int tool_main(int argc, char** argv) {
SetupCrashHandler();
SkCommandLineFlags::Parse(argc, argv);
#if SK_ENABLE_INST_COUNT
if (FLAGS_leaks) {
gPrintInstCount = true;
}
#endif
SkAutoGraphics ag;
// First, parse some flags.
BenchLogger logger;
if (FLAGS_logFile.count()) {
logger.SetLogFile(FLAGS_logFile[0]);
}
LoggerResultsWriter logWriter(logger, FLAGS_timeFormat[0]);
MultiResultsWriter writer;
writer.add(&logWriter);
SkAutoTDelete<JSONResultsWriter> jsonWriter;
if (FLAGS_outResultsFile.count()) {
jsonWriter.reset(SkNEW(JSONResultsWriter(FLAGS_outResultsFile[0])));
writer.add(jsonWriter.get());
}
// Instantiate after all the writers have been added to writer so that we
// call close() before their destructors are called on the way out.
CallEnd<MultiResultsWriter> ender(writer);
const uint8_t alpha = FLAGS_forceBlend ? 0x80 : 0xFF;
SkTriState::State dither = SkTriState::kDefault;
for (size_t i = 0; i < 3; i++) {
if (strcmp(SkTriState::Name[i], FLAGS_forceDither[0]) == 0) {
dither = static_cast<SkTriState::State>(i);
}
}
BenchMode benchMode = kNormal_BenchMode;
for (size_t i = 0; i < SK_ARRAY_COUNT(BenchMode_Name); i++) {
if (strcmp(FLAGS_mode[0], BenchMode_Name[i]) == 0) {
benchMode = static_cast<BenchMode>(i);
}
}
SkTDArray<int> configs;
bool runDefaultConfigs = false;
// Try user-given configs first.
for (int i = 0; i < FLAGS_config.count(); i++) {
for (int j = 0; j < static_cast<int>(SK_ARRAY_COUNT(gConfigs)); ++j) {
if (0 == strcmp(FLAGS_config[i], gConfigs[j].name)) {
*configs.append() = j;
} else if (0 == strcmp(FLAGS_config[i], kDefaultsConfigStr)) {
runDefaultConfigs = true;
}
}
}
// If there weren't any, fill in with defaults.
if (runDefaultConfigs) {
for (int i = 0; i < static_cast<int>(SK_ARRAY_COUNT(gConfigs)); ++i) {
if (gConfigs[i].runByDefault) {
*configs.append() = i;
}
}
}
// Filter out things we can't run.
if (kNormal_BenchMode != benchMode) {
// Non-rendering configs only run in normal mode
for (int i = 0; i < configs.count(); ++i) {
const Config& config = gConfigs[configs[i]];
if (Benchmark::kNonRendering_Backend == config.backend) {
configs.remove(i, 1);
--i;
}
}
}
#if SK_SUPPORT_GPU
for (int i = 0; i < configs.count(); ++i) {
const Config& config = gConfigs[configs[i]];
if (Benchmark::kGPU_Backend == config.backend) {
GrContext* context = gContextFactory.get(config.contextType);
if (NULL == context) {
SkDebugf("GrContext could not be created for config %s. Config will be skipped.\n",
config.name);
configs.remove(i);
--i;
continue;
}
if (config.sampleCount > context->getMaxSampleCount()){
SkDebugf(
"Sample count (%d) for config %s is not supported. Config will be skipped.\n",
config.sampleCount, config.name);
configs.remove(i);
--i;
continue;
}
}
}
#endif
// All flags should be parsed now. Report our settings.
if (FLAGS_runOnce) {
logger.logError("bench was run with --runOnce, so we're going to hide the times."
" It's for your own good!\n");
}
writer.option("mode", FLAGS_mode[0]);
writer.option("alpha", SkStringPrintf("0x%02X", alpha).c_str());
writer.option("antialias", SkStringPrintf("%d", FLAGS_forceAA).c_str());
writer.option("filter", SkStringPrintf("%d", FLAGS_forceFilter).c_str());
writer.option("dither", SkTriState::Name[dither]);
writer.option("rotate", SkStringPrintf("%d", FLAGS_rotate).c_str());
writer.option("scale", SkStringPrintf("%d", FLAGS_scale).c_str());
writer.option("clip", SkStringPrintf("%d", FLAGS_clip).c_str());
#if defined(SK_BUILD_FOR_WIN32)
writer.option("system", "WIN32");
#elif defined(SK_BUILD_FOR_MAC)
writer.option("system", "MAC");
#elif defined(SK_BUILD_FOR_ANDROID)
writer.option("system", "ANDROID");
#elif defined(SK_BUILD_FOR_UNIX)
writer.option("system", "UNIX");
#else
writer.option("system", "other");
#endif
#if defined(SK_DEBUG)
writer.option("build", "DEBUG");
#else
writer.option("build", "RELEASE");
#endif
// Set texture cache limits if non-default.
for (size_t i = 0; i < SK_ARRAY_COUNT(gConfigs); ++i) {
#if SK_SUPPORT_GPU
const Config& config = gConfigs[i];
if (Benchmark::kGPU_Backend != config.backend) {
continue;
}
GrContext* context = gContextFactory.get(config.contextType);
if (NULL == context) {
continue;
}
size_t bytes;
int count;
context->getResourceCacheLimits(&count, &bytes);
if (-1 != FLAGS_gpuCacheBytes) {
bytes = static_cast<size_t>(FLAGS_gpuCacheBytes);
}
if (-1 != FLAGS_gpuCacheCount) {
count = FLAGS_gpuCacheCount;
}
context->setResourceCacheLimits(count, bytes);
#endif
}
// Run each bench in each configuration it supports and we asked for.
Iter iter;
Benchmark* bench;
while ((bench = iter.next()) != NULL) {
SkAutoTUnref<Benchmark> benchUnref(bench);
if (SkCommandLineFlags::ShouldSkip(FLAGS_match, bench->getName())) {
continue;
}
bench->setForceAlpha(alpha);
bench->setForceAA(FLAGS_forceAA);
bench->setForceFilter(FLAGS_forceFilter);
bench->setDither(dither);
bench->preDraw();
bool loggedBenchName = false;
for (int i = 0; i < configs.count(); ++i) {
const int configIndex = configs[i];
const Config& config = gConfigs[configIndex];
if (!bench->isSuitableFor(config.backend)) {
continue;
}
GrContext* context = NULL;
#if SK_SUPPORT_GPU
SkGLContextHelper* glContext = NULL;
if (Benchmark::kGPU_Backend == config.backend) {
context = gContextFactory.get(config.contextType);
if (NULL == context) {
continue;
}
glContext = gContextFactory.getGLContext(config.contextType);
}
#endif
SkAutoTUnref<SkCanvas> canvas;
SkAutoTUnref<SkPicture> recordFrom;
SkPictureRecorder recorderTo;
const SkIPoint dim = bench->getSize();
SkAutoTUnref<SkSurface> surface;
if (Benchmark::kNonRendering_Backend != config.backend) {
surface.reset(make_surface(config.fColorType,
dim,
config.backend,
config.sampleCount,
context));
if (!surface.get()) {
logger.logError(SkStringPrintf(
"Device creation failure for config %s. Will skip.\n", config.name));
continue;
}
switch(benchMode) {
case kDeferredSilent_BenchMode:
case kDeferred_BenchMode:
canvas.reset(SkDeferredCanvas::Create(surface.get()));
break;
case kRecord_BenchMode:
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
break;
case kPictureRecord_BenchMode: {
SkPictureRecorder recorderFrom;
bench->draw(1, recorderFrom.beginRecording(dim.fX, dim.fY));
recordFrom.reset(recorderFrom.endRecording());
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
break;
}
case kNormal_BenchMode:
canvas.reset(SkRef(surface->getCanvas()));
break;
default:
SkASSERT(false);
}
}
if (NULL != canvas) {
canvas->clear(SK_ColorWHITE);
if (FLAGS_clip) {
perform_clip(canvas, dim.fX, dim.fY);
}
if (FLAGS_scale) {
perform_scale(canvas, dim.fX, dim.fY);
}
if (FLAGS_rotate) {
perform_rotate(canvas, dim.fX, dim.fY);
}
}
if (!loggedBenchName) {
loggedBenchName = true;
writer.bench(bench->getName(), dim.fX, dim.fY);
}
#if SK_SUPPORT_GPU
SkGLContextHelper* contextHelper = NULL;
if (Benchmark::kGPU_Backend == config.backend) {
contextHelper = gContextFactory.getGLContext(config.contextType);
}
BenchTimer timer(contextHelper);
#else
BenchTimer timer;
#endif
double previous = std::numeric_limits<double>::infinity();
bool converged = false;
// variables used to compute loopsPerFrame
double frameIntervalTime = 0.0f;
int frameIntervalTotalLoops = 0;
bool frameIntervalComputed = false;
int loopsPerFrame = 0;
int loopsPerIter = 0;
if (FLAGS_verbose) { SkDebugf("%s %s: ", bench->getName(), config.name); }
if (!FLAGS_dryRun) {
do {
// Ramp up 1 -> 2 -> 4 -> 8 -> 16 -> ... -> ~1 billion.
loopsPerIter = (loopsPerIter == 0) ? 1 : loopsPerIter * 2;
if (loopsPerIter >= (1<<30) || timer.fWall > FLAGS_maxMs) {
// If you find it takes more than a billion loops to get up to 20ms of runtime,
// you've got a computer clocked at several THz or have a broken benchmark. ;)
// "1B ought to be enough for anybody."
logger.logError(SkStringPrintf(
"\nCan't get %s %s to converge in %dms (%d loops)",
bench->getName(), config.name, FLAGS_maxMs, loopsPerIter));
break;
}
if ((benchMode == kRecord_BenchMode || benchMode == kPictureRecord_BenchMode)) {
// Clear the recorded commands so that they do not accumulate.
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
}
timer.start();
// Inner loop that allows us to break the run into smaller
// chunks (e.g. frames). This is especially useful for the GPU
// as we can flush and/or swap buffers to keep the GPU from
// queuing up too much work.
for (int loopCount = loopsPerIter; loopCount > 0; ) {
// Save and restore around each call to draw() to guarantee a pristine canvas.
SkAutoCanvasRestore saveRestore(canvas, true/*also save*/);
int loops;
if (frameIntervalComputed && loopCount > loopsPerFrame) {
loops = loopsPerFrame;
loopCount -= loopsPerFrame;
} else {
loops = loopCount;
loopCount = 0;
}
if (benchMode == kPictureRecord_BenchMode) {
recordFrom->draw(canvas);
} else {
bench->draw(loops, canvas);
}
if (kDeferredSilent_BenchMode == benchMode) {
static_cast<SkDeferredCanvas*>(canvas.get())->silentFlush();
} else if (NULL != canvas) {
canvas->flush();
}
#if SK_SUPPORT_GPU
// swap drawing buffers on each frame to prevent the GPU
// from queuing up too much work
if (NULL != glContext) {
glContext->swapBuffers();
}
#endif
}
// Stop truncated timers before GL calls complete, and stop the full timers after.
timer.truncatedEnd();
#if SK_SUPPORT_GPU
if (NULL != glContext) {
context->flush();
SK_GL(*glContext, Finish());
}
#endif
timer.end();
// setup the frame interval for subsequent iterations
if (!frameIntervalComputed) {
frameIntervalTime += timer.fWall;
frameIntervalTotalLoops += loopsPerIter;
if (frameIntervalTime >= FLAGS_minMs) {
frameIntervalComputed = true;
loopsPerFrame =
(int)(((double)frameIntervalTotalLoops / frameIntervalTime) * FLAGS_minMs);
if (loopsPerFrame < 1) {
loopsPerFrame = 1;
}
// SkDebugf(" %s has %d loops in %f ms (normalized to %d)\n",
// bench->getName(), frameIntervalTotalLoops,
// timer.fWall, loopsPerFrame);
}
}
const double current = timer.fWall / loopsPerIter;
if (FLAGS_verbose && current > previous) { SkDebugf("↑"); }
if (FLAGS_verbose) { SkDebugf("%.3g ", current); }
converged = HasConverged(previous, current, timer.fWall);
previous = current;
} while (!FLAGS_runOnce && !converged);
}
if (FLAGS_verbose) { SkDebugf("\n"); }
if (!FLAGS_dryRun && FLAGS_outDir.count() && Benchmark::kNonRendering_Backend != config.backend) {
SkAutoTUnref<SkImage> image(surface->newImageSnapshot());
if (image.get()) {
saveFile(bench->getName(), config.name, FLAGS_outDir[0],
image);
}
}
if (FLAGS_runOnce) {
// Let's not mislead ourselves by looking at Debug build or single iteration bench times!
continue;
}
// Normalize to ms per 1000 iterations.
const double normalize = 1000.0 / loopsPerIter;
const struct { char shortName; const char* longName; double ms; } times[] = {
{'w', "msecs", normalize * timer.fWall},
{'W', "Wmsecs", normalize * timer.fTruncatedWall},
{'c', "cmsecs", normalize * timer.fCpu},
{'C', "Cmsecs", normalize * timer.fTruncatedCpu},
{'g', "gmsecs", normalize * timer.fGpu},
};
writer.config(config.name);
for (size_t i = 0; i < SK_ARRAY_COUNT(times); i++) {
if (strchr(FLAGS_timers[0], times[i].shortName) && times[i].ms > 0) {
writer.timer(times[i].longName, times[i].ms);
}
}
}
}
#if SK_SUPPORT_GPU
gContextFactory.destroyContexts();
#endif
return 0;
}
