Home | History | Annotate | Download | only in bench
      1 /*
      2  * Copyright 2014 Google Inc.
      3  *
      4  * Use of this source code is governed by a BSD-style license that can be
      5  * found in the LICENSE file.
      6  */
      7 
      8 #include <ctype.h>
      9 
     10 #include "nanobench.h"
     11 
     12 #include "AndroidCodecBench.h"
     13 #include "Benchmark.h"
     14 #include "BitmapRegionDecoderBench.h"
     15 #include "CodecBench.h"
     16 #include "CodecBenchPriv.h"
     17 #include "CrashHandler.h"
     18 #include "GMBench.h"
     19 #include "ProcStats.h"
     20 #include "ResultsWriter.h"
     21 #include "RecordingBench.h"
     22 #include "SKPAnimationBench.h"
     23 #include "SKPBench.h"
     24 #include "Stats.h"
     25 
     26 #include "SkAndroidCodec.h"
     27 #include "SkBitmapRegionDecoder.h"
     28 #include "SkBBoxHierarchy.h"
     29 #include "SkCanvas.h"
     30 #include "SkCodec.h"
     31 #include "SkCommonFlags.h"
     32 #include "SkCommonFlagsConfig.h"
     33 #include "SkData.h"
     34 #include "SkForceLinking.h"
     35 #include "SkGraphics.h"
     36 #include "SkOSFile.h"
     37 #include "SkPictureRecorder.h"
     38 #include "SkPictureUtils.h"
     39 #include "SkString.h"
     40 #include "SkSurface.h"
     41 #include "SkTaskGroup.h"
     42 #include "SkThreadUtils.h"
     43 #include "ThermalManager.h"
     44 
     45 #include <stdlib.h>
     46 
     47 #ifndef SK_BUILD_FOR_WIN32
     48     #include <unistd.h>
     49 #endif
     50 
     51 #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK
     52     #include "nanobenchAndroid.h"
     53 #endif
     54 
     55 #if SK_SUPPORT_GPU
     56     #include "gl/GrGLDefines.h"
     57     #include "GrCaps.h"
     58     #include "GrContextFactory.h"
     59     SkAutoTDelete<GrContextFactory> gGrFactory;
     60 #endif
     61 
     62     struct GrContextOptions;
     63 
     64 __SK_FORCE_IMAGE_DECODER_LINKING;
     65 
     66 static const int kAutoTuneLoops = 0;
     67 
     68 static const int kDefaultLoops =
     69 #ifdef SK_DEBUG
     70     1;
     71 #else
     72     kAutoTuneLoops;
     73 #endif
     74 
     75 static SkString loops_help_txt() {
     76     SkString help;
     77     help.printf("Number of times to run each bench. Set this to %d to auto-"
     78                 "tune for each bench. Timings are only reported when auto-tuning.",
     79                 kAutoTuneLoops);
     80     return help;
     81 }
     82 
     83 static SkString to_string(int n) {
     84     SkString str;
     85     str.appendS32(n);
     86     return str;
     87 }
     88 
     89 DEFINE_int32(loops, kDefaultLoops, loops_help_txt().c_str());
     90 
     91 DEFINE_int32(samples, 10, "Number of samples to measure for each bench.");
     92 DEFINE_int32(ms, 0, "If >0, run each bench for this many ms instead of obeying --samples.");
     93 DEFINE_int32(overheadLoops, 100000, "Loops to estimate timer overhead.");
     94 DEFINE_double(overheadGoal, 0.0001,
     95               "Loop until timer overhead is at most this fraction of our measurments.");
     96 DEFINE_double(gpuMs, 5, "Target bench time in millseconds for GPU.");
     97 DEFINE_int32(gpuFrameLag, 5, "If unknown, estimated maximum number of frames GPU allows to lag.");
     98 DEFINE_bool(gpuCompressAlphaMasks, false, "Compress masks generated from falling back to "
     99                                           "software path rendering.");
    100 
    101 DEFINE_string(outResultsFile, "", "If given, write results here as JSON.");
    102 DEFINE_int32(maxCalibrationAttempts, 3,
    103              "Try up to this many times to guess loops for a bench, or skip the bench.");
    104 DEFINE_int32(maxLoops, 1000000, "Never run a bench more times than this.");
    105 DEFINE_string(clip, "0,0,1000,1000", "Clip for SKPs.");
    106 DEFINE_string(scales, "1.0", "Space-separated scales for SKPs.");
    107 DEFINE_string(zoom, "1.0,0", "Comma-separated zoomMax,zoomPeriodMs factors for a periodic SKP zoom "
    108                              "function that ping-pongs between 1.0 and zoomMax.");
    109 DEFINE_bool(bbh, true, "Build a BBH for SKPs?");
    110 DEFINE_bool(mpd, true, "Use MultiPictureDraw for the SKPs?");
    111 DEFINE_bool(loopSKP, true, "Loop SKPs like we do for micro benches?");
    112 DEFINE_int32(flushEvery, 10, "Flush --outResultsFile every Nth run.");
    113 DEFINE_bool(resetGpuContext, true, "Reset the GrContext before running each test.");
    114 DEFINE_bool(gpuStats, false, "Print GPU stats after each gpu benchmark?");
    115 DEFINE_bool(gpuStatsDump, false, "Dump GPU states after each benchmark to json");
    116 DEFINE_bool(keepAlive, false, "Print a message every so often so that we don't time out");
    117 DEFINE_string(useThermalManager, "0,1,10,1000", "enabled,threshold,sleepTimeMs,TimeoutMs for "
    118                                                 "thermalManager\n");
    119 
    120 DEFINE_string(sourceType, "",
    121         "Apply usual --match rules to source type: bench, gm, skp, image, etc.");
    122 DEFINE_string(benchType,  "",
    123         "Apply usual --match rules to bench type: micro, recording, playback, skcodec, etc.");
    124 
    125 static double now_ms() { return SkTime::GetNSecs() * 1e-6; }
    126 
    127 static SkString humanize(double ms) {
    128     if (FLAGS_verbose) return SkStringPrintf("%llu", (uint64_t)(ms*1e6));
    129     return HumanizeMs(ms);
    130 }
    131 #define HUMANIZE(ms) humanize(ms).c_str()
    132 
    133 bool Target::init(SkImageInfo info, Benchmark* bench) {
    134     if (Benchmark::kRaster_Backend == config.backend) {
    135         this->surface.reset(SkSurface::NewRaster(info));
    136         if (!this->surface.get()) {
    137             return false;
    138         }
    139     }
    140     return true;
    141 }
    142 bool Target::capturePixels(SkBitmap* bmp) {
    143     SkCanvas* canvas = this->getCanvas();
    144     if (!canvas) {
    145         return false;
    146     }
    147     bmp->setInfo(canvas->imageInfo());
    148     if (!canvas->readPixels(bmp, 0, 0)) {
    149         SkDebugf("Can't read canvas pixels.\n");
    150         return false;
    151     }
    152     return true;
    153 }
    154 
    155 #if SK_SUPPORT_GPU
    156 struct GPUTarget : public Target {
    157     explicit GPUTarget(const Config& c) : Target(c), gl(nullptr) { }
    158     SkGLContext* gl;
    159 
    160     void setup() override {
    161         this->gl->makeCurrent();
    162         // Make sure we're done with whatever came before.
    163         SK_GL(*this->gl, Finish());
    164     }
    165     void endTiming() override {
    166         if (this->gl) {
    167             SK_GL(*this->gl, Flush());
    168             this->gl->waitOnSyncOrSwap();
    169         }
    170     }
    171     void fence() override {
    172         SK_GL(*this->gl, Finish());
    173     }
    174 
    175     bool needsFrameTiming(int* maxFrameLag) const override {
    176         if (!this->gl->getMaxGpuFrameLag(maxFrameLag)) {
    177             // Frame lag is unknown.
    178             *maxFrameLag = FLAGS_gpuFrameLag;
    179         }
    180         return true;
    181     }
    182     bool init(SkImageInfo info, Benchmark* bench) override {
    183         uint32_t flags = this->config.useDFText ? SkSurfaceProps::kUseDeviceIndependentFonts_Flag :
    184                                                   0;
    185         SkSurfaceProps props(flags, SkSurfaceProps::kLegacyFontHost_InitType);
    186         this->surface.reset(SkSurface::NewRenderTarget(gGrFactory->get(this->config.ctxType,
    187                                                                        this->config.ctxOptions),
    188                                                          SkBudgeted::kNo, info,
    189                                                          this->config.samples, &props));
    190         this->gl = gGrFactory->getContextInfo(this->config.ctxType,
    191                                               this->config.ctxOptions).fGLContext;
    192         if (!this->surface.get()) {
    193             return false;
    194         }
    195         if (!this->gl->fenceSyncSupport()) {
    196             SkDebugf("WARNING: GL context for config \"%s\" does not support fence sync. "
    197                      "Timings might not be accurate.\n", this->config.name.c_str());
    198         }
    199         return true;
    200     }
    201     void fillOptions(ResultsWriter* log) override {
    202         const GrGLubyte* version;
    203         SK_GL_RET(*this->gl, version, GetString(GR_GL_VERSION));
    204         log->configOption("GL_VERSION", (const char*)(version));
    205 
    206         SK_GL_RET(*this->gl, version, GetString(GR_GL_RENDERER));
    207         log->configOption("GL_RENDERER", (const char*) version);
    208 
    209         SK_GL_RET(*this->gl, version, GetString(GR_GL_VENDOR));
    210         log->configOption("GL_VENDOR", (const char*) version);
    211 
    212         SK_GL_RET(*this->gl, version, GetString(GR_GL_SHADING_LANGUAGE_VERSION));
    213         log->configOption("GL_SHADING_LANGUAGE_VERSION", (const char*) version);
    214     }
    215 };
    216 
    217 #endif
    218 
    219 static double time(int loops, Benchmark* bench, Target* target) {
    220     SkCanvas* canvas = target->getCanvas();
    221     if (canvas) {
    222         canvas->clear(SK_ColorWHITE);
    223     }
    224     bench->preDraw(canvas);
    225     double start = now_ms();
    226     canvas = target->beginTiming(canvas);
    227     bench->draw(loops, canvas);
    228     if (canvas) {
    229         canvas->flush();
    230     }
    231     target->endTiming();
    232     double elapsed = now_ms() - start;
    233     bench->postDraw(canvas);
    234     return elapsed;
    235 }
    236 
    237 static double estimate_timer_overhead() {
    238     double overhead = 0;
    239     for (int i = 0; i < FLAGS_overheadLoops; i++) {
    240         double start = now_ms();
    241         overhead += now_ms() - start;
    242     }
    243     return overhead / FLAGS_overheadLoops;
    244 }
    245 
    246 static int detect_forever_loops(int loops) {
    247     // look for a magic run-forever value
    248     if (loops < 0) {
    249         loops = SK_MaxS32;
    250     }
    251     return loops;
    252 }
    253 
    254 static int clamp_loops(int loops) {
    255     if (loops < 1) {
    256         SkDebugf("ERROR: clamping loops from %d to 1. "
    257                  "There's probably something wrong with the bench.\n", loops);
    258         return 1;
    259     }
    260     if (loops > FLAGS_maxLoops) {
    261         SkDebugf("WARNING: clamping loops from %d to FLAGS_maxLoops, %d.\n", loops, FLAGS_maxLoops);
    262         return FLAGS_maxLoops;
    263     }
    264     return loops;
    265 }
    266 
    267 static bool write_canvas_png(Target* target, const SkString& filename) {
    268 
    269     if (filename.isEmpty()) {
    270         return false;
    271     }
    272     if (target->getCanvas() &&
    273         kUnknown_SkColorType == target->getCanvas()->imageInfo().colorType()) {
    274         return false;
    275     }
    276 
    277     SkBitmap bmp;
    278 
    279     if (!target->capturePixels(&bmp)) {
    280         return false;
    281     }
    282 
    283     SkString dir = SkOSPath::Dirname(filename.c_str());
    284     if (!sk_mkdir(dir.c_str())) {
    285         SkDebugf("Can't make dir %s.\n", dir.c_str());
    286         return false;
    287     }
    288     SkFILEWStream stream(filename.c_str());
    289     if (!stream.isValid()) {
    290         SkDebugf("Can't write %s.\n", filename.c_str());
    291         return false;
    292     }
    293     if (!SkImageEncoder::EncodeStream(&stream, bmp, SkImageEncoder::kPNG_Type, 100)) {
    294         SkDebugf("Can't encode a PNG.\n");
    295         return false;
    296     }
    297     return true;
    298 }
    299 
    300 static int kFailedLoops = -2;
    301 static int setup_cpu_bench(const double overhead, Target* target, Benchmark* bench) {
    302     // First figure out approximately how many loops of bench it takes to make overhead negligible.
    303     double bench_plus_overhead = 0.0;
    304     int round = 0;
    305     int loops = bench->calculateLoops(FLAGS_loops);
    306     if (kAutoTuneLoops == loops) {
    307         while (bench_plus_overhead < overhead) {
    308             if (round++ == FLAGS_maxCalibrationAttempts) {
    309                 SkDebugf("WARNING: Can't estimate loops for %s (%s vs. %s); skipping.\n",
    310                          bench->getUniqueName(), HUMANIZE(bench_plus_overhead), HUMANIZE(overhead));
    311                 return kFailedLoops;
    312             }
    313             bench_plus_overhead = time(1, bench, target);
    314         }
    315     }
    316 
    317     // Later we'll just start and stop the timer once but loop N times.
    318     // We'll pick N to make timer overhead negligible:
    319     //
    320     //          overhead
    321     //  -------------------------  < FLAGS_overheadGoal
    322     //  overhead + N * Bench Time
    323     //
    324     // where bench_plus_overhead  overhead + Bench Time.
    325     //
    326     // Doing some math, we get:
    327     //
    328     //  (overhead / FLAGS_overheadGoal) - overhead
    329     //  ------------------------------------------  < N
    330     //       bench_plus_overhead - overhead)
    331     //
    332     // Luckily, this also works well in practice. :)
    333     if (kAutoTuneLoops == loops) {
    334         const double numer = overhead / FLAGS_overheadGoal - overhead;
    335         const double denom = bench_plus_overhead - overhead;
    336         loops = (int)ceil(numer / denom);
    337         loops = clamp_loops(loops);
    338     } else {
    339         loops = detect_forever_loops(loops);
    340     }
    341 
    342     return loops;
    343 }
    344 
    345 static int setup_gpu_bench(Target* target, Benchmark* bench, int maxGpuFrameLag) {
    346     // First, figure out how many loops it'll take to get a frame up to FLAGS_gpuMs.
    347     int loops = bench->calculateLoops(FLAGS_loops);
    348     if (kAutoTuneLoops == loops) {
    349         loops = 1;
    350         double elapsed = 0;
    351         do {
    352             if (1<<30 == loops) {
    353                 // We're about to wrap.  Something's wrong with the bench.
    354                 loops = 0;
    355                 break;
    356             }
    357             loops *= 2;
    358             // If the GPU lets frames lag at all, we need to make sure we're timing
    359             // _this_ round, not still timing last round.
    360             for (int i = 0; i < maxGpuFrameLag; i++) {
    361                 elapsed = time(loops, bench, target);
    362             }
    363         } while (elapsed < FLAGS_gpuMs);
    364 
    365         // We've overshot at least a little.  Scale back linearly.
    366         loops = (int)ceil(loops * FLAGS_gpuMs / elapsed);
    367         loops = clamp_loops(loops);
    368 
    369         // Make sure we're not still timing our calibration.
    370         target->fence();
    371     } else {
    372         loops = detect_forever_loops(loops);
    373     }
    374 
    375     // Pretty much the same deal as the calibration: do some warmup to make
    376     // sure we're timing steady-state pipelined frames.
    377     for (int i = 0; i < maxGpuFrameLag - 1; i++) {
    378         time(loops, bench, target);
    379     }
    380 
    381     return loops;
    382 }
    383 
    384 #if SK_SUPPORT_GPU
    385 #define kBogusGLContextType GrContextFactory::kNative_GLContextType
    386 #define kBogusGLContextOptions GrContextFactory::kNone_GLContextOptions
    387 #else
    388 #define kBogusGLContextType 0
    389 #define kBogusGLContextOptions 0
    390 #endif
    391 
    392 static void create_config(const SkCommandLineConfig* config, SkTArray<Config>* configs) {
    393 
    394 #if SK_SUPPORT_GPU
    395     if (const auto* gpuConfig = config->asConfigGpu()) {
    396         if (!FLAGS_gpu)
    397             return;
    398 
    399         const auto ctxOptions = gpuConfig->getUseNVPR() ? GrContextFactory::kEnableNVPR_GLContextOptions
    400                                                         : GrContextFactory::kNone_GLContextOptions;
    401         const auto ctxType = gpuConfig->getContextType();
    402         const auto sampleCount = gpuConfig->getSamples();
    403 
    404         if (const GrContext* ctx = gGrFactory->get(ctxType, ctxOptions)) {
    405             const auto maxSampleCount = ctx->caps()->maxSampleCount();
    406             if (sampleCount > ctx->caps()->maxSampleCount()) {
    407                 SkDebugf("Configuration sample count %d exceeds maximum %d.\n",
    408                     sampleCount, maxSampleCount);
    409                 return;
    410             }
    411         } else {
    412             SkDebugf("No context was available matching config type and options.\n");
    413             return;
    414         }
    415 
    416         Config target = {
    417             config->getTag(),
    418             Benchmark::kGPU_Backend,
    419             kN32_SkColorType,
    420             kPremul_SkAlphaType,
    421             sampleCount,
    422             ctxType,
    423             ctxOptions,
    424             false };
    425 
    426         configs->push_back(target);
    427         return;
    428     }
    429 #endif
    430 
    431     #define CPU_CONFIG(name, backend, color, alpha)                                  \
    432         if (config->getTag().equals(#name)) {                                        \
    433             Config config = { SkString(#name), Benchmark::backend, color, alpha, 0,  \
    434                               kBogusGLContextType, kBogusGLContextOptions,           \
    435                               false };                                               \
    436             configs->push_back(config);                                              \
    437             return;                                                                  \
    438         }
    439 
    440     if (FLAGS_cpu) {
    441         CPU_CONFIG(nonrendering, kNonRendering_Backend, kUnknown_SkColorType, kUnpremul_SkAlphaType)
    442         CPU_CONFIG(8888, kRaster_Backend, kN32_SkColorType, kPremul_SkAlphaType)
    443         CPU_CONFIG(565, kRaster_Backend, kRGB_565_SkColorType, kOpaque_SkAlphaType)
    444     }
    445 
    446     #undef CPU_CONFIG
    447 
    448 #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK
    449     if (config->getTag().equals("hwui")) {
    450         Config config = { SkString("hwui"), Benchmark::kHWUI_Backend, kRGBA_8888_SkColorType,
    451                           kPremul_SkAlphaType, 0, kBogusGLContextType, kBogusGLContextOptions,
    452                           false };
    453         configs->push_back(config);
    454     }
    455 #endif
    456 }
    457 
    458 // Append all configs that are enabled and supported.
    459 void create_configs(SkTArray<Config>* configs) {
    460     SkCommandLineConfigArray array;
    461     ParseConfigs(FLAGS_config, &array);
    462     for (int i = 0; i < array.count(); ++i) {
    463         create_config(array[i], configs);
    464     }
    465 }
    466 
    467 // If bench is enabled for config, returns a Target* for it, otherwise nullptr.
    468 static Target* is_enabled(Benchmark* bench, const Config& config) {
    469     if (!bench->isSuitableFor(config.backend)) {
    470         return nullptr;
    471     }
    472 
    473     SkImageInfo info = SkImageInfo::Make(bench->getSize().fX, bench->getSize().fY,
    474                                          config.color, config.alpha);
    475 
    476     Target* target = nullptr;
    477 
    478     switch (config.backend) {
    479 #if SK_SUPPORT_GPU
    480     case Benchmark::kGPU_Backend:
    481         target = new GPUTarget(config);
    482         break;
    483 #endif
    484 #ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK
    485     case Benchmark::kHWUI_Backend:
    486         target = new HWUITarget(config, bench);
    487         break;
    488 #endif
    489     default:
    490         target = new Target(config);
    491         break;
    492     }
    493 
    494     if (!target->init(info, bench)) {
    495         delete target;
    496         return nullptr;
    497     }
    498     return target;
    499 }
    500 
    501 static bool valid_brd_bench(SkData* encoded, SkBitmapRegionDecoder::Strategy strategy,
    502         SkColorType colorType, uint32_t sampleSize, uint32_t minOutputSize, int* width,
    503         int* height) {
    504     SkAutoTDelete<SkBitmapRegionDecoder> brd(
    505             SkBitmapRegionDecoder::Create(encoded, strategy));
    506     if (nullptr == brd.get()) {
    507         // This is indicates that subset decoding is not supported for a particular image format.
    508         return false;
    509     }
    510 
    511     SkBitmap bitmap;
    512     if (!brd->decodeRegion(&bitmap, nullptr, SkIRect::MakeXYWH(0, 0, brd->width(), brd->height()),
    513             1, colorType, false)) {
    514         return false;
    515     }
    516 
    517     if (sampleSize * minOutputSize > (uint32_t) brd->width() || sampleSize * minOutputSize >
    518             (uint32_t) brd->height()) {
    519         // This indicates that the image is not large enough to decode a
    520         // minOutputSize x minOutputSize subset at the given sampleSize.
    521         return false;
    522     }
    523 
    524     // Set the image width and height.  The calling code will use this to choose subsets to decode.
    525     *width = brd->width();
    526     *height = brd->height();
    527     return true;
    528 }
    529 
    530 static void cleanup_run(Target* target) {
    531     delete target;
    532 #if SK_SUPPORT_GPU
    533     if (FLAGS_abandonGpuContext) {
    534         gGrFactory->abandonContexts();
    535     }
    536     if (FLAGS_resetGpuContext || FLAGS_abandonGpuContext) {
    537         gGrFactory->destroyContexts();
    538     }
    539 #endif
    540 }
    541 
    542 class BenchmarkStream {
    543 public:
    544     BenchmarkStream() : fBenches(BenchRegistry::Head())
    545                       , fGMs(skiagm::GMRegistry::Head())
    546                       , fCurrentRecording(0)
    547                       , fCurrentScale(0)
    548                       , fCurrentSKP(0)
    549                       , fCurrentUseMPD(0)
    550                       , fCurrentCodec(0)
    551                       , fCurrentAndroidCodec(0)
    552                       , fCurrentBRDImage(0)
    553                       , fCurrentColorType(0)
    554                       , fCurrentAlphaType(0)
    555                       , fCurrentSubsetType(0)
    556                       , fCurrentBRDStrategy(0)
    557                       , fCurrentSampleSize(0)
    558                       , fCurrentAnimSKP(0) {
    559         for (int i = 0; i < FLAGS_skps.count(); i++) {
    560             if (SkStrEndsWith(FLAGS_skps[i], ".skp")) {
    561                 fSKPs.push_back() = FLAGS_skps[i];
    562             } else {
    563                 SkOSFile::Iter it(FLAGS_skps[i], ".skp");
    564                 SkString path;
    565                 while (it.next(&path)) {
    566                     fSKPs.push_back() = SkOSPath::Join(FLAGS_skps[0], path.c_str());
    567                 }
    568             }
    569         }
    570 
    571         if (4 != sscanf(FLAGS_clip[0], "%d,%d,%d,%d",
    572                         &fClip.fLeft, &fClip.fTop, &fClip.fRight, &fClip.fBottom)) {
    573             SkDebugf("Can't parse %s from --clip as an SkIRect.\n", FLAGS_clip[0]);
    574             exit(1);
    575         }
    576 
    577         for (int i = 0; i < FLAGS_scales.count(); i++) {
    578             if (1 != sscanf(FLAGS_scales[i], "%f", &fScales.push_back())) {
    579                 SkDebugf("Can't parse %s from --scales as an SkScalar.\n", FLAGS_scales[i]);
    580                 exit(1);
    581             }
    582         }
    583 
    584         if (2 != sscanf(FLAGS_zoom[0], "%f,%lf", &fZoomMax, &fZoomPeriodMs)) {
    585             SkDebugf("Can't parse %s from --zoom as a zoomMax,zoomPeriodMs.\n", FLAGS_zoom[0]);
    586             exit(1);
    587         }
    588 
    589         if (FLAGS_mpd) {
    590             fUseMPDs.push_back() = true;
    591         }
    592         fUseMPDs.push_back() = false;
    593 
    594         // Prepare the images for decoding
    595         if (!CollectImages(&fImages)) {
    596             exit(1);
    597         }
    598 
    599         // Choose the candidate color types for image decoding
    600         const SkColorType colorTypes[] =
    601             { kN32_SkColorType,
    602               kRGB_565_SkColorType,
    603               kAlpha_8_SkColorType,
    604               kIndex_8_SkColorType,
    605               kGray_8_SkColorType };
    606         fColorTypes.reset(colorTypes, SK_ARRAY_COUNT(colorTypes));
    607     }
    608 
    609     static bool ReadPicture(const char* path, SkAutoTUnref<SkPicture>* pic) {
    610         // Not strictly necessary, as it will be checked again later,
    611         // but helps to avoid a lot of pointless work if we're going to skip it.
    612         if (SkCommandLineFlags::ShouldSkip(FLAGS_match, SkOSPath::Basename(path).c_str())) {
    613             return false;
    614         }
    615 
    616         SkAutoTDelete<SkStream> stream(SkStream::NewFromFile(path));
    617         if (stream.get() == nullptr) {
    618             SkDebugf("Could not read %s.\n", path);
    619             return false;
    620         }
    621 
    622         pic->reset(SkPicture::CreateFromStream(stream.get()));
    623         if (pic->get() == nullptr) {
    624             SkDebugf("Could not read %s as an SkPicture.\n", path);
    625             return false;
    626         }
    627         return true;
    628     }
    629 
    630     Benchmark* next() {
    631         SkAutoTDelete<Benchmark> bench;
    632         do {
    633             bench.reset(this->rawNext());
    634             if (!bench) {
    635                 return nullptr;
    636             }
    637         } while(SkCommandLineFlags::ShouldSkip(FLAGS_sourceType, fSourceType) ||
    638                 SkCommandLineFlags::ShouldSkip(FLAGS_benchType,  fBenchType));
    639         return bench.detach();
    640     }
    641 
    642     Benchmark* rawNext() {
    643         if (fBenches) {
    644             Benchmark* bench = fBenches->factory()(nullptr);
    645             fBenches = fBenches->next();
    646             fSourceType = "bench";
    647             fBenchType  = "micro";
    648             return bench;
    649         }
    650 
    651         while (fGMs) {
    652             SkAutoTDelete<skiagm::GM> gm(fGMs->factory()(nullptr));
    653             fGMs = fGMs->next();
    654             if (gm->runAsBench()) {
    655                 fSourceType = "gm";
    656                 fBenchType  = "micro";
    657                 return new GMBench(gm.detach());
    658             }
    659         }
    660 
    661         // First add all .skps as RecordingBenches.
    662         while (fCurrentRecording < fSKPs.count()) {
    663             const SkString& path = fSKPs[fCurrentRecording++];
    664             SkAutoTUnref<SkPicture> pic;
    665             if (!ReadPicture(path.c_str(), &pic)) {
    666                 continue;
    667             }
    668             SkString name = SkOSPath::Basename(path.c_str());
    669             fSourceType = "skp";
    670             fBenchType  = "recording";
    671             fSKPBytes = static_cast<double>(SkPictureUtils::ApproximateBytesUsed(pic));
    672             fSKPOps   = pic->approximateOpCount();
    673             return new RecordingBench(name.c_str(), pic.get(), FLAGS_bbh);
    674         }
    675 
    676         // Then once each for each scale as SKPBenches (playback).
    677         while (fCurrentScale < fScales.count()) {
    678             while (fCurrentSKP < fSKPs.count()) {
    679                 const SkString& path = fSKPs[fCurrentSKP];
    680                 SkAutoTUnref<SkPicture> pic;
    681                 if (!ReadPicture(path.c_str(), &pic)) {
    682                     fCurrentSKP++;
    683                     continue;
    684                 }
    685 
    686                 while (fCurrentUseMPD < fUseMPDs.count()) {
    687                     if (FLAGS_bbh) {
    688                         // The SKP we read off disk doesn't have a BBH.  Re-record so it grows one.
    689                         SkRTreeFactory factory;
    690                         SkPictureRecorder recorder;
    691                         static const int kFlags = SkPictureRecorder::kComputeSaveLayerInfo_RecordFlag;
    692                         pic->playback(recorder.beginRecording(pic->cullRect().width(),
    693                                                               pic->cullRect().height(),
    694                                                               &factory,
    695                                                               fUseMPDs[fCurrentUseMPD] ? kFlags : 0));
    696                         pic.reset(recorder.endRecording());
    697                     }
    698                     SkString name = SkOSPath::Basename(path.c_str());
    699                     fSourceType = "skp";
    700                     fBenchType = "playback";
    701                     return new SKPBench(name.c_str(), pic.get(), fClip, fScales[fCurrentScale],
    702                                         fUseMPDs[fCurrentUseMPD++], FLAGS_loopSKP);
    703                 }
    704                 fCurrentUseMPD = 0;
    705                 fCurrentSKP++;
    706             }
    707             fCurrentSKP = 0;
    708             fCurrentScale++;
    709         }
    710 
    711         // Now loop over each skp again if we have an animation
    712         if (fZoomMax != 1.0f && fZoomPeriodMs > 0) {
    713             while (fCurrentAnimSKP < fSKPs.count()) {
    714                 const SkString& path = fSKPs[fCurrentAnimSKP];
    715                 SkAutoTUnref<SkPicture> pic;
    716                 if (!ReadPicture(path.c_str(), &pic)) {
    717                     fCurrentAnimSKP++;
    718                     continue;
    719                 }
    720 
    721                 fCurrentAnimSKP++;
    722                 SkString name = SkOSPath::Basename(path.c_str());
    723                 SkAutoTUnref<SKPAnimationBench::Animation> animation(
    724                     SKPAnimationBench::CreateZoomAnimation(fZoomMax, fZoomPeriodMs));
    725                 return new SKPAnimationBench(name.c_str(), pic.get(), fClip, animation,
    726                                              FLAGS_loopSKP);
    727             }
    728         }
    729 
    730         for (; fCurrentCodec < fImages.count(); fCurrentCodec++) {
    731             fSourceType = "image";
    732             fBenchType = "skcodec";
    733             const SkString& path = fImages[fCurrentCodec];
    734             if (SkCommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
    735                 continue;
    736             }
    737             SkAutoTUnref<SkData> encoded(SkData::NewFromFileName(path.c_str()));
    738             SkAutoTDelete<SkCodec> codec(SkCodec::NewFromData(encoded));
    739             if (!codec) {
    740                 // Nothing to time.
    741                 SkDebugf("Cannot find codec for %s\n", path.c_str());
    742                 continue;
    743             }
    744 
    745             while (fCurrentColorType < fColorTypes.count()) {
    746                 const SkColorType colorType = fColorTypes[fCurrentColorType];
    747 
    748                 SkAlphaType alphaType = codec->getInfo().alphaType();
    749                 switch (alphaType) {
    750                     case kOpaque_SkAlphaType:
    751                         // We only need to test one alpha type (opaque).
    752                         fCurrentColorType++;
    753                         break;
    754                     case kUnpremul_SkAlphaType:
    755                     case kPremul_SkAlphaType:
    756                         if (0 == fCurrentAlphaType) {
    757                             // Test unpremul first.
    758                             alphaType = kUnpremul_SkAlphaType;
    759                             fCurrentAlphaType++;
    760                         } else {
    761                             // Test premul.
    762                             alphaType = kPremul_SkAlphaType;
    763                             fCurrentAlphaType = 0;
    764                             fCurrentColorType++;
    765                         }
    766                         break;
    767                     default:
    768                         SkASSERT(false);
    769                         fCurrentColorType++;
    770                         break;
    771                 }
    772 
    773                 // Make sure we can decode to this color type and alpha type.
    774                 SkImageInfo info =
    775                         codec->getInfo().makeColorType(colorType).makeAlphaType(alphaType);
    776                 const size_t rowBytes = info.minRowBytes();
    777                 SkAutoMalloc storage(info.getSafeSize(rowBytes));
    778 
    779                 // Used if fCurrentColorType is kIndex_8_SkColorType
    780                 int colorCount = 256;
    781                 SkPMColor colors[256];
    782 
    783                 const SkCodec::Result result = codec->getPixels(
    784                         info, storage.get(), rowBytes, nullptr, colors,
    785                         &colorCount);
    786                 switch (result) {
    787                     case SkCodec::kSuccess:
    788                     case SkCodec::kIncompleteInput:
    789                         return new CodecBench(SkOSPath::Basename(path.c_str()),
    790                                 encoded, colorType, alphaType);
    791                     case SkCodec::kInvalidConversion:
    792                         // This is okay. Not all conversions are valid.
    793                         break;
    794                     default:
    795                         // This represents some sort of failure.
    796                         SkASSERT(false);
    797                         break;
    798                 }
    799             }
    800             fCurrentColorType = 0;
    801         }
    802 
    803         // Run AndroidCodecBenches
    804         const int sampleSizes[] = { 2, 4, 8 };
    805         for (; fCurrentAndroidCodec < fImages.count(); fCurrentAndroidCodec++) {
    806             fSourceType = "image";
    807             fBenchType = "skandroidcodec";
    808 
    809             const SkString& path = fImages[fCurrentAndroidCodec];
    810             if (SkCommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
    811                 continue;
    812             }
    813             SkAutoTUnref<SkData> encoded(SkData::NewFromFileName(path.c_str()));
    814             SkAutoTDelete<SkAndroidCodec> codec(SkAndroidCodec::NewFromData(encoded));
    815             if (!codec) {
    816                 // Nothing to time.
    817                 SkDebugf("Cannot find codec for %s\n", path.c_str());
    818                 continue;
    819             }
    820 
    821             while (fCurrentSampleSize < (int) SK_ARRAY_COUNT(sampleSizes)) {
    822                 int sampleSize = sampleSizes[fCurrentSampleSize];
    823                 fCurrentSampleSize++;
    824                 if (10 * sampleSize > SkTMin(codec->getInfo().width(), codec->getInfo().height())) {
    825                     // Avoid benchmarking scaled decodes of already small images.
    826                     break;
    827                 }
    828 
    829                 return new AndroidCodecBench(SkOSPath::Basename(path.c_str()), encoded, sampleSize);
    830             }
    831             fCurrentSampleSize = 0;
    832         }
    833 
    834         // Run the BRDBenches
    835         // We will benchmark multiple BRD strategies.
    836         static const struct {
    837             SkBitmapRegionDecoder::Strategy    fStrategy;
    838             const char*                        fName;
    839         } strategies[] = {
    840             { SkBitmapRegionDecoder::kCanvas_Strategy,       "BRD_canvas" },
    841             { SkBitmapRegionDecoder::kAndroidCodec_Strategy, "BRD_android_codec" },
    842         };
    843 
    844         // We intend to create benchmarks that model the use cases in
    845         // android/libraries/social/tiledimage.  In this library, an image is decoded in 512x512
    846         // tiles.  The image can be translated freely, so the location of a tile may be anywhere in
    847         // the image.  For that reason, we will benchmark decodes in five representative locations
    848         // in the image.  Additionally, this use case utilizes power of two scaling, so we will
    849         // test on power of two sample sizes.  The output tile is always 512x512, so, when a
    850         // sampleSize is used, the size of the subset that is decoded is always
    851         // (sampleSize*512)x(sampleSize*512).
    852         // There are a few good reasons to only test on power of two sample sizes at this time:
    853         //     All use cases we are aware of only scale by powers of two.
    854         //     PNG decodes use the indicated sampling strategy regardless of the sample size, so
    855         //         these tests are sufficient to provide good coverage of our scaling options.
    856         const uint32_t brdSampleSizes[] = { 1, 2, 4, 8, 16 };
    857         const uint32_t minOutputSize = 512;
    858         for (; fCurrentBRDImage < fImages.count(); fCurrentBRDImage++) {
    859             const SkString& path = fImages[fCurrentBRDImage];
    860             if (SkCommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
    861                 continue;
    862             }
    863             while (fCurrentBRDStrategy < (int) SK_ARRAY_COUNT(strategies)) {
    864                 fSourceType = "image";
    865                 fBenchType = strategies[fCurrentBRDStrategy].fName;
    866 
    867                 const SkBitmapRegionDecoder::Strategy strategy =
    868                         strategies[fCurrentBRDStrategy].fStrategy;
    869 
    870                 while (fCurrentColorType < fColorTypes.count()) {
    871                     while (fCurrentSampleSize < (int) SK_ARRAY_COUNT(brdSampleSizes)) {
    872                         while (fCurrentSubsetType <= kLastSingle_SubsetType) {
    873 
    874                             SkAutoTUnref<SkData> encoded(SkData::NewFromFileName(path.c_str()));
    875                             const SkColorType colorType = fColorTypes[fCurrentColorType];
    876                             uint32_t sampleSize = brdSampleSizes[fCurrentSampleSize];
    877                             int currentSubsetType = fCurrentSubsetType++;
    878 
    879                             int width = 0;
    880                             int height = 0;
    881                             if (!valid_brd_bench(encoded.get(), strategy, colorType, sampleSize,
    882                                     minOutputSize, &width, &height)) {
    883                                 break;
    884                             }
    885 
    886                             SkString basename = SkOSPath::Basename(path.c_str());
    887                             SkIRect subset;
    888                             const uint32_t subsetSize = sampleSize * minOutputSize;
    889                             switch (currentSubsetType) {
    890                                 case kTopLeft_SubsetType:
    891                                     basename.append("_TopLeft");
    892                                     subset = SkIRect::MakeXYWH(0, 0, subsetSize, subsetSize);
    893                                     break;
    894                                 case kTopRight_SubsetType:
    895                                     basename.append("_TopRight");
    896                                     subset = SkIRect::MakeXYWH(width - subsetSize, 0, subsetSize,
    897                                             subsetSize);
    898                                     break;
    899                                 case kMiddle_SubsetType:
    900                                     basename.append("_Middle");
    901                                     subset = SkIRect::MakeXYWH((width - subsetSize) / 2,
    902                                             (height - subsetSize) / 2, subsetSize, subsetSize);
    903                                     break;
    904                                 case kBottomLeft_SubsetType:
    905                                     basename.append("_BottomLeft");
    906                                     subset = SkIRect::MakeXYWH(0, height - subsetSize, subsetSize,
    907                                             subsetSize);
    908                                     break;
    909                                 case kBottomRight_SubsetType:
    910                                     basename.append("_BottomRight");
    911                                     subset = SkIRect::MakeXYWH(width - subsetSize,
    912                                             height - subsetSize, subsetSize, subsetSize);
    913                                     break;
    914                                 default:
    915                                     SkASSERT(false);
    916                             }
    917 
    918                             return new BitmapRegionDecoderBench(basename.c_str(), encoded.get(),
    919                                     strategy, colorType, sampleSize, subset);
    920                         }
    921                         fCurrentSubsetType = 0;
    922                         fCurrentSampleSize++;
    923                     }
    924                     fCurrentSampleSize = 0;
    925                     fCurrentColorType++;
    926                 }
    927                 fCurrentColorType = 0;
    928                 fCurrentBRDStrategy++;
    929             }
    930             fCurrentBRDStrategy = 0;
    931         }
    932 
    933         return nullptr;
    934     }
    935 
    936     void fillCurrentOptions(ResultsWriter* log) const {
    937         log->configOption("source_type", fSourceType);
    938         log->configOption("bench_type",  fBenchType);
    939         if (0 == strcmp(fSourceType, "skp")) {
    940             log->configOption("clip",
    941                     SkStringPrintf("%d %d %d %d", fClip.fLeft, fClip.fTop,
    942                                                   fClip.fRight, fClip.fBottom).c_str());
    943             SkASSERT_RELEASE(fCurrentScale < fScales.count());  // debugging paranoia
    944             log->configOption("scale", SkStringPrintf("%.2g", fScales[fCurrentScale]).c_str());
    945             if (fCurrentUseMPD > 0) {
    946                 SkASSERT(1 == fCurrentUseMPD || 2 == fCurrentUseMPD);
    947                 log->configOption("multi_picture_draw", fUseMPDs[fCurrentUseMPD-1] ? "true" : "false");
    948             }
    949         }
    950         if (0 == strcmp(fBenchType, "recording")) {
    951             log->metric("bytes", fSKPBytes);
    952             log->metric("ops",   fSKPOps);
    953         }
    954     }
    955 
    956 private:
    957     enum SubsetType {
    958         kTopLeft_SubsetType     = 0,
    959         kTopRight_SubsetType    = 1,
    960         kMiddle_SubsetType      = 2,
    961         kBottomLeft_SubsetType  = 3,
    962         kBottomRight_SubsetType = 4,
    963         kTranslate_SubsetType   = 5,
    964         kZoom_SubsetType        = 6,
    965         kLast_SubsetType        = kZoom_SubsetType,
    966         kLastSingle_SubsetType  = kBottomRight_SubsetType,
    967     };
    968 
    969     const BenchRegistry* fBenches;
    970     const skiagm::GMRegistry* fGMs;
    971     SkIRect            fClip;
    972     SkTArray<SkScalar> fScales;
    973     SkTArray<SkString> fSKPs;
    974     SkTArray<bool>     fUseMPDs;
    975     SkTArray<SkString> fImages;
    976     SkTArray<SkColorType, true> fColorTypes;
    977     SkScalar           fZoomMax;
    978     double             fZoomPeriodMs;
    979 
    980     double fSKPBytes, fSKPOps;
    981 
    982     const char* fSourceType;  // What we're benching: bench, GM, SKP, ...
    983     const char* fBenchType;   // How we bench it: micro, recording, playback, ...
    984     int fCurrentRecording;
    985     int fCurrentScale;
    986     int fCurrentSKP;
    987     int fCurrentUseMPD;
    988     int fCurrentCodec;
    989     int fCurrentAndroidCodec;
    990     int fCurrentBRDImage;
    991     int fCurrentColorType;
    992     int fCurrentAlphaType;
    993     int fCurrentSubsetType;
    994     int fCurrentBRDStrategy;
    995     int fCurrentSampleSize;
    996     int fCurrentAnimSKP;
    997 };
    998 
    999 // Some runs (mostly, Valgrind) are so slow that the bot framework thinks we've hung.
   1000 // This prints something every once in a while so that it knows we're still working.
   1001 static void start_keepalive() {
   1002     struct Loop {
   1003         static void forever(void*) {
   1004             for (;;) {
   1005                 static const int kSec = 1200;
   1006             #if defined(SK_BUILD_FOR_WIN)
   1007                 Sleep(kSec * 1000);
   1008             #else
   1009                 sleep(kSec);
   1010             #endif
   1011                 SkDebugf("\nBenchmarks still running...\n");
   1012             }
   1013         }
   1014     };
   1015     static SkThread* intentionallyLeaked = new SkThread(Loop::forever);
   1016     intentionallyLeaked->start();
   1017 }
   1018 
   1019 int nanobench_main();
   1020 int nanobench_main() {
   1021     SetupCrashHandler();
   1022     SkAutoGraphics ag;
   1023     SkTaskGroup::Enabler enabled(FLAGS_threads);
   1024 
   1025 #if SK_SUPPORT_GPU
   1026     GrContextOptions grContextOpts;
   1027     grContextOpts.fDrawPathToCompressedTexture = FLAGS_gpuCompressAlphaMasks;
   1028     gGrFactory.reset(new GrContextFactory(grContextOpts));
   1029 #endif
   1030 
   1031     if (FLAGS_veryVerbose) {
   1032         FLAGS_verbose = true;
   1033     }
   1034 
   1035     if (kAutoTuneLoops != FLAGS_loops) {
   1036         FLAGS_samples     = 1;
   1037         FLAGS_gpuFrameLag = 0;
   1038     }
   1039 
   1040     if (!FLAGS_writePath.isEmpty()) {
   1041         SkDebugf("Writing files to %s.\n", FLAGS_writePath[0]);
   1042         if (!sk_mkdir(FLAGS_writePath[0])) {
   1043             SkDebugf("Could not create %s. Files won't be written.\n", FLAGS_writePath[0]);
   1044             FLAGS_writePath.set(0, nullptr);
   1045         }
   1046     }
   1047 
   1048     SkAutoTDelete<ResultsWriter> log(new ResultsWriter);
   1049     if (!FLAGS_outResultsFile.isEmpty()) {
   1050 #if defined(SK_RELEASE)
   1051         log.reset(new NanoJSONResultsWriter(FLAGS_outResultsFile[0]));
   1052 #else
   1053         SkDebugf("I'm ignoring --outResultsFile because this is a Debug build.");
   1054         return 1;
   1055 #endif
   1056     }
   1057 
   1058     if (1 == FLAGS_properties.count() % 2) {
   1059         SkDebugf("ERROR: --properties must be passed with an even number of arguments.\n");
   1060         return 1;
   1061     }
   1062     for (int i = 1; i < FLAGS_properties.count(); i += 2) {
   1063         log->property(FLAGS_properties[i-1], FLAGS_properties[i]);
   1064     }
   1065 
   1066     if (1 == FLAGS_key.count() % 2) {
   1067         SkDebugf("ERROR: --key must be passed with an even number of arguments.\n");
   1068         return 1;
   1069     }
   1070     for (int i = 1; i < FLAGS_key.count(); i += 2) {
   1071         log->key(FLAGS_key[i-1], FLAGS_key[i]);
   1072     }
   1073 
   1074     const double overhead = estimate_timer_overhead();
   1075     SkDebugf("Timer overhead: %s\n", HUMANIZE(overhead));
   1076 
   1077     SkTArray<double> samples;
   1078 
   1079     if (kAutoTuneLoops != FLAGS_loops) {
   1080         SkDebugf("Fixed number of loops; times would only be misleading so we won't print them.\n");
   1081     } else if (FLAGS_quiet) {
   1082         SkDebugf("! -> high variance, ? -> moderate variance\n");
   1083         SkDebugf("    micros   \tbench\n");
   1084     } else if (FLAGS_ms) {
   1085         SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\tsamples\tconfig\tbench\n");
   1086     } else {
   1087         SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\t%-*s\tconfig\tbench\n",
   1088                  FLAGS_samples, "samples");
   1089     }
   1090 
   1091     SkTArray<Config> configs;
   1092     create_configs(&configs);
   1093 
   1094 #ifdef THERMAL_MANAGER_SUPPORTED
   1095     int tmEnabled, tmThreshold, tmSleepTimeMs, tmTimeoutMs;
   1096     if (4 != sscanf(FLAGS_useThermalManager[0], "%d,%d,%d,%d",
   1097                     &tmEnabled, &tmThreshold, &tmSleepTimeMs, &tmTimeoutMs)) {
   1098         SkDebugf("Can't parse %s from --useThermalManager.\n", FLAGS_useThermalManager[0]);
   1099         exit(1);
   1100     }
   1101     ThermalManager tm(tmThreshold, tmSleepTimeMs, tmTimeoutMs);
   1102 #endif
   1103 
   1104     if (FLAGS_keepAlive) {
   1105         start_keepalive();
   1106     }
   1107 
   1108     int runs = 0;
   1109     BenchmarkStream benchStream;
   1110     while (Benchmark* b = benchStream.next()) {
   1111         SkAutoTDelete<Benchmark> bench(b);
   1112         if (SkCommandLineFlags::ShouldSkip(FLAGS_match, bench->getUniqueName())) {
   1113             continue;
   1114         }
   1115 
   1116         if (!configs.empty()) {
   1117             log->bench(bench->getUniqueName(), bench->getSize().fX, bench->getSize().fY);
   1118             bench->delayedSetup();
   1119         }
   1120         for (int i = 0; i < configs.count(); ++i) {
   1121 #ifdef THERMAL_MANAGER_SUPPORTED
   1122             if (tmEnabled && !tm.coolOffIfNecessary()) {
   1123                 SkDebugf("Could not cool off, timings will be throttled\n");
   1124             }
   1125 #endif
   1126             Target* target = is_enabled(b, configs[i]);
   1127             if (!target) {
   1128                 continue;
   1129             }
   1130 
   1131             // During HWUI output this canvas may be nullptr.
   1132             SkCanvas* canvas = target->getCanvas();
   1133             const char* config = target->config.name.c_str();
   1134 
   1135             if (FLAGS_pre_log || FLAGS_dryRun) {
   1136                 SkDebugf("Running %s\t%s\n"
   1137                          , bench->getUniqueName()
   1138                          , config);
   1139                 if (FLAGS_dryRun) {
   1140                     continue;
   1141                 }
   1142             }
   1143 
   1144             target->setup();
   1145             bench->perCanvasPreDraw(canvas);
   1146 
   1147             int maxFrameLag;
   1148             int loops = target->needsFrameTiming(&maxFrameLag)
   1149                 ? setup_gpu_bench(target, bench.get(), maxFrameLag)
   1150                 : setup_cpu_bench(overhead, target, bench.get());
   1151 
   1152             if (FLAGS_ms) {
   1153                 samples.reset();
   1154                 auto stop = now_ms() + FLAGS_ms;
   1155                 do {
   1156                     samples.push_back(time(loops, bench, target) / loops);
   1157                 } while (now_ms() < stop);
   1158             } else {
   1159                 samples.reset(FLAGS_samples);
   1160                 for (int s = 0; s < FLAGS_samples; s++) {
   1161                     samples[s] = time(loops, bench, target) / loops;
   1162                 }
   1163             }
   1164 
   1165 #if SK_SUPPORT_GPU
   1166             SkTArray<SkString> keys;
   1167             SkTArray<double> values;
   1168             bool gpuStatsDump = FLAGS_gpuStatsDump && Benchmark::kGPU_Backend == configs[i].backend;
   1169             if (gpuStatsDump) {
   1170                 // TODO cache stats
   1171                 bench->getGpuStats(canvas, &keys, &values);
   1172             }
   1173 #endif
   1174 
   1175             bench->perCanvasPostDraw(canvas);
   1176 
   1177             if (Benchmark::kNonRendering_Backend != target->config.backend &&
   1178                 !FLAGS_writePath.isEmpty() && FLAGS_writePath[0]) {
   1179                 SkString pngFilename = SkOSPath::Join(FLAGS_writePath[0], config);
   1180                 pngFilename = SkOSPath::Join(pngFilename.c_str(), bench->getUniqueName());
   1181                 pngFilename.append(".png");
   1182                 write_canvas_png(target, pngFilename);
   1183             }
   1184 
   1185             if (kFailedLoops == loops) {
   1186                 // Can't be timed.  A warning note has already been printed.
   1187                 cleanup_run(target);
   1188                 continue;
   1189             }
   1190 
   1191             Stats stats(samples);
   1192             log->config(config);
   1193             log->configOption("name", bench->getName());
   1194             benchStream.fillCurrentOptions(log.get());
   1195             target->fillOptions(log.get());
   1196             log->metric("min_ms",    stats.min);
   1197             log->metric("median_ms", stats.median);
   1198 #if SK_SUPPORT_GPU
   1199             if (gpuStatsDump) {
   1200                 // dump to json, only SKPBench currently returns valid keys / values
   1201                 SkASSERT(keys.count() == values.count());
   1202                 for (int i = 0; i < keys.count(); i++) {
   1203                     log->metric(keys[i].c_str(), values[i]);
   1204                 }
   1205             }
   1206 #endif
   1207 
   1208             if (runs++ % FLAGS_flushEvery == 0) {
   1209                 log->flush();
   1210             }
   1211 
   1212             if (kAutoTuneLoops != FLAGS_loops) {
   1213                 if (configs.count() == 1) {
   1214                     config = ""; // Only print the config if we run the same bench on more than one.
   1215                 }
   1216                 SkDebugf("%4d/%-4dMB\t%s\t%s\n"
   1217                          , sk_tools::getCurrResidentSetSizeMB()
   1218                          , sk_tools::getMaxResidentSetSizeMB()
   1219                          , bench->getUniqueName()
   1220                          , config);
   1221             } else if (FLAGS_quiet) {
   1222                 const char* mark = " ";
   1223                 const double stddev_percent = 100 * sqrt(stats.var) / stats.mean;
   1224                 if (stddev_percent >  5) mark = "?";
   1225                 if (stddev_percent > 10) mark = "!";
   1226 
   1227                 SkDebugf("%10.2f %s\t%s\t%s\n",
   1228                          stats.median*1e3, mark, bench->getUniqueName(), config);
   1229             } else {
   1230                 const double stddev_percent = 100 * sqrt(stats.var) / stats.mean;
   1231                 SkDebugf("%4d/%-4dMB\t%d\t%s\t%s\t%s\t%s\t%.0f%%\t%s\t%s\t%s\n"
   1232                         , sk_tools::getCurrResidentSetSizeMB()
   1233                         , sk_tools::getMaxResidentSetSizeMB()
   1234                         , loops
   1235                         , HUMANIZE(stats.min)
   1236                         , HUMANIZE(stats.median)
   1237                         , HUMANIZE(stats.mean)
   1238                         , HUMANIZE(stats.max)
   1239                         , stddev_percent
   1240                         , FLAGS_ms ? to_string(samples.count()).c_str() : stats.plot.c_str()
   1241                         , config
   1242                         , bench->getUniqueName()
   1243                         );
   1244             }
   1245 
   1246 #if SK_SUPPORT_GPU
   1247             if (FLAGS_gpuStats && Benchmark::kGPU_Backend == configs[i].backend) {
   1248                 GrContext* context = gGrFactory->get(configs[i].ctxType,
   1249                                                      configs[i].ctxOptions);
   1250                 context->printCacheStats();
   1251                 context->printGpuStats();
   1252             }
   1253 #endif
   1254 
   1255             if (FLAGS_verbose) {
   1256                 SkDebugf("Samples:  ");
   1257                 for (int i = 0; i < samples.count(); i++) {
   1258                     SkDebugf("%s  ", HUMANIZE(samples[i]));
   1259                 }
   1260                 SkDebugf("%s\n", bench->getUniqueName());
   1261             }
   1262             cleanup_run(target);
   1263         }
   1264     }
   1265 
   1266     log->bench("memory_usage", 0,0);
   1267     log->config("meta");
   1268     log->metric("max_rss_mb", sk_tools::getMaxResidentSetSizeMB());
   1269 
   1270 #if SK_SUPPORT_GPU
   1271     // Make sure we clean up the global GrContextFactory here, otherwise we might race with the
   1272     // SkEventTracer destructor
   1273     gGrFactory.reset(nullptr);
   1274 #endif
   1275 
   1276     return 0;
   1277 }
   1278 
   1279 #if !defined SK_BUILD_FOR_IOS
   1280 int main(int argc, char** argv) {
   1281     SkCommandLineFlags::Parse(argc, argv);
   1282     return nanobench_main();
   1283 }
   1284 #endif
   1285