Benchmark-mode for macrobench
Adds googlebench output format support
Adds offscreen rendering for >60fps benchmarking
Adds 'all' alias to run all registered TestScenes
Change-Id: I2579e40f2f4c941bfbd90c75efbee384c08a116b
diff --git a/libs/hwui/tests/macrobench/TestSceneRunner.cpp b/libs/hwui/tests/macrobench/TestSceneRunner.cpp
index 2d78437..f03dcbf 100644
--- a/libs/hwui/tests/macrobench/TestSceneRunner.cpp
+++ b/libs/hwui/tests/macrobench/TestSceneRunner.cpp
@@ -22,6 +22,7 @@
#include "renderthread/RenderProxy.h"
#include "renderthread/RenderTask.h"
+#include <benchmark/benchmark.h>
#include <cutils/log.h>
#include <gui/Surface.h>
#include <ui/PixelFormat.h>
@@ -62,13 +63,62 @@
T mAverage;
};
-void run(const TestScene::Info& info, const TestScene::Options& opts) {
+void outputBenchmarkReport(const TestScene::Info& info, const TestScene::Options& opts,
+ benchmark::BenchmarkReporter* reporter, RenderProxy* proxy,
+ double durationInS) {
+ using namespace benchmark;
+
+ struct ReportInfo {
+ int percentile;
+ const char* suffix;
+ };
+
+ static std::array<ReportInfo, 4> REPORTS = {
+ ReportInfo { 50, "_50th" },
+ ReportInfo { 90, "_90th" },
+ ReportInfo { 95, "_95th" },
+ ReportInfo { 99, "_99th" },
+ };
+
+ // Although a vector is used, it must stay with only a single element
+ // otherwise the BenchmarkReporter will automatically compute
+ // mean and stddev which doesn't make sense for our usage
+ std::vector<BenchmarkReporter::Run> reports;
+ BenchmarkReporter::Run report;
+ report.benchmark_name = info.name;
+ report.iterations = static_cast<int64_t>(opts.count);
+ report.real_accumulated_time = durationInS;
+ report.cpu_accumulated_time = durationInS;
+ report.items_per_second = opts.count / durationInS;
+ reports.push_back(report);
+ reporter->ReportRuns(reports);
+
+ // Pretend the percentiles are single-iteration runs of the test
+ // If rendering offscreen skip this as it's fps that's more interesting
+ // in that test case than percentiles.
+ if (!opts.renderOffscreen) {
+ for (auto& ri : REPORTS) {
+ reports[0].benchmark_name = info.name;
+ reports[0].benchmark_name += ri.suffix;
+ durationInS = proxy->frameTimePercentile(ri.percentile) / 1000.0;
+ reports[0].real_accumulated_time = durationInS;
+ reports[0].cpu_accumulated_time = durationInS;
+ reports[0].iterations = 1;
+ reports[0].items_per_second = 0;
+ reporter->ReportRuns(reports);
+ }
+ }
+}
+
+void run(const TestScene::Info& info, const TestScene::Options& opts,
+ benchmark::BenchmarkReporter* reporter) {
// Switch to the real display
gDisplay = getBuiltInDisplay();
std::unique_ptr<TestScene> scene(info.createScene(opts));
TestContext testContext;
+ testContext.setRenderOffscreen(opts.renderOffscreen);
// create the native surface
const int width = gDisplay.w;
@@ -91,7 +141,12 @@
proxy->setLightCenter((Vector3){lightX, dp(-200.0f), dp(800.0f)});
// Do a few cold runs then reset the stats so that the caches are all hot
- for (int i = 0; i < 5; i++) {
+ int warmupFrameCount = 5;
+ if (opts.renderOffscreen) {
+ // Do a few more warmups to try and boost the clocks up
+ warmupFrameCount = 10;
+ }
+ for (int i = 0; i < warmupFrameCount; i++) {
testContext.waitForVsync();
nsecs_t vsync = systemTime(CLOCK_MONOTONIC);
UiFrameInfoBuilder(proxy->frameInfo()).setVsync(vsync, vsync);
@@ -103,6 +158,7 @@
ModifiedMovingAverage<double> avgMs(opts.reportFrametimeWeight);
+ nsecs_t start = systemTime(CLOCK_MONOTONIC);
for (int i = 0; i < opts.count; i++) {
testContext.waitForVsync();
nsecs_t vsync = systemTime(CLOCK_MONOTONIC);
@@ -121,6 +177,13 @@
}
}
}
+ proxy->fence();
+ nsecs_t end = systemTime(CLOCK_MONOTONIC);
- proxy->dumpProfileInfo(STDOUT_FILENO, DumpFlags::JankStats);
+ if (reporter) {
+ outputBenchmarkReport(info, opts, reporter, proxy.get(),
+ (end - start) / (double) s2ns(1));
+ } else {
+ proxy->dumpProfileInfo(STDOUT_FILENO, DumpFlags::JankStats);
+ }
}