProcessStatsBuffer();
static void setProcessStatsBuffer(int fd);
diff --git a/libs/hwui/renderthread/RenderThread.cpp b/libs/hwui/renderthread/RenderThread.cpp
index 3421e01..308352d 100644
--- a/libs/hwui/renderthread/RenderThread.cpp
+++ b/libs/hwui/renderthread/RenderThread.cpp
@@ -302,30 +302,25 @@
return *mVkManager.get();
}
-void RenderThread::dumpGraphicsMemory(int fd) {
- globalProfileData()->dump(fd);
-
- String8 cachesOutput;
- String8 pipeline;
- auto renderType = Properties::getRenderPipelineType();
- switch (renderType) {
- case RenderPipelineType::SkiaGL: {
- mCacheManager->dumpMemoryUsage(cachesOutput, mRenderState);
- pipeline.appendFormat("Skia (OpenGL)");
- break;
- }
- case RenderPipelineType::SkiaVulkan: {
- mCacheManager->dumpMemoryUsage(cachesOutput, mRenderState);
- pipeline.appendFormat("Skia (Vulkan)");
- break;
- }
+static const char* pipelineToString() {
+ switch (auto renderType = Properties::getRenderPipelineType()) {
+ case RenderPipelineType::SkiaGL:
+ return "Skia (OpenGL)";
+ case RenderPipelineType::SkiaVulkan:
+ return "Skia (Vulkan)";
default:
LOG_ALWAYS_FATAL("canvas context type %d not supported", (int32_t)renderType);
- break;
+ }
+}
+
+void RenderThread::dumpGraphicsMemory(int fd, bool includeProfileData) {
+ if (includeProfileData) {
+ globalProfileData()->dump(fd);
}
- dprintf(fd, "\n%s\n", cachesOutput.string());
- dprintf(fd, "\nPipeline=%s\n", pipeline.string());
+ String8 cachesOutput;
+ mCacheManager->dumpMemoryUsage(cachesOutput, mRenderState);
+ dprintf(fd, "\nPipeline=%s\n%s\n", pipelineToString(), cachesOutput.string());
}
Readback& RenderThread::readback() {
diff --git a/libs/hwui/renderthread/RenderThread.h b/libs/hwui/renderthread/RenderThread.h
index cd9b923..afd5750 100644
--- a/libs/hwui/renderthread/RenderThread.h
+++ b/libs/hwui/renderthread/RenderThread.h
@@ -150,7 +150,7 @@
VulkanManager& vulkanManager();
sk_sp<Bitmap> allocateHardwareBitmap(SkBitmap& skBitmap);
- void dumpGraphicsMemory(int fd);
+ void dumpGraphicsMemory(int fd, bool includeProfileData);
void requireGlContext();
void requireVkContext();
diff --git a/libs/hwui/tests/common/TestScene.cpp b/libs/hwui/tests/common/TestScene.cpp
index 02bcd47..2c532b0 100644
--- a/libs/hwui/tests/common/TestScene.cpp
+++ b/libs/hwui/tests/common/TestScene.cpp
@@ -22,8 +22,9 @@
// Not a static global because we need to force the map to be constructed
// before we try to add things to it.
-std::unordered_map<std::string, TestScene::Info>& TestScene::testMap() {
- static std::unordered_map<std::string, TestScene::Info> testMap;
+// std::map because tests sorted by name is a prettier output
+std::map<std::string, TestScene::Info>& TestScene::testMap() {
+ static std::map<std::string, TestScene::Info> testMap;
return testMap;
}
diff --git a/libs/hwui/tests/common/TestScene.h b/libs/hwui/tests/common/TestScene.h
index eaf5988..781884a 100644
--- a/libs/hwui/tests/common/TestScene.h
+++ b/libs/hwui/tests/common/TestScene.h
@@ -19,8 +19,8 @@
#include <gui/Surface.h>
#include <utils/StrongPointer.h>
+#include <map>
#include <string>
-#include <unordered_map>
namespace android {
@@ -39,6 +39,7 @@
int repeatCount = 1;
int reportFrametimeWeight = 0;
bool renderOffscreen = true;
+ bool reportGpuMemoryUsage = false;
};
template <class T>
@@ -68,7 +69,7 @@
virtual void createContent(int width, int height, Canvas& renderer) = 0;
virtual void doFrame(int frameNr) = 0;
- static std::unordered_map<std::string, Info>& testMap();
+ static std::map<std::string, Info>& testMap();
static void registerScene(const Info& info);
sp<Surface> renderTarget;
diff --git a/libs/hwui/tests/macrobench/TestSceneRunner.cpp b/libs/hwui/tests/macrobench/TestSceneRunner.cpp
index cf9b0c5..9d3b732 100644
--- a/libs/hwui/tests/macrobench/TestSceneRunner.cpp
+++ b/libs/hwui/tests/macrobench/TestSceneRunner.cpp
@@ -162,6 +162,11 @@
void run(const TestScene::Info& info, const TestScene::Options& opts,
benchmark::BenchmarkReporter* reporter) {
+ if (opts.reportGpuMemoryUsage) {
+ // If we're reporting GPU memory usage we need to first start with a clean slate
+ // All repetitions of the same test will share a single memory usage report
+ RenderProxy::trimMemory(100);
+ }
BenchmarkResults results;
for (int i = 0; i < opts.repeatCount; i++) {
doRun(info, opts, i, reporter ? &results : nullptr);
@@ -172,4 +177,7 @@
// TODO: Report summary
}
}
+ if (opts.reportGpuMemoryUsage) {
+ RenderProxy::dumpGraphicsMemory(STDOUT_FILENO, false);
+ }
}
diff --git a/libs/hwui/tests/macrobench/main.cpp b/libs/hwui/tests/macrobench/main.cpp
index acbbb95..e9e962a 100644
--- a/libs/hwui/tests/macrobench/main.cpp
+++ b/libs/hwui/tests/macrobench/main.cpp
@@ -23,6 +23,7 @@
#include "renderthread/RenderProxy.h"
#include <benchmark/benchmark.h>
+#include <fnmatch.h>
#include <getopt.h>
#include <pthread.h>
#include <stdio.h>
@@ -70,6 +71,7 @@
--benchmark_format Set output format. Possible values are tabular, json, csv
--renderer=TYPE Sets the render pipeline to use. May be skiagl or skiavk
--skip-leak-check Skips the memory leak check
+ --report-gpu-memory Dumps the GPU memory usage after each test run
)");
}
@@ -172,6 +174,7 @@
Offscreen,
Renderer,
SkipLeakCheck,
+ ReportGpuMemory,
};
}
@@ -188,6 +191,7 @@
{"offscreen", no_argument, nullptr, LongOpts::Offscreen},
{"renderer", required_argument, nullptr, LongOpts::Renderer},
{"skip-leak-check", no_argument, nullptr, LongOpts::SkipLeakCheck},
+ {"report-gpu-memory", no_argument, nullptr, LongOpts::ReportGpuMemory},
{0, 0, 0, 0}};
static const char* SHORT_OPTIONS = "c:r:h";
@@ -290,6 +294,10 @@
gRunLeakCheck = false;
break;
+ case LongOpts::ReportGpuMemory:
+ gOpts.reportGpuMemoryUsage = true;
+ break;
+
case 'h':
printHelp();
exit(EXIT_SUCCESS);
@@ -313,12 +321,21 @@
if (optind < argc) {
do {
const char* test = argv[optind++];
- auto pos = TestScene::testMap().find(test);
- if (pos == TestScene::testMap().end()) {
- fprintf(stderr, "Unknown test '%s'\n", test);
- exit(EXIT_FAILURE);
+ if (strchr(test, '*')) {
+ // Glob match
+ for (auto& iter : TestScene::testMap()) {
+ if (!fnmatch(test, iter.first.c_str(), 0)) {
+ gRunTests.push_back(iter.second);
+ }
+ }
} else {
- gRunTests.push_back(pos->second);
+ auto pos = TestScene::testMap().find(test);
+ if (pos == TestScene::testMap().end()) {
+ fprintf(stderr, "Unknown test '%s'\n", test);
+ exit(EXIT_FAILURE);
+ } else {
+ gRunTests.push_back(pos->second);
+ }
}
} while (optind < argc);
} else {