add autotuning to visualbench
BUG=skia:
Review URL: https://codereview.chromium.org/1201003002
diff --git a/tools/VisualBench.cpp b/tools/VisualBench.cpp
index 60174a4..cbc8994 100644
--- a/tools/VisualBench.cpp
+++ b/tools/VisualBench.cpp
@@ -24,9 +24,15 @@
__SK_FORCE_IMAGE_DECODER_LINKING;
+// Between samples we reset context
+// Between frames we swap buffers
+// Between flushes we call flush on GrContext
+
DEFINE_int32(gpuFrameLag, 5, "Overestimate of maximum number of frames GPU allows to lag.");
-DEFINE_int32(samples, 10, "Number of times to render each skp.");
-DEFINE_int32(loops, 5, "Number of times to time.");
+DEFINE_int32(samples, 10, "Number of times to time each skp.");
+DEFINE_int32(frames, 5, "Number of frames of each skp to render per sample.");
+DEFINE_double(flushMs, 20, "Target flush time in millseconds.");
+DEFINE_double(loopMs, 5, "Target loop time in millseconds.");
DEFINE_int32(msaa, 0, "Number of msaa samples.");
DEFINE_bool2(fullscreen, f, true, "Run fullscreen.");
@@ -41,10 +47,12 @@
VisualBench::VisualBench(void* hwnd, int argc, char** argv)
: INHERITED(hwnd)
- , fLoop(0)
, fCurrentPictureIdx(-1)
, fCurrentSample(0)
- , fState(kPreWarm_State) {
+ , fCurrentFrame(0)
+ , fFlushes(1)
+ , fLoops(1)
+ , fState(kPreWarmLoops_State) {
SkCommandLineFlags::Parse(argc, argv);
// read all the skp file names.
@@ -66,6 +74,9 @@
this->setTitle();
this->setupBackend();
+
+ // Print header
+ SkDebugf("curr/maxrss\tloops\tflushes\tmin\tmedian\tmean\tmax\tstddev\tbench\n");
}
VisualBench::~VisualBench() {
@@ -122,8 +133,12 @@
}
inline void VisualBench::renderFrame(SkCanvas* canvas) {
- canvas->drawPicture(fPicture);
- canvas->flush();
+ for (int flush = 0; flush < fFlushes; flush++) {
+ for (int loop = 0; loop < fLoops; loop++) {
+ canvas->drawPicture(fPicture);
+ }
+ canvas->flush();
+ }
INHERITED::present();
}
@@ -139,9 +154,11 @@
SkASSERT(measurements.count());
Stats stats(measurements.begin(), measurements.count());
const double stdDevPercent = 100 * sqrt(stats.var) / stats.mean;
- SkDebugf("%4d/%-4dMB\t%s\t%s\t%s\t%s\t%.0f%%\t%s\n",
+ SkDebugf("%4d/%-4dMB\t%d\t%d\t%s\t%s\t%s\t%s\t%.0f%%\t%s\n",
sk_tools::getCurrResidentSetSizeMB(),
sk_tools::getMaxResidentSetSizeMB(),
+ fLoops,
+ fFlushes,
HUMANIZE(stats.min),
HUMANIZE(stats.median),
HUMANIZE(stats.mean),
@@ -180,6 +197,17 @@
return false;
}
+void VisualBench::preWarm(State nextState) {
+ if (fCurrentFrame >= FLAGS_gpuFrameLag) {
+ // we currently time across all frames to make sure we capture all GPU work
+ fState = nextState;
+ fCurrentFrame = 0;
+ fTimer.start();
+ } else {
+ fCurrentFrame++;
+ }
+}
+
void VisualBench::draw(SkCanvas* canvas) {
if (!this->advanceRecordIfNecessary()) {
this->closeWindow();
@@ -187,34 +215,59 @@
}
this->renderFrame(canvas);
switch (fState) {
- case kPreWarm_State: {
- if (fCurrentSample >= FLAGS_gpuFrameLag) {
- // TODO we currently time across all frames to make sure we capture all GPU work
- // We should also rendering an empty SKP to get a baseline to subtract from
- // our timing
- fState = kTiming_State;
- fCurrentSample -= FLAGS_gpuFrameLag;
- fTimer.start();
+ case kPreWarmLoops_State: {
+ this->preWarm(kTuneLoops_State);
+ break;
+ }
+ case kTuneLoops_State: {
+ if (1 << 30 == fLoops) {
+ // We're about to wrap. Something's wrong with the bench.
+ SkDebugf("InnerLoops wrapped\n");
+ fLoops = 0;
} else {
- fCurrentSample++;
+ fTimer.end();
+ double elapsed = fTimer.fWall;
+ if (elapsed > FLAGS_loopMs) {
+ fState = kPreWarmTiming_State;
+
+ // Scale back the number of loops
+ fLoops = (int)ceil(fLoops * FLAGS_loopMs / elapsed);
+ fFlushes = (int)ceil(FLAGS_flushMs / elapsed);
+ } else {
+ fState = kPreWarmLoops_State;
+ fLoops *= 2;
+ }
+
+ fCurrentFrame = 0;
+ fTimer = WallTimer();
+ this->resetContext();
}
break;
}
+ case kPreWarmTiming_State: {
+ this->preWarm(kTiming_State);
+ break;
+ }
case kTiming_State: {
- if (fCurrentSample >= FLAGS_samples) {
+ if (fCurrentFrame >= FLAGS_frames) {
fTimer.end();
- fRecords[fCurrentPictureIdx].fMeasurements.push_back(fTimer.fWall / FLAGS_samples);
- this->resetContext();
- fTimer = WallTimer();
- fState = kPreWarm_State;
- fCurrentSample = 0;
- if (fLoop++ > FLAGS_loops) {
+ fRecords[fCurrentPictureIdx].fMeasurements.push_back(
+ fTimer.fWall / (FLAGS_frames * fLoops * fFlushes));
+ if (fCurrentSample++ >= FLAGS_samples) {
+ fState = kPreWarmLoops_State;
this->printStats();
fPicture.reset(NULL);
- fLoop = 0;
+ fCurrentSample = 0;
+ fFlushes = 1;
+ fLoops = 1;
+ } else {
+ fState = kPreWarmTiming_State;
}
+ fTimer = WallTimer();
+ this->resetContext();
+ fCurrentFrame = 0;
} else {
- fCurrentSample++;
+ fCurrentFrame++;
}
break;
}
diff --git a/tools/VisualBench.h b/tools/VisualBench.h
index 9d7b946..332fe82 100644
--- a/tools/VisualBench.h
+++ b/tools/VisualBench.h
@@ -54,14 +54,19 @@
};
enum State {
- kPreWarm_State,
+ kPreWarmLoops_State,
+ kTuneLoops_State,
+ kPreWarmTiming_State,
kTiming_State,
};
+ void preWarm(State nextState);
- int fLoop;
int fCurrentPictureIdx;
SkAutoTUnref<SkPicture> fPicture;
int fCurrentSample;
+ int fCurrentFrame;
+ int fFlushes;
+ int fLoops;
SkTArray<Record> fRecords;
WallTimer fTimer;
State fState;
