xref: /aosp_15_r20/external/skia/bench/nanobench.cpp (revision c8dee2aa9b3f27cf6c858bd81872bdeb2c07ed17)
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 "bench/nanobench.h"
11 
12 #include "bench/AndroidCodecBench.h"
13 #include "bench/Benchmark.h"
14 #include "bench/CodecBench.h"
15 #include "bench/CodecBenchPriv.h"
16 #include "bench/GMBench.h"
17 #include "bench/MSKPBench.h"
18 #include "bench/RecordingBench.h"
19 #include "bench/ResultsWriter.h"
20 #include "bench/SKPAnimationBench.h"
21 #include "bench/SKPBench.h"
22 #include "bench/SkGlyphCacheBench.h"
23 #include "bench/SkSLBench.h"
24 #include "include/codec/SkAndroidCodec.h"
25 #include "include/codec/SkCodec.h"
26 #include "include/codec/SkJpegDecoder.h"
27 #include "include/codec/SkPngDecoder.h"
28 #include "include/core/SkBBHFactory.h"
29 #include "include/core/SkCanvas.h"
30 #include "include/core/SkData.h"
31 #include "include/core/SkGraphics.h"
32 #include "include/core/SkPictureRecorder.h"
33 #include "include/core/SkString.h"
34 #include "include/core/SkSurface.h"
35 #include "include/encode/SkPngEncoder.h"
36 #include "include/private/base/SkMacros.h"
37 #include "src/base/SkAutoMalloc.h"
38 #include "src/base/SkLeanWindows.h"
39 #include "src/base/SkTime.h"
40 #include "src/core/SkColorSpacePriv.h"
41 #include "src/core/SkOSFile.h"
42 #include "src/core/SkTaskGroup.h"
43 #include "src/core/SkTraceEvent.h"
44 #include "src/utils/SkJSONWriter.h"
45 #include "src/utils/SkOSPath.h"
46 #include "src/utils/SkShaderUtils.h"
47 #include "tools/AutoreleasePool.h"
48 #include "tools/CrashHandler.h"
49 #include "tools/MSKPPlayer.h"
50 #include "tools/ProcStats.h"
51 #include "tools/Stats.h"
52 #include "tools/ToolUtils.h"
53 #include "tools/flags/CommonFlags.h"
54 #include "tools/flags/CommonFlagsConfig.h"
55 #include "tools/flags/CommonFlagsGanesh.h"
56 #include "tools/fonts/FontToolUtils.h"
57 #include "tools/ios_utils.h"
58 #include "tools/trace/EventTracingPriv.h"
59 #include "tools/trace/SkDebugfTracer.h"
60 
61 #if defined(SK_ENABLE_SVG)
62 #include "modules/skshaper/utils/FactoryHelpers.h"
63 #include "modules/svg/include/SkSVGDOM.h"
64 #include "modules/svg/include/SkSVGNode.h"
65 #endif
66 
67 #ifdef SK_ENABLE_ANDROID_UTILS
68 #include "bench/BitmapRegionDecoderBench.h"
69 #include "client_utils/android/BitmapRegionDecoder.h"
70 #endif
71 
72 #if defined(SK_GRAPHITE)
73 #include "include/gpu/graphite/Context.h"
74 #include "include/gpu/graphite/Recorder.h"
75 #include "include/gpu/graphite/Recording.h"
76 #include "include/gpu/graphite/Surface.h"
77 #include "tools/flags/CommonFlagsGraphite.h"
78 #include "tools/graphite/ContextFactory.h"
79 #include "tools/graphite/GraphiteTestContext.h"
80 #include "tools/graphite/GraphiteToolUtils.h"
81 #endif
82 
83 #include <cinttypes>
84 #include <memory>
85 #include <optional>
86 #include <stdlib.h>
87 #include <thread>
88 
89 extern bool gSkForceRasterPipelineBlitter;
90 extern bool gForceHighPrecisionRasterPipeline;
91 
92 #ifndef SK_BUILD_FOR_WIN
93 #include <unistd.h>
94 #endif
95 
96 #include "include/gpu/ganesh/GrDirectContext.h"
97 #include "include/gpu/ganesh/SkSurfaceGanesh.h"
98 #include "src/gpu/ganesh/GrCaps.h"
99 #include "src/gpu/ganesh/GrDirectContextPriv.h"
100 #include "src/gpu/ganesh/SkGr.h"
101 #include "tools/gpu/GrContextFactory.h"
102 
103 using namespace skia_private;
104 
105 using sk_gpu_test::ContextInfo;
106 using sk_gpu_test::GrContextFactory;
107 using sk_gpu_test::TestContext;
108 
109 GrContextOptions grContextOpts;
110 
111 #if defined(SK_GRAPHITE)
112 skiatest::graphite::TestOptions gTestOptions;
113 #endif
114 
115 static const int kAutoTuneLoops = 0;
116 
loops_help_txt()117 static SkString loops_help_txt() {
118     SkString help;
119     help.printf("Number of times to run each bench. Set this to %d to auto-"
120                 "tune for each bench. Timings are only reported when auto-tuning.",
121                 kAutoTuneLoops);
122     return help;
123 }
124 
to_string(int n)125 static SkString to_string(int n) {
126     SkString str;
127     str.appendS32(n);
128     return str;
129 }
130 
131 static DEFINE_int(loops, kAutoTuneLoops, loops_help_txt().c_str());
132 
133 static DEFINE_int(samples, 10, "Number of samples to measure for each bench.");
134 static DEFINE_int(ms, 0, "If >0, run each bench for this many ms instead of obeying --samples.");
135 static DEFINE_int(overheadLoops, 100000, "Loops to estimate timer overhead.");
136 static DEFINE_double(overheadGoal, 0.0001,
137               "Loop until timer overhead is at most this fraction of our measurments.");
138 static DEFINE_double(gpuMs, 5, "Target bench time in millseconds for GPU.");
139 static DEFINE_int(gpuFrameLag, 5,
140                     "If unknown, estimated maximum number of frames GPU allows to lag.");
141 
142 static DEFINE_string(outResultsFile, "", "If given, write results here as JSON.");
143 static DEFINE_int(maxCalibrationAttempts, 3,
144              "Try up to this many times to guess loops for a bench, or skip the bench.");
145 static DEFINE_int(maxLoops, 1000000, "Never run a bench more times than this.");
146 static DEFINE_string(clip, "0,0,1000,1000", "Clip for SKPs.");
147 static DEFINE_string(scales, "1.0", "Space-separated scales for SKPs.");
148 static DEFINE_string(zoom, "1.0,0",
149                      "Comma-separated zoomMax,zoomPeriodMs factors for a periodic SKP zoom "
150                      "function that ping-pongs between 1.0 and zoomMax.");
151 static DEFINE_bool(bbh, true, "Build a BBH for SKPs?");
152 static DEFINE_bool(loopSKP, true, "Loop SKPs like we do for micro benches?");
153 static DEFINE_int(flushEvery, 10, "Flush --outResultsFile every Nth run.");
154 static DEFINE_bool(gpuStats, false, "Print GPU stats after each gpu benchmark?");
155 static DEFINE_bool(gpuStatsDump, false, "Dump GPU stats after each benchmark to json");
156 static DEFINE_bool(dmsaaStatsDump, false, "Dump DMSAA stats after each benchmark to json");
157 static DEFINE_bool(keepAlive, false, "Print a message every so often so that we don't time out");
158 static DEFINE_bool(csv, false, "Print status in CSV format");
159 static DEFINE_string(sourceType, "",
160         "Apply usual --match rules to source type: bench, gm, skp, image, etc.");
161 static DEFINE_string(benchType,  "",
162         "Apply usual --match rules to bench type: micro, recording, "
163         "piping, playback, skcodec, etc.");
164 
165 static DEFINE_bool(forceRasterPipeline, false, "sets gSkForceRasterPipelineBlitter");
166 static DEFINE_bool(forceRasterPipelineHP, false, "sets gSkForceRasterPipelineBlitter and gForceHighPrecisionRasterPipeline");
167 
168 static DEFINE_bool2(pre_log, p, false,
169                     "Log before running each test. May be incomprehensible when threading");
170 
171 static DEFINE_bool(cpu, true, "Run CPU-bound work?");
172 static DEFINE_bool(gpu, true, "Run GPU-bound work?");
173 static DEFINE_bool(dryRun, false,
174                    "just print the tests that would be run, without actually running them.");
175 static DEFINE_string(images, "",
176                      "List of images and/or directories to decode. A directory with no images"
177                      " is treated as a fatal error.");
178 static DEFINE_bool(simpleCodec, false,
179                    "Runs of a subset of the codec tests, always N32, Premul or Opaque");
180 
181 static DEFINE_string2(match, m, nullptr,
182                "[~][^]substring[$] [...] of name to run.\n"
183                "Multiple matches may be separated by spaces.\n"
184                "~ causes a matching name to always be skipped\n"
185                "^ requires the start of the name to match\n"
186                "$ requires the end of the name to match\n"
187                "^ and $ requires an exact match\n"
188                "If a name does not match any list entry,\n"
189                "it is skipped unless some list entry starts with ~");
190 
191 static DEFINE_bool2(quiet, q, false, "if true, don't print status updates.");
192 static DEFINE_bool2(verbose, v, false, "enable verbose output from the test driver.");
193 
194 
195 static DEFINE_string(skps, "skps", "Directory to read skps from.");
196 static DEFINE_string(mskps, "mskps", "Directory to read mskps from.");
197 static DEFINE_string(svgs, "", "Directory to read SVGs from, or a single SVG file.");
198 static DEFINE_string(texttraces, "", "Directory to read TextBlobTrace files from.");
199 
200 static DEFINE_int_2(threads, j, -1,
201                "Run threadsafe tests on a threadpool with this many extra threads, "
202                "defaulting to one extra thread per core.");
203 
204 static DEFINE_string2(writePath, w, "", "If set, write bitmaps here as .pngs.");
205 
206 static DEFINE_string(key, "",
207                      "Space-separated key/value pairs to add to JSON identifying this builder.");
208 static DEFINE_string(properties, "",
209                      "Space-separated key/value pairs to add to JSON identifying this run.");
210 
211 static DEFINE_bool(purgeBetweenBenches, false,
212                    "Call SkGraphics::PurgeAllCaches() between each benchmark?");
213 
214 static DEFINE_bool(splitPerfettoTracesByBenchmark, true,
215                   "Create separate perfetto trace files for each benchmark?\n"
216                   "Will only take effect if perfetto tracing is enabled. See --trace.");
217 
218 static DEFINE_bool(runtimeCPUDetection, true, "Skip runtime CPU detection and optimization");
219 
now_ms()220 static double now_ms() { return SkTime::GetNSecs() * 1e-6; }
221 
humanize(double ms)222 static SkString humanize(double ms) {
223     if (FLAGS_verbose) return SkStringPrintf("%" PRIu64, (uint64_t)(ms*1e6));
224     return HumanizeMs(ms);
225 }
226 #define HUMANIZE(ms) humanize(ms).c_str()
227 
init(SkImageInfo info,Benchmark * bench)228 bool Target::init(SkImageInfo info, Benchmark* bench) {
229     if (Benchmark::Backend::kRaster == config.backend) {
230         this->surface = SkSurfaces::Raster(info);
231         if (!this->surface) {
232             return false;
233         }
234     }
235     return true;
236 }
capturePixels(SkBitmap * bmp)237 bool Target::capturePixels(SkBitmap* bmp) {
238     SkCanvas* canvas = this->getCanvas();
239     if (!canvas) {
240         return false;
241     }
242     bmp->allocPixels(canvas->imageInfo());
243     if (!canvas->readPixels(*bmp, 0, 0)) {
244         SkDebugf("Can't read canvas pixels.\n");
245         return false;
246     }
247     return true;
248 }
249 
250 struct GPUTarget : public Target {
GPUTargetGPUTarget251     explicit GPUTarget(const Config& c) : Target(c) {}
252     ContextInfo contextInfo;
253     std::unique_ptr<GrContextFactory> factory;
254 
~GPUTargetGPUTarget255     ~GPUTarget() override {
256         // For Vulkan we need to release all our refs to the GrContext before destroy the vulkan
257         // context which happens at the end of this destructor. Thus we need to release the surface
258         // here which holds a ref to the GrContext.
259         surface.reset();
260     }
261 
onSetupGPUTarget262     void onSetup() override {
263         this->contextInfo.testContext()->makeCurrent();
264     }
endTimingGPUTarget265     void endTiming() override {
266         if (this->contextInfo.testContext()) {
267             this->contextInfo.testContext()->flushAndWaitOnSync(contextInfo.directContext());
268         }
269     }
submitWorkAndSyncCPUGPUTarget270     void submitWorkAndSyncCPU() override {
271         if (this->contextInfo.testContext()) {
272             this->contextInfo.testContext()->flushAndSyncCpu(contextInfo.directContext());
273         }
274     }
275 
needsFrameTimingGPUTarget276     bool needsFrameTiming(int* maxFrameLag) const override {
277         if (!this->contextInfo.testContext()->getMaxGpuFrameLag(maxFrameLag)) {
278             // Frame lag is unknown.
279             *maxFrameLag = FLAGS_gpuFrameLag;
280         }
281         return true;
282     }
initGPUTarget283     bool init(SkImageInfo info, Benchmark* bench) override {
284         GrContextOptions options = grContextOpts;
285         bench->modifyGrContextOptions(&options);
286         this->factory = std::make_unique<GrContextFactory>(options);
287         SkSurfaceProps props(this->config.surfaceFlags, kRGB_H_SkPixelGeometry);
288         this->surface = SkSurfaces::RenderTarget(
289                 this->factory->get(this->config.ctxType, this->config.ctxOverrides),
290                 skgpu::Budgeted::kNo,
291                 info,
292                 this->config.samples,
293                 &props);
294         this->contextInfo =
295                 this->factory->getContextInfo(this->config.ctxType, this->config.ctxOverrides);
296         if (!this->surface) {
297             return false;
298         }
299         if (!this->contextInfo.testContext()->fenceSyncSupport()) {
300             SkDebugf("WARNING: GL context for config \"%s\" does not support fence sync. "
301                      "Timings might not be accurate.\n", this->config.name.c_str());
302         }
303         return true;
304     }
305 
dumpStatsGPUTarget306     void dumpStats() override {
307         auto context = this->contextInfo.directContext();
308 
309         context->priv().printCacheStats();
310         context->priv().printGpuStats();
311         context->priv().printContextStats();
312     }
313 };
314 
315 #if defined(SK_GRAPHITE)
316 struct GraphiteTarget : public Target {
GraphiteTargetGraphiteTarget317     explicit GraphiteTarget(const Config& c) : Target(c) {}
318     using TestContext = skiatest::graphite::GraphiteTestContext;
319     using ContextFactory = skiatest::graphite::ContextFactory;
320 
321     std::unique_ptr<ContextFactory> factory;
322 
323     TestContext* testContext;
324     skgpu::graphite::Context* context;
325     std::unique_ptr<skgpu::graphite::Recorder> recorder;
326 
~GraphiteTargetGraphiteTarget327     ~GraphiteTarget() override {
328         // For Vulkan we need to release all our refs before we destroy the vulkan context which
329         // happens at the end of this destructor. Thus we need to release the surface here which
330         // holds a ref to the Graphite device
331         surface.reset();
332     }
333 
endTimingGraphiteTarget334     void endTiming() override {
335         if (context && recorder) {
336             std::unique_ptr<skgpu::graphite::Recording> recording = this->recorder->snap();
337             if (recording) {
338                 this->testContext->submitRecordingAndWaitOnSync(this->context, recording.get());
339             }
340         }
341     }
submitWorkAndSyncCPUGraphiteTarget342     void submitWorkAndSyncCPU() override {
343         if (context && recorder) {
344             // TODO: have a way to sync work with out submitting a Recording which is currently
345             // required. Probably need to get to the point where the backend command buffers are
346             // stored on the Context and not Recordings before this is feasible.
347             std::unique_ptr<skgpu::graphite::Recording> recording = this->recorder->snap();
348             if (recording) {
349                 skgpu::graphite::InsertRecordingInfo info;
350                 info.fRecording = recording.get();
351                 this->context->insertRecording(info);
352             }
353             this->context->submit(skgpu::graphite::SyncToCpu::kYes);
354         }
355     }
356 
needsFrameTimingGraphiteTarget357     bool needsFrameTiming(int* maxFrameLag) const override {
358         SkAssertResult(this->testContext->getMaxGpuFrameLag(maxFrameLag));
359         return true;
360     }
initGraphiteTarget361     bool init(SkImageInfo info, Benchmark* bench) override {
362         skiatest::graphite::TestOptions testOptions = gTestOptions;
363         bench->modifyGraphiteContextOptions(&testOptions.fContextOptions);
364 
365         this->factory = std::make_unique<ContextFactory>(testOptions);
366 
367         skiatest::graphite::ContextInfo ctxInfo =
368                 this->factory->getContextInfo(this->config.ctxType);
369         if (!ctxInfo.fContext) {
370             return false;
371         }
372         this->testContext = ctxInfo.fTestContext;
373         this->context = ctxInfo.fContext;
374 
375         this->recorder = this->context->makeRecorder(ToolUtils::CreateTestingRecorderOptions());
376         if (!this->recorder) {
377             return false;
378         }
379 
380         this->surface = SkSurfaces::RenderTarget(this->recorder.get(), info);
381         if (!this->surface) {
382             return false;
383         }
384         // TODO: get fence stuff working
385 #if 0
386         if (!this->contextInfo.testContext()->fenceSyncSupport()) {
387             SkDebugf("WARNING: GL context for config \"%s\" does not support fence sync. "
388                      "Timings might not be accurate.\n", this->config.name.c_str());
389         }
390 #endif
391         return true;
392     }
393 
dumpStatsGraphiteTarget394     void dumpStats() override {
395     }
396 };
397 #endif // SK_GRAPHITE
398 
time(int loops,Benchmark * bench,Target * target)399 static double time(int loops, Benchmark* bench, Target* target) {
400     SkCanvas* canvas = target->getCanvas();
401     if (canvas) {
402         canvas->clear(SK_ColorWHITE);
403     }
404     bench->preDraw(canvas);
405     double start = now_ms();
406     canvas = target->beginTiming(canvas);
407 
408     bench->draw(loops, canvas);
409 
410     target->endTiming();
411     double elapsed = now_ms() - start;
412     bench->postDraw(canvas);
413     return elapsed;
414 }
415 
estimate_timer_overhead()416 static double estimate_timer_overhead() {
417     double overhead = 0;
418     for (int i = 0; i < FLAGS_overheadLoops; i++) {
419         double start = now_ms();
420         overhead += now_ms() - start;
421     }
422     return overhead / FLAGS_overheadLoops;
423 }
424 
detect_forever_loops(int loops)425 static int detect_forever_loops(int loops) {
426     // look for a magic run-forever value
427     if (loops < 0) {
428         loops = SK_MaxS32;
429     }
430     return loops;
431 }
432 
clamp_loops(int loops)433 static int clamp_loops(int loops) {
434     if (loops < 1) {
435         SkDebugf("ERROR: clamping loops from %d to 1. "
436                  "There's probably something wrong with the bench.\n", loops);
437         return 1;
438     }
439     if (loops > FLAGS_maxLoops) {
440         SkDebugf("WARNING: clamping loops from %d to FLAGS_maxLoops, %d.\n", loops, FLAGS_maxLoops);
441         return FLAGS_maxLoops;
442     }
443     return loops;
444 }
445 
write_canvas_png(Target * target,const SkString & filename)446 static bool write_canvas_png(Target* target, const SkString& filename) {
447 
448     if (filename.isEmpty()) {
449         return false;
450     }
451     if (target->getCanvas() &&
452         kUnknown_SkColorType == target->getCanvas()->imageInfo().colorType()) {
453         return false;
454     }
455 
456     SkBitmap bmp;
457 
458     if (!target->capturePixels(&bmp)) {
459         return false;
460     }
461 
462     SkString dir = SkOSPath::Dirname(filename.c_str());
463     if (!sk_mkdir(dir.c_str())) {
464         SkDebugf("Can't make dir %s.\n", dir.c_str());
465         return false;
466     }
467     SkFILEWStream stream(filename.c_str());
468     if (!stream.isValid()) {
469         SkDebugf("Can't write %s.\n", filename.c_str());
470         return false;
471     }
472     if (!SkPngEncoder::Encode(&stream, bmp.pixmap(), {})) {
473         SkDebugf("Can't encode a PNG.\n");
474         return false;
475     }
476     return true;
477 }
478 
479 static int kFailedLoops = -2;
setup_cpu_bench(const double overhead,Target * target,Benchmark * bench)480 static int setup_cpu_bench(const double overhead, Target* target, Benchmark* bench) {
481     // First figure out approximately how many loops of bench it takes to make overhead negligible.
482     double bench_plus_overhead = 0.0;
483     int round = 0;
484     int loops = bench->shouldLoop() ? FLAGS_loops : 1;
485     if (kAutoTuneLoops == loops) {
486         while (bench_plus_overhead < overhead) {
487             if (round++ == FLAGS_maxCalibrationAttempts) {
488                 SkDebugf("WARNING: Can't estimate loops for %s (%s vs. %s); skipping.\n",
489                          bench->getUniqueName(), HUMANIZE(bench_plus_overhead), HUMANIZE(overhead));
490                 return kFailedLoops;
491             }
492             bench_plus_overhead = time(1, bench, target);
493         }
494     }
495 
496     // Later we'll just start and stop the timer once but loop N times.
497     // We'll pick N to make timer overhead negligible:
498     //
499     //          overhead
500     //  -------------------------  < FLAGS_overheadGoal
501     //  overhead + N * Bench Time
502     //
503     // where bench_plus_overhead ~=~ overhead + Bench Time.
504     //
505     // Doing some math, we get:
506     //
507     //  (overhead / FLAGS_overheadGoal) - overhead
508     //  ------------------------------------------  < N
509     //       bench_plus_overhead - overhead)
510     //
511     // Luckily, this also works well in practice. :)
512     if (kAutoTuneLoops == loops) {
513         const double numer = overhead / FLAGS_overheadGoal - overhead;
514         const double denom = bench_plus_overhead - overhead;
515         loops = (int)ceil(numer / denom);
516         loops = clamp_loops(loops);
517     } else {
518         loops = detect_forever_loops(loops);
519     }
520 
521     return loops;
522 }
523 
setup_gpu_bench(Target * target,Benchmark * bench,int maxGpuFrameLag)524 static int setup_gpu_bench(Target* target, Benchmark* bench, int maxGpuFrameLag) {
525     // First, figure out how many loops it'll take to get a frame up to FLAGS_gpuMs.
526     int loops = bench->shouldLoop() ? FLAGS_loops : 1;
527     if (kAutoTuneLoops == loops) {
528         loops = 1;
529         double elapsed = 0;
530         do {
531             if (1<<30 == loops) {
532                 // We're about to wrap.  Something's wrong with the bench.
533                 loops = 0;
534                 break;
535             }
536             loops *= 2;
537             // If the GPU lets frames lag at all, we need to make sure we're timing
538             // _this_ round, not still timing last round.
539             for (int i = 0; i < maxGpuFrameLag; i++) {
540                 elapsed = time(loops, bench, target);
541             }
542         } while (elapsed < FLAGS_gpuMs);
543 
544         // We've overshot at least a little.  Scale back linearly.
545         loops = (int)ceil(loops * FLAGS_gpuMs / elapsed);
546         loops = clamp_loops(loops);
547 
548         // Make sure we're not still timing our calibration.
549         target->submitWorkAndSyncCPU();
550     } else {
551         loops = detect_forever_loops(loops);
552     }
553     // Pretty much the same deal as the calibration: do some warmup to make
554     // sure we're timing steady-state pipelined frames.
555     for (int i = 0; i < maxGpuFrameLag; i++) {
556         time(loops, bench, target);
557     }
558 
559     return loops;
560 }
561 
562 #define kBogusContextType skgpu::ContextType::kGL
563 #define kBogusContextOverrides GrContextFactory::ContextOverrides::kNone
564 
create_config(const SkCommandLineConfig * config)565 static std::optional<Config> create_config(const SkCommandLineConfig* config) {
566     if (const auto* gpuConfig = config->asConfigGpu()) {
567         if (!FLAGS_gpu) {
568             SkDebugf("Skipping config '%s' as requested.\n", config->getTag().c_str());
569             return std::nullopt;
570         }
571 
572         const auto ctxType = gpuConfig->getContextType();
573         const auto ctxOverrides = gpuConfig->getContextOverrides();
574         const auto sampleCount = gpuConfig->getSamples();
575         const auto colorType = gpuConfig->getColorType();
576         if (gpuConfig->getSurfType() != SkCommandLineConfigGpu::SurfType::kDefault) {
577             SkDebugf("This tool only supports the default surface type.");
578             return std::nullopt;
579         }
580 
581         GrContextFactory factory(grContextOpts);
582         if (const auto ctx = factory.get(ctxType, ctxOverrides)) {
583             GrBackendFormat format = ctx->defaultBackendFormat(colorType, GrRenderable::kYes);
584             int supportedSampleCount =
585                     ctx->priv().caps()->getRenderTargetSampleCount(sampleCount, format);
586             if (sampleCount != supportedSampleCount) {
587                 SkDebugf("Configuration '%s' sample count %d is not a supported sample count.\n",
588                          config->getTag().c_str(),
589                          sampleCount);
590                 return std::nullopt;
591             }
592         } else {
593             SkDebugf("No context was available matching config '%s'.\n", config->getTag().c_str());
594             return std::nullopt;
595         }
596 
597         return Config{gpuConfig->getTag(),
598                       Benchmark::Backend::kGanesh,
599                       colorType,
600                       kPremul_SkAlphaType,
601                       config->refColorSpace(),
602                       sampleCount,
603                       ctxType,
604                       ctxOverrides,
605                       gpuConfig->getSurfaceFlags()};
606     }
607 #if defined(SK_GRAPHITE)
608     if (const auto* gpuConfig = config->asConfigGraphite()) {
609         if (!FLAGS_gpu) {
610             SkDebugf("Skipping config '%s' as requested.\n", config->getTag().c_str());
611             return std::nullopt;
612         }
613 
614         const auto graphiteCtxType = gpuConfig->getContextType();
615         const auto sampleCount = 1; // TODO: gpuConfig->getSamples();
616         const auto colorType = gpuConfig->getColorType();
617 
618         using ContextFactory = skiatest::graphite::ContextFactory;
619 
620         ContextFactory factory(gTestOptions);
621         skiatest::graphite::ContextInfo ctxInfo = factory.getContextInfo(graphiteCtxType);
622         skgpu::graphite::Context* ctx = ctxInfo.fContext;
623         if (ctx) {
624             // TODO: Add graphite ctx queries for supported sample count by color type.
625 #if 0
626             GrBackendFormat format = ctx->defaultBackendFormat(colorType, GrRenderable::kYes);
627             int supportedSampleCount =
628                     ctx->priv().caps()->getRenderTargetSampleCount(sampleCount, format);
629             if (sampleCount != supportedSampleCount) {
630                 SkDebugf("Configuration '%s' sample count %d is not a supported sample count.\n",
631                          config->getTag().c_str(),
632                          sampleCount);
633                 return std::nullopt;
634             }
635 #else
636             if (sampleCount > 1) {
637                 SkDebugf("Configuration '%s' sample count %d is not a supported sample count.\n",
638                          config->getTag().c_str(),
639                          sampleCount);
640                 return std::nullopt;
641             }
642 #endif
643         } else {
644             SkDebugf("No context was available matching config '%s'.\n", config->getTag().c_str());
645             return std::nullopt;
646         }
647 
648         return Config{gpuConfig->getTag(),
649                       Benchmark::Backend::kGraphite,
650                       colorType,
651                       kPremul_SkAlphaType,
652                       config->refColorSpace(),
653                       sampleCount,
654                       graphiteCtxType,
655                       kBogusContextOverrides,
656                       0};
657     }
658 #endif
659 
660 #define CPU_CONFIG(name, backend, color, alpha)                                         \
661     if (config->getBackend().equals(name)) {                                            \
662         if (!FLAGS_cpu) {                                                               \
663             SkDebugf("Skipping config '%s' as requested.\n", config->getTag().c_str()); \
664             return std::nullopt;                                                      \
665         }                                                                               \
666         return Config{SkString(name),                                                   \
667                       Benchmark::backend,                                               \
668                       color,                                                            \
669                       alpha,                                                            \
670                       config->refColorSpace(),                                          \
671                       0,                                                                \
672                       kBogusContextType,                                                \
673                       kBogusContextOverrides,                                           \
674                       0};                                                               \
675     }
676 
677     CPU_CONFIG("nonrendering", Backend::kNonRendering, kUnknown_SkColorType, kUnpremul_SkAlphaType)
678 
679     CPU_CONFIG("a8",    Backend::kRaster,    kAlpha_8_SkColorType, kPremul_SkAlphaType)
680     CPU_CONFIG("565",   Backend::kRaster,    kRGB_565_SkColorType, kOpaque_SkAlphaType)
681     CPU_CONFIG("8888",  Backend::kRaster,        kN32_SkColorType, kPremul_SkAlphaType)
682     CPU_CONFIG("rgba",  Backend::kRaster,  kRGBA_8888_SkColorType, kPremul_SkAlphaType)
683     CPU_CONFIG("bgra",  Backend::kRaster,  kBGRA_8888_SkColorType, kPremul_SkAlphaType)
684     CPU_CONFIG("f16",   Backend::kRaster,   kRGBA_F16_SkColorType, kPremul_SkAlphaType)
685     CPU_CONFIG("srgba", Backend::kRaster, kSRGBA_8888_SkColorType, kPremul_SkAlphaType)
686 
687 #undef CPU_CONFIG
688 
689     SkDebugf("Unknown config '%s'.\n", config->getTag().c_str());
690     return std::nullopt;
691 }
692 
693 // Append all configs that are enabled and supported.
create_configs(TArray<Config> * configs)694 void create_configs(TArray<Config>* configs) {
695     SkCommandLineConfigArray array;
696     ParseConfigs(FLAGS_config, &array);
697     for (int i = 0; i < array.size(); ++i) {
698         if (std::optional<Config> config = create_config(array[i].get())) {
699             configs->push_back(*config);
700         }
701     }
702 
703     // If no just default configs were requested, then we're okay.
704     if (array.size() == 0 || FLAGS_config.size() == 0 ||
705         // Otherwise, make sure that all specified configs have been created.
706         array.size() == configs->size()) {
707         return;
708     }
709     exit(1);
710 }
711 
712 // disable warning : switch statement contains default but no 'case' labels
713 #if defined _WIN32
714 #pragma warning ( push )
715 #pragma warning ( disable : 4065 )
716 #endif
717 
718 // If bench is enabled for config, returns a Target* for it, otherwise nullptr.
is_enabled(Benchmark * bench,const Config & config)719 static Target* is_enabled(Benchmark* bench, const Config& config) {
720     if (!bench->isSuitableFor(config.backend)) {
721         return nullptr;
722     }
723 
724     SkImageInfo info =
725             SkImageInfo::Make(bench->getSize(), config.color, config.alpha, config.colorSpace);
726 
727     Target* target = nullptr;
728 
729     switch (config.backend) {
730     case Benchmark::Backend::kGanesh:
731         target = new GPUTarget(config);
732         break;
733 #if defined(SK_GRAPHITE)
734     case Benchmark::Backend::kGraphite:
735         target = new GraphiteTarget(config);
736         break;
737 #endif
738     default:
739         target = new Target(config);
740         break;
741     }
742 
743     if (!target->init(info, bench)) {
744         delete target;
745         return nullptr;
746     }
747     return target;
748 }
749 
750 #if defined _WIN32
751 #pragma warning ( pop )
752 #endif
753 
754 #ifdef SK_ENABLE_ANDROID_UTILS
valid_brd_bench(sk_sp<SkData> encoded,SkColorType colorType,uint32_t sampleSize,uint32_t minOutputSize,int * width,int * height)755 static bool valid_brd_bench(sk_sp<SkData> encoded, SkColorType colorType, uint32_t sampleSize,
756         uint32_t minOutputSize, int* width, int* height) {
757     auto brd = android::skia::BitmapRegionDecoder::Make(encoded);
758     if (nullptr == brd) {
759         // This is indicates that subset decoding is not supported for a particular image format.
760         return false;
761     }
762 
763     if (sampleSize * minOutputSize > (uint32_t) brd->width() || sampleSize * minOutputSize >
764             (uint32_t) brd->height()) {
765         // This indicates that the image is not large enough to decode a
766         // minOutputSize x minOutputSize subset at the given sampleSize.
767         return false;
768     }
769 
770     // Set the image width and height.  The calling code will use this to choose subsets to decode.
771     *width = brd->width();
772     *height = brd->height();
773     return true;
774 }
775 #endif
776 
cleanup_run(Target * target)777 static void cleanup_run(Target* target) {
778     delete target;
779 }
780 
collect_files(const CommandLineFlags::StringArray & paths,const char * ext,TArray<SkString> * list)781 static void collect_files(const CommandLineFlags::StringArray& paths,
782                           const char*                          ext,
783                           TArray<SkString>*                  list) {
784     for (int i = 0; i < paths.size(); ++i) {
785         if (SkStrEndsWith(paths[i], ext)) {
786             list->push_back(SkString(paths[i]));
787         } else {
788             SkOSFile::Iter it(paths[i], ext);
789             SkString path;
790             while (it.next(&path)) {
791                 list->push_back(SkOSPath::Join(paths[i], path.c_str()));
792             }
793         }
794     }
795 }
796 
797 class BenchmarkStream {
798 public:
BenchmarkStream()799     BenchmarkStream() : fBenches(BenchRegistry::Head())
800                       , fGMs(skiagm::GMRegistry::Head()) {
801         collect_files(FLAGS_skps, ".skp", &fSKPs);
802         collect_files(FLAGS_mskps, ".mskp", &fMSKPs);
803         collect_files(FLAGS_svgs, ".svg", &fSVGs);
804         collect_files(FLAGS_texttraces, ".trace", &fTextBlobTraces);
805 
806         if (4 != sscanf(FLAGS_clip[0], "%d,%d,%d,%d",
807                         &fClip.fLeft, &fClip.fTop, &fClip.fRight, &fClip.fBottom)) {
808             SkDebugf("Can't parse %s from --clip as an SkIRect.\n", FLAGS_clip[0]);
809             exit(1);
810         }
811 
812         for (int i = 0; i < FLAGS_scales.size(); i++) {
813             if (1 != sscanf(FLAGS_scales[i], "%f", &fScales.push_back())) {
814                 SkDebugf("Can't parse %s from --scales as an SkScalar.\n", FLAGS_scales[i]);
815                 exit(1);
816             }
817         }
818 
819         if (2 != sscanf(FLAGS_zoom[0], "%f,%lf", &fZoomMax, &fZoomPeriodMs)) {
820             SkDebugf("Can't parse %s from --zoom as a zoomMax,zoomPeriodMs.\n", FLAGS_zoom[0]);
821             exit(1);
822         }
823 
824         // Prepare the images for decoding
825         if (!CommonFlags::CollectImages(FLAGS_images, &fImages)) {
826             exit(1);
827         }
828 
829         // Choose the candidate color types for image decoding
830         fColorTypes.push_back(kN32_SkColorType);
831         if (!FLAGS_simpleCodec) {
832             fColorTypes.push_back(kRGB_565_SkColorType);
833             fColorTypes.push_back(kAlpha_8_SkColorType);
834             fColorTypes.push_back(kGray_8_SkColorType);
835         }
836     }
837 
ReadPicture(const char * path)838     static sk_sp<SkPicture> ReadPicture(const char* path) {
839         // Not strictly necessary, as it will be checked again later,
840         // but helps to avoid a lot of pointless work if we're going to skip it.
841         if (CommandLineFlags::ShouldSkip(FLAGS_match, SkOSPath::Basename(path).c_str())) {
842             return nullptr;
843         }
844 
845         std::unique_ptr<SkStream> stream = SkStream::MakeFromFile(path);
846         if (!stream) {
847             SkDebugf("Could not read %s.\n", path);
848             return nullptr;
849         }
850 
851         return SkPicture::MakeFromStream(stream.get());
852     }
853 
ReadMSKP(const char * path)854     static std::unique_ptr<MSKPPlayer> ReadMSKP(const char* path) {
855         // Not strictly necessary, as it will be checked again later,
856         // but helps to avoid a lot of pointless work if we're going to skip it.
857         if (CommandLineFlags::ShouldSkip(FLAGS_match, SkOSPath::Basename(path).c_str())) {
858             return nullptr;
859         }
860 
861         std::unique_ptr<SkStreamSeekable> stream = SkStream::MakeFromFile(path);
862         if (!stream) {
863             SkDebugf("Could not read %s.\n", path);
864             return nullptr;
865         }
866 
867         return MSKPPlayer::Make(stream.get());
868     }
869 
ReadSVGPicture(const char * path)870     static sk_sp<SkPicture> ReadSVGPicture(const char* path) {
871         if (CommandLineFlags::ShouldSkip(FLAGS_match, SkOSPath::Basename(path).c_str())) {
872             return nullptr;
873         }
874         sk_sp<SkData> data(SkData::MakeFromFileName(path));
875         if (!data) {
876             SkDebugf("Could not read %s.\n", path);
877             return nullptr;
878         }
879 
880 #if defined(SK_ENABLE_SVG)
881         SkMemoryStream stream(std::move(data));
882         sk_sp<SkSVGDOM> svgDom = SkSVGDOM::Builder()
883                                          .setFontManager(ToolUtils::TestFontMgr())
884                                          .setTextShapingFactory(SkShapers::BestAvailable())
885                                          .make(stream);
886         if (!svgDom) {
887             SkDebugf("Could not parse %s.\n", path);
888             return nullptr;
889         }
890 
891         // Use the intrinsic SVG size if available, otherwise fall back to a default value.
892         static const SkSize kDefaultContainerSize = SkSize::Make(128, 128);
893         if (svgDom->containerSize().isEmpty()) {
894             svgDom->setContainerSize(kDefaultContainerSize);
895         }
896 
897         SkPictureRecorder recorder;
898         svgDom->render(recorder.beginRecording(svgDom->containerSize().width(),
899                                                svgDom->containerSize().height()));
900         return recorder.finishRecordingAsPicture();
901 #else
902         return nullptr;
903 #endif  // defined(SK_ENABLE_SVG)
904     }
905 
next()906     Benchmark* next() {
907         std::unique_ptr<Benchmark> bench;
908         do {
909             bench.reset(this->rawNext());
910             if (!bench) {
911                 return nullptr;
912             }
913         } while (CommandLineFlags::ShouldSkip(FLAGS_sourceType, fSourceType) ||
914                  CommandLineFlags::ShouldSkip(FLAGS_benchType, fBenchType));
915         return bench.release();
916     }
917 
rawNext()918     Benchmark* rawNext() {
919         if (fBenches) {
920             Benchmark* bench = fBenches->get()(nullptr);
921             fBenches = fBenches->next();
922             fSourceType = "bench";
923             fBenchType  = "micro";
924             return bench;
925         }
926 
927         while (fGMs) {
928             std::unique_ptr<skiagm::GM> gm = fGMs->get()();
929             if (gm->isBazelOnly()) {
930                 // We skip Bazel-only GMs because they might not be regular GMs. The Bazel build
931                 // reuses the notion of GMs to replace the notion of DM sources of various kinds,
932                 // such as codec sources and image generation sources. See comments in the
933                 // skiagm::GM::isBazelOnly function declaration for context.
934                 continue;
935             }
936             fGMs = fGMs->next();
937             if (gm->runAsBench()) {
938                 fSourceType = "gm";
939                 fBenchType  = "micro";
940                 return new GMBench(std::move(gm));
941             }
942         }
943 
944         while (fCurrentTextBlobTrace < fTextBlobTraces.size()) {
945             SkString path = fTextBlobTraces[fCurrentTextBlobTrace++];
946             SkString basename = SkOSPath::Basename(path.c_str());
947             static constexpr char kEnding[] = ".trace";
948             if (basename.endsWith(kEnding)) {
949                 basename.remove(basename.size() - strlen(kEnding), strlen(kEnding));
950             }
951             fSourceType = "texttrace";
952             fBenchType  = "micro";
953             return CreateDiffCanvasBench(
954                     SkStringPrintf("SkDiffBench-%s", basename.c_str()),
955                     [path](){ return SkStream::MakeFromFile(path.c_str()); });
956         }
957 
958         // First add all .skps as RecordingBenches.
959         while (fCurrentRecording < fSKPs.size()) {
960             const SkString& path = fSKPs[fCurrentRecording++];
961             sk_sp<SkPicture> pic = ReadPicture(path.c_str());
962             if (!pic) {
963                 continue;
964             }
965             SkString name = SkOSPath::Basename(path.c_str());
966             fSourceType = "skp";
967             fBenchType  = "recording";
968             fSKPBytes = static_cast<double>(pic->approximateBytesUsed());
969             fSKPOps   = pic->approximateOpCount();
970             return new RecordingBench(name.c_str(), pic.get(), FLAGS_bbh);
971         }
972 
973         // Add all .skps as DeserializePictureBenchs.
974         while (fCurrentDeserialPicture < fSKPs.size()) {
975             const SkString& path = fSKPs[fCurrentDeserialPicture++];
976             sk_sp<SkData> data = SkData::MakeFromFileName(path.c_str());
977             if (!data) {
978                 continue;
979             }
980             SkString name = SkOSPath::Basename(path.c_str());
981             fSourceType = "skp";
982             fBenchType  = "deserial";
983             fSKPBytes = static_cast<double>(data->size());
984             fSKPOps   = 0;
985             return new DeserializePictureBench(name.c_str(), std::move(data));
986         }
987 
988         // Then once each for each scale as SKPBenches (playback).
989         while (fCurrentScale < fScales.size()) {
990             while (fCurrentSKP < fSKPs.size()) {
991                 const SkString& path = fSKPs[fCurrentSKP++];
992                 sk_sp<SkPicture> pic = ReadPicture(path.c_str());
993                 if (!pic) {
994                     continue;
995                 }
996 
997                 if (FLAGS_bbh) {
998                     // The SKP we read off disk doesn't have a BBH.  Re-record so it grows one.
999                     SkRTreeFactory factory;
1000                     SkPictureRecorder recorder;
1001                     pic->playback(recorder.beginRecording(pic->cullRect().width(),
1002                                                           pic->cullRect().height(),
1003                                                           &factory));
1004                     pic = recorder.finishRecordingAsPicture();
1005                 }
1006                 SkString name = SkOSPath::Basename(path.c_str());
1007                 fSourceType = "skp";
1008                 fBenchType = "playback";
1009                 return new SKPBench(name.c_str(), pic.get(), fClip, fScales[fCurrentScale],
1010                                     FLAGS_loopSKP);
1011             }
1012 
1013             while (fCurrentSVG < fSVGs.size()) {
1014                 const char* path = fSVGs[fCurrentSVG++].c_str();
1015                 if (sk_sp<SkPicture> pic = ReadSVGPicture(path)) {
1016                     fSourceType = "svg";
1017                     fBenchType = "playback";
1018                     return new SKPBench(SkOSPath::Basename(path).c_str(), pic.get(), fClip,
1019                                         fScales[fCurrentScale], FLAGS_loopSKP);
1020                 }
1021             }
1022 
1023             fCurrentSKP = 0;
1024             fCurrentSVG = 0;
1025             fCurrentScale++;
1026         }
1027 
1028         // Now loop over each skp again if we have an animation
1029         if (fZoomMax != 1.0f && fZoomPeriodMs > 0) {
1030             while (fCurrentAnimSKP < fSKPs.size()) {
1031                 const SkString& path = fSKPs[fCurrentAnimSKP];
1032                 sk_sp<SkPicture> pic = ReadPicture(path.c_str());
1033                 if (!pic) {
1034                     fCurrentAnimSKP++;
1035                     continue;
1036                 }
1037 
1038                 fCurrentAnimSKP++;
1039                 SkString name = SkOSPath::Basename(path.c_str());
1040                 sk_sp<SKPAnimationBench::Animation> animation =
1041                     SKPAnimationBench::MakeZoomAnimation(fZoomMax, fZoomPeriodMs);
1042                 return new SKPAnimationBench(name.c_str(), pic.get(), fClip, std::move(animation),
1043                                              FLAGS_loopSKP);
1044             }
1045         }
1046 
1047         // Read all MSKPs as benches
1048         while (fCurrentMSKP < fMSKPs.size()) {
1049             const SkString& path = fMSKPs[fCurrentMSKP++];
1050             std::unique_ptr<MSKPPlayer> player = ReadMSKP(path.c_str());
1051             if (!player) {
1052                 continue;
1053             }
1054             SkString name = SkOSPath::Basename(path.c_str());
1055             fSourceType = "mskp";
1056             fBenchType = "mskp";
1057             return new MSKPBench(std::move(name), std::move(player));
1058         }
1059 
1060         for (; fCurrentCodec < fImages.size(); fCurrentCodec++) {
1061             fSourceType = "image";
1062             fBenchType = "skcodec";
1063             const SkString& path = fImages[fCurrentCodec];
1064             if (CommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
1065                 continue;
1066             }
1067             sk_sp<SkData> encoded(SkData::MakeFromFileName(path.c_str()));
1068             std::unique_ptr<SkCodec> codec(SkCodec::MakeFromData(encoded));
1069             if (!codec) {
1070                 // Nothing to time.
1071                 SkDebugf("Cannot find codec for %s\n", path.c_str());
1072                 continue;
1073             }
1074 
1075             while (fCurrentColorType < fColorTypes.size()) {
1076                 const SkColorType colorType = fColorTypes[fCurrentColorType];
1077 
1078                 SkAlphaType alphaType = codec->getInfo().alphaType();
1079                 if (FLAGS_simpleCodec) {
1080                     if (kUnpremul_SkAlphaType == alphaType) {
1081                         alphaType = kPremul_SkAlphaType;
1082                     }
1083 
1084                     fCurrentColorType++;
1085                 } else {
1086                     switch (alphaType) {
1087                         case kOpaque_SkAlphaType:
1088                             // We only need to test one alpha type (opaque).
1089                             fCurrentColorType++;
1090                             break;
1091                         case kUnpremul_SkAlphaType:
1092                         case kPremul_SkAlphaType:
1093                             if (0 == fCurrentAlphaType) {
1094                                 // Test unpremul first.
1095                                 alphaType = kUnpremul_SkAlphaType;
1096                                 fCurrentAlphaType++;
1097                             } else {
1098                                 // Test premul.
1099                                 alphaType = kPremul_SkAlphaType;
1100                                 fCurrentAlphaType = 0;
1101                                 fCurrentColorType++;
1102                             }
1103                             break;
1104                         default:
1105                             SkASSERT(false);
1106                             fCurrentColorType++;
1107                             break;
1108                     }
1109                 }
1110 
1111                 // Make sure we can decode to this color type and alpha type.
1112                 SkImageInfo info =
1113                         codec->getInfo().makeColorType(colorType).makeAlphaType(alphaType);
1114                 const size_t rowBytes = info.minRowBytes();
1115                 SkAutoMalloc storage(info.computeByteSize(rowBytes));
1116 
1117                 const SkCodec::Result result = codec->getPixels(
1118                         info, storage.get(), rowBytes);
1119                 switch (result) {
1120                     case SkCodec::kSuccess:
1121                     case SkCodec::kIncompleteInput:
1122                         return new CodecBench(SkOSPath::Basename(path.c_str()),
1123                                               encoded.get(), colorType, alphaType);
1124                     case SkCodec::kInvalidConversion:
1125                         // This is okay. Not all conversions are valid.
1126                         break;
1127                     default:
1128                         // This represents some sort of failure.
1129                         SkASSERT(false);
1130                         break;
1131                 }
1132             }
1133             fCurrentColorType = 0;
1134         }
1135 
1136         // Run AndroidCodecBenches
1137         const int sampleSizes[] = { 2, 4, 8 };
1138         for (; fCurrentAndroidCodec < fImages.size(); fCurrentAndroidCodec++) {
1139             fSourceType = "image";
1140             fBenchType = "skandroidcodec";
1141 
1142             const SkString& path = fImages[fCurrentAndroidCodec];
1143             if (CommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
1144                 continue;
1145             }
1146             sk_sp<SkData> encoded(SkData::MakeFromFileName(path.c_str()));
1147             std::unique_ptr<SkAndroidCodec> codec(SkAndroidCodec::MakeFromData(encoded));
1148             if (!codec) {
1149                 // Nothing to time.
1150                 SkDebugf("Cannot find codec for %s\n", path.c_str());
1151                 continue;
1152             }
1153 
1154             while (fCurrentSampleSize < (int) std::size(sampleSizes)) {
1155                 int sampleSize = sampleSizes[fCurrentSampleSize];
1156                 fCurrentSampleSize++;
1157                 if (10 * sampleSize > std::min(codec->getInfo().width(), codec->getInfo().height())) {
1158                     // Avoid benchmarking scaled decodes of already small images.
1159                     break;
1160                 }
1161 
1162                 return new AndroidCodecBench(SkOSPath::Basename(path.c_str()),
1163                                              encoded.get(), sampleSize);
1164             }
1165             fCurrentSampleSize = 0;
1166         }
1167 
1168 #ifdef SK_ENABLE_ANDROID_UTILS
1169         // Run the BRDBenches
1170         // We intend to create benchmarks that model the use cases in
1171         // android/libraries/social/tiledimage.  In this library, an image is decoded in 512x512
1172         // tiles.  The image can be translated freely, so the location of a tile may be anywhere in
1173         // the image.  For that reason, we will benchmark decodes in five representative locations
1174         // in the image.  Additionally, this use case utilizes power of two scaling, so we will
1175         // test on power of two sample sizes.  The output tile is always 512x512, so, when a
1176         // sampleSize is used, the size of the subset that is decoded is always
1177         // (sampleSize*512)x(sampleSize*512).
1178         // There are a few good reasons to only test on power of two sample sizes at this time:
1179         //     All use cases we are aware of only scale by powers of two.
1180         //     PNG decodes use the indicated sampling strategy regardless of the sample size, so
1181         //         these tests are sufficient to provide good coverage of our scaling options.
1182         const uint32_t brdSampleSizes[] = { 1, 2, 4, 8, 16 };
1183         const uint32_t minOutputSize = 512;
1184         for (; fCurrentBRDImage < fImages.size(); fCurrentBRDImage++) {
1185             fSourceType = "image";
1186             fBenchType = "BRD";
1187 
1188             const SkString& path = fImages[fCurrentBRDImage];
1189             if (CommandLineFlags::ShouldSkip(FLAGS_match, path.c_str())) {
1190                 continue;
1191             }
1192 
1193             while (fCurrentColorType < fColorTypes.size()) {
1194                 while (fCurrentSampleSize < (int) std::size(brdSampleSizes)) {
1195                     while (fCurrentSubsetType <= kLastSingle_SubsetType) {
1196 
1197                         sk_sp<SkData> encoded(SkData::MakeFromFileName(path.c_str()));
1198                         const SkColorType colorType = fColorTypes[fCurrentColorType];
1199                         uint32_t sampleSize = brdSampleSizes[fCurrentSampleSize];
1200                         int currentSubsetType = fCurrentSubsetType++;
1201 
1202                         int width = 0;
1203                         int height = 0;
1204                         if (!valid_brd_bench(encoded, colorType, sampleSize, minOutputSize,
1205                                 &width, &height)) {
1206                             break;
1207                         }
1208 
1209                         SkString basename = SkOSPath::Basename(path.c_str());
1210                         SkIRect subset;
1211                         const uint32_t subsetSize = sampleSize * minOutputSize;
1212                         switch (currentSubsetType) {
1213                             case kTopLeft_SubsetType:
1214                                 basename.append("_TopLeft");
1215                                 subset = SkIRect::MakeXYWH(0, 0, subsetSize, subsetSize);
1216                                 break;
1217                             case kTopRight_SubsetType:
1218                                 basename.append("_TopRight");
1219                                 subset = SkIRect::MakeXYWH(width - subsetSize, 0, subsetSize,
1220                                         subsetSize);
1221                                 break;
1222                             case kMiddle_SubsetType:
1223                                 basename.append("_Middle");
1224                                 subset = SkIRect::MakeXYWH((width - subsetSize) / 2,
1225                                         (height - subsetSize) / 2, subsetSize, subsetSize);
1226                                 break;
1227                             case kBottomLeft_SubsetType:
1228                                 basename.append("_BottomLeft");
1229                                 subset = SkIRect::MakeXYWH(0, height - subsetSize, subsetSize,
1230                                         subsetSize);
1231                                 break;
1232                             case kBottomRight_SubsetType:
1233                                 basename.append("_BottomRight");
1234                                 subset = SkIRect::MakeXYWH(width - subsetSize,
1235                                         height - subsetSize, subsetSize, subsetSize);
1236                                 break;
1237                             default:
1238                                 SkASSERT(false);
1239                         }
1240 
1241                         return new BitmapRegionDecoderBench(basename.c_str(), encoded.get(),
1242                                 colorType, sampleSize, subset);
1243                     }
1244                     fCurrentSubsetType = 0;
1245                     fCurrentSampleSize++;
1246                 }
1247                 fCurrentSampleSize = 0;
1248                 fCurrentColorType++;
1249             }
1250             fCurrentColorType = 0;
1251         }
1252 #endif // SK_ENABLE_ANDROID_UTILS
1253 
1254         return nullptr;
1255     }
1256 
fillCurrentOptions(NanoJSONResultsWriter & log) const1257     void fillCurrentOptions(NanoJSONResultsWriter& log) const {
1258         log.appendCString("source_type", fSourceType);
1259         log.appendCString("bench_type",  fBenchType);
1260         if (0 == strcmp(fSourceType, "skp")) {
1261             log.appendString("clip",
1262                     SkStringPrintf("%d %d %d %d", fClip.fLeft, fClip.fTop,
1263                                                   fClip.fRight, fClip.fBottom));
1264             SkASSERT_RELEASE(fCurrentScale < fScales.size());  // debugging paranoia
1265             log.appendString("scale", SkStringPrintf("%.2g", fScales[fCurrentScale]));
1266         }
1267     }
1268 
fillCurrentMetrics(NanoJSONResultsWriter & log) const1269     void fillCurrentMetrics(NanoJSONResultsWriter& log) const {
1270         if (0 == strcmp(fBenchType, "recording")) {
1271             log.appendMetric("bytes", fSKPBytes);
1272             log.appendMetric("ops", fSKPOps);
1273         }
1274     }
1275 
1276 private:
1277 #ifdef SK_ENABLE_ANDROID_UTILS
1278     enum SubsetType {
1279         kTopLeft_SubsetType     = 0,
1280         kTopRight_SubsetType    = 1,
1281         kMiddle_SubsetType      = 2,
1282         kBottomLeft_SubsetType  = 3,
1283         kBottomRight_SubsetType = 4,
1284         kTranslate_SubsetType   = 5,
1285         kZoom_SubsetType        = 6,
1286         kLast_SubsetType        = kZoom_SubsetType,
1287         kLastSingle_SubsetType  = kBottomRight_SubsetType,
1288     };
1289 #endif
1290 
1291     const BenchRegistry* fBenches;
1292     const skiagm::GMRegistry* fGMs;
1293     SkIRect            fClip;
1294     TArray<SkScalar> fScales;
1295     TArray<SkString> fSKPs;
1296     TArray<SkString> fMSKPs;
1297     TArray<SkString> fSVGs;
1298     TArray<SkString> fTextBlobTraces;
1299     TArray<SkString> fImages;
1300     TArray<SkColorType, true> fColorTypes;
1301     SkScalar           fZoomMax;
1302     double             fZoomPeriodMs;
1303 
1304     double fSKPBytes, fSKPOps;
1305 
1306     const char* fSourceType;  // What we're benching: bench, GM, SKP, ...
1307     const char* fBenchType;   // How we bench it: micro, recording, playback, ...
1308     int fCurrentRecording = 0;
1309     int fCurrentDeserialPicture = 0;
1310     int fCurrentMSKP = 0;
1311     int fCurrentScale = 0;
1312     int fCurrentSKP = 0;
1313     int fCurrentSVG = 0;
1314     int fCurrentTextBlobTrace = 0;
1315     int fCurrentCodec = 0;
1316     int fCurrentAndroidCodec = 0;
1317 #ifdef SK_ENABLE_ANDROID_UTILS
1318     int fCurrentBRDImage = 0;
1319     int fCurrentSubsetType = 0;
1320 #endif
1321     int fCurrentColorType = 0;
1322     int fCurrentAlphaType = 0;
1323     int fCurrentSampleSize = 0;
1324     int fCurrentAnimSKP = 0;
1325 };
1326 
1327 // Some runs (mostly, Valgrind) are so slow that the bot framework thinks we've hung.
1328 // This prints something every once in a while so that it knows we're still working.
start_keepalive()1329 static void start_keepalive() {
1330     static std::thread* intentionallyLeaked = new std::thread([]{
1331         for (;;) {
1332             static const int kSec = 1200;
1333         #if defined(SK_BUILD_FOR_WIN)
1334             Sleep(kSec * 1000);
1335         #else
1336             sleep(kSec);
1337         #endif
1338             SkDebugf("\nBenchmarks still running...\n");
1339         }
1340     });
1341     (void)intentionallyLeaked;
1342     SK_INTENTIONALLY_LEAKED(intentionallyLeaked);
1343 }
1344 
1345 class NanobenchShaderErrorHandler : public GrContextOptions::ShaderErrorHandler {
compileError(const char * shader,const char * errors)1346     void compileError(const char* shader, const char* errors) override {
1347         // Nanobench should abort if any shader can't compile. Failure is much better than
1348         // reporting meaningless performance metrics.
1349         std::string message = SkShaderUtils::BuildShaderErrorMessage(shader, errors);
1350         SK_ABORT("\n%s", message.c_str());
1351     }
1352 };
1353 
main(int argc,char ** argv)1354 int main(int argc, char** argv) {
1355     CommandLineFlags::Parse(argc, argv);
1356 
1357     initializeEventTracingForTools();
1358 
1359 #if defined(SK_BUILD_FOR_IOS)
1360     cd_Documents();
1361 #endif
1362     SetupCrashHandler();
1363     if (FLAGS_runtimeCPUDetection) {
1364         SkGraphics::Init();
1365     }
1366 
1367     // Our benchmarks only currently decode .png or .jpg files
1368     SkCodecs::Register(SkPngDecoder::Decoder());
1369     SkCodecs::Register(SkJpegDecoder::Decoder());
1370 
1371     SkTaskGroup::Enabler enabled(FLAGS_threads);
1372 
1373     CommonFlags::SetCtxOptions(&grContextOpts);
1374 
1375 #if defined(SK_GRAPHITE)
1376     CommonFlags::SetTestOptions(&gTestOptions);
1377 #endif
1378 
1379     NanobenchShaderErrorHandler errorHandler;
1380     grContextOpts.fShaderErrorHandler = &errorHandler;
1381 
1382     if (kAutoTuneLoops != FLAGS_loops) {
1383         FLAGS_samples     = 1;
1384         FLAGS_gpuFrameLag = 0;
1385     }
1386 
1387     if (!FLAGS_writePath.isEmpty()) {
1388         SkDebugf("Writing files to %s.\n", FLAGS_writePath[0]);
1389         if (!sk_mkdir(FLAGS_writePath[0])) {
1390             SkDebugf("Could not create %s. Files won't be written.\n", FLAGS_writePath[0]);
1391             FLAGS_writePath.set(0, nullptr);
1392         }
1393     }
1394 
1395     std::unique_ptr<SkWStream> logStream(new SkNullWStream);
1396     if (!FLAGS_outResultsFile.isEmpty()) {
1397 #if defined(SK_RELEASE)
1398         logStream.reset(new SkFILEWStream(FLAGS_outResultsFile[0]));
1399 #else
1400         SkDebugf("I'm ignoring --outResultsFile because this is a Debug build.");
1401         return 1;
1402 #endif
1403     }
1404     NanoJSONResultsWriter log(logStream.get(), SkJSONWriter::Mode::kPretty);
1405     log.beginObject(); // root
1406 
1407     if (1 == FLAGS_properties.size() % 2) {
1408         SkDebugf("ERROR: --properties must be passed with an even number of arguments.\n");
1409         return 1;
1410     }
1411     for (int i = 1; i < FLAGS_properties.size(); i += 2) {
1412         log.appendCString(FLAGS_properties[i-1], FLAGS_properties[i]);
1413     }
1414 
1415     if (1 == FLAGS_key.size() % 2) {
1416         SkDebugf("ERROR: --key must be passed with an even number of arguments.\n");
1417         return 1;
1418     }
1419     if (FLAGS_key.size()) {
1420         log.beginObject("key");
1421         for (int i = 1; i < FLAGS_key.size(); i += 2) {
1422             log.appendCString(FLAGS_key[i - 1], FLAGS_key[i]);
1423         }
1424         log.endObject(); // key
1425     }
1426 
1427     const double overhead = estimate_timer_overhead();
1428     if (!FLAGS_quiet && !FLAGS_csv) {
1429         SkDebugf("Timer overhead: %s\n", HUMANIZE(overhead));
1430     }
1431 
1432     TArray<double> samples;
1433 
1434     if (kAutoTuneLoops != FLAGS_loops) {
1435         SkDebugf("Fixed number of loops; times would only be misleading so we won't print them.\n");
1436     } else if (FLAGS_quiet) {
1437         SkDebugf("! -> high variance, ? -> moderate variance\n");
1438         SkDebugf("    micros   \tbench\n");
1439     } else if (FLAGS_csv) {
1440         SkDebugf("min,median,mean,max,stddev,config,bench\n");
1441     } else if (FLAGS_ms) {
1442         SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\tsamples\tconfig\tbench\n");
1443     } else {
1444         SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\t%-*s\tconfig\tbench\n",
1445                  FLAGS_samples, "samples");
1446     }
1447 
1448     GrRecordingContextPriv::DMSAAStats combinedDMSAAStats;
1449 
1450     TArray<Config> configs;
1451     create_configs(&configs);
1452 
1453     if (FLAGS_keepAlive) {
1454         start_keepalive();
1455     }
1456 
1457     gSkForceRasterPipelineBlitter     = FLAGS_forceRasterPipelineHP || FLAGS_forceRasterPipeline;
1458     gForceHighPrecisionRasterPipeline = FLAGS_forceRasterPipelineHP;
1459 
1460     // The SkSL memory benchmark must run before any GPU painting occurs. SkSL allocates memory for
1461     // its modules the first time they are accessed, and this test is trying to measure the size of
1462     // those allocations. If a paint has already occurred, some modules will have already been
1463     // loaded, so we won't be able to capture a delta for them.
1464     log.beginObject("results");
1465     RunSkSLModuleBenchmarks(&log);
1466 
1467     int runs = 0;
1468     BenchmarkStream benchStream;
1469     AutoreleasePool pool;
1470     while (Benchmark* b = benchStream.next()) {
1471         std::unique_ptr<Benchmark> bench(b);
1472         if (CommandLineFlags::ShouldSkip(FLAGS_match, bench->getUniqueName())) {
1473             continue;
1474         }
1475 
1476         if (!configs.empty()) {
1477             log.beginBench(
1478                     bench->getUniqueName(), bench->getSize().width(), bench->getSize().height());
1479             bench->delayedSetup();
1480         }
1481         for (int i = 0; i < configs.size(); ++i) {
1482             Target* target = is_enabled(b, configs[i]);
1483             if (!target) {
1484                 continue;
1485             }
1486 
1487             // During HWUI output this canvas may be nullptr.
1488             SkCanvas* canvas = target->getCanvas();
1489             const char* config = target->config.name.c_str();
1490 
1491             if (FLAGS_pre_log || FLAGS_dryRun) {
1492                 SkDebugf("Running %s\t%s\n"
1493                          , bench->getUniqueName()
1494                          , config);
1495                 if (FLAGS_dryRun) {
1496                     continue;
1497                 }
1498             }
1499 
1500             if (FLAGS_purgeBetweenBenches) {
1501                 SkGraphics::PurgeAllCaches();
1502             }
1503 
1504             if (FLAGS_splitPerfettoTracesByBenchmark) {
1505                 TRACE_EVENT_API_NEW_TRACE_SECTION(TRACE_STR_COPY(bench->getUniqueName()));
1506             }
1507             TRACE_EVENT2("skia", "Benchmark", "name", TRACE_STR_COPY(bench->getUniqueName()),
1508                                               "config", TRACE_STR_COPY(config));
1509 
1510             target->setup();
1511             bench->perCanvasPreDraw(canvas);
1512 
1513             int maxFrameLag;
1514             int loops = target->needsFrameTiming(&maxFrameLag)
1515                 ? setup_gpu_bench(target, bench.get(), maxFrameLag)
1516                 : setup_cpu_bench(overhead, target, bench.get());
1517 
1518             if (kFailedLoops == loops) {
1519                 // Can't be timed.  A warning note has already been printed.
1520                 cleanup_run(target);
1521                 continue;
1522             }
1523 
1524             if (runs == 0 && FLAGS_ms < 1000) {
1525                 // Run the first bench for 1000ms to warm up the nanobench if FLAGS_ms < 1000.
1526                 // Otherwise, the first few benches' measurements will be inaccurate.
1527                 auto stop = now_ms() + 1000;
1528                 do {
1529                     time(loops, bench.get(), target);
1530                     pool.drain();
1531                 } while (now_ms() < stop);
1532             }
1533 
1534             if (FLAGS_ms) {
1535                 samples.clear();
1536                 auto stop = now_ms() + FLAGS_ms;
1537                 do {
1538                     samples.push_back(time(loops, bench.get(), target) / loops);
1539                     pool.drain();
1540                 } while (now_ms() < stop);
1541             } else {
1542                 samples.reset(FLAGS_samples);
1543                 for (int s = 0; s < FLAGS_samples; s++) {
1544                     samples[s] = time(loops, bench.get(), target) / loops;
1545                     pool.drain();
1546                 }
1547             }
1548 
1549             // Scale each result to the benchmark's own units, time/unit.
1550             for (double& sample : samples) {
1551                 sample *= (1.0 / bench->getUnits());
1552             }
1553 
1554             TArray<SkString> keys;
1555             TArray<double> values;
1556             if (configs[i].backend == Benchmark::Backend::kGanesh) {
1557                 if (FLAGS_gpuStatsDump) {
1558                     // TODO cache stats
1559                     bench->getGpuStats(canvas, &keys, &values);
1560                 }
1561                 if (FLAGS_dmsaaStatsDump && bench->getDMSAAStats(canvas->recordingContext())) {
1562                     const auto& dmsaaStats = canvas->recordingContext()->priv().dmsaaStats();
1563                     dmsaaStats.dumpKeyValuePairs(&keys, &values);
1564                     dmsaaStats.dump();
1565                     combinedDMSAAStats.merge(dmsaaStats);
1566                 }
1567             }
1568 
1569             bench->perCanvasPostDraw(canvas);
1570 
1571             if (Benchmark::Backend::kNonRendering != target->config.backend &&
1572                 !FLAGS_writePath.isEmpty() && FLAGS_writePath[0]) {
1573                 SkString pngFilename = SkOSPath::Join(FLAGS_writePath[0], config);
1574                 pngFilename = SkOSPath::Join(pngFilename.c_str(), bench->getUniqueName());
1575                 pngFilename.append(".png");
1576                 write_canvas_png(target, pngFilename);
1577             }
1578 
1579             // Building stats.plot often shows up in profiles,
1580             // so skip building it when we're not going to print it anyway.
1581             const bool want_plot = !FLAGS_quiet && !FLAGS_ms;
1582 
1583             Stats stats(samples, want_plot);
1584             log.beginObject(config);
1585 
1586             log.beginObject("options");
1587             log.appendCString("name", bench->getName());
1588             benchStream.fillCurrentOptions(log);
1589             log.endObject(); // options
1590 
1591             // Metrics
1592             log.appendMetric("min_ms", stats.min);
1593             log.appendMetric("min_ratio", sk_ieee_double_divide(stats.median, stats.min));
1594             log.beginArray("samples");
1595             for (double sample : samples) {
1596                 log.appendDoubleDigits(sample, 16);
1597             }
1598             log.endArray(); // samples
1599             benchStream.fillCurrentMetrics(log);
1600             if (!keys.empty()) {
1601                 // dump to json, only SKPBench currently returns valid keys / values
1602                 SkASSERT(keys.size() == values.size());
1603                 for (int j = 0; j < keys.size(); j++) {
1604                     log.appendMetric(keys[j].c_str(), values[j]);
1605                 }
1606             }
1607 
1608             log.endObject(); // config
1609 
1610             if (runs++ % FLAGS_flushEvery == 0) {
1611                 log.flush();
1612             }
1613 
1614             if (kAutoTuneLoops != FLAGS_loops) {
1615                 if (configs.size() == 1) {
1616                     config = ""; // Only print the config if we run the same bench on more than one.
1617                 }
1618                 SkDebugf("%4d/%-4dMB\t%s\t%s "
1619                          , sk_tools::getCurrResidentSetSizeMB()
1620                          , sk_tools::getMaxResidentSetSizeMB()
1621                          , bench->getUniqueName()
1622                          , config);
1623                 SkDebugf("\n");
1624             } else if (FLAGS_quiet) {
1625                 const char* mark = " ";
1626                 const double stddev_percent =
1627                     sk_ieee_double_divide(100 * sqrt(stats.var), stats.mean);
1628                 if (stddev_percent >  5) mark = "?";
1629                 if (stddev_percent > 10) mark = "!";
1630 
1631                 SkDebugf("%10.2f %s\t%s\t%s\n",
1632                          stats.median*1e3, mark, bench->getUniqueName(), config);
1633             } else if (FLAGS_csv) {
1634                 const double stddev_percent =
1635                     sk_ieee_double_divide(100 * sqrt(stats.var), stats.mean);
1636                 SkDebugf("%g,%g,%g,%g,%g,%s,%s\n"
1637                          , stats.min
1638                          , stats.median
1639                          , stats.mean
1640                          , stats.max
1641                          , stddev_percent
1642                          , config
1643                          , bench->getUniqueName()
1644                          );
1645             } else {
1646                 const double stddev_percent =
1647                     sk_ieee_double_divide(100 * sqrt(stats.var), stats.mean);
1648                 SkDebugf("%4d/%-4dMB\t%d\t%s\t%s\t%s\t%s\t%.0f%%\t%s\t%s\t%s\n"
1649                         , sk_tools::getCurrResidentSetSizeMB()
1650                         , sk_tools::getMaxResidentSetSizeMB()
1651                         , loops
1652                         , HUMANIZE(stats.min)
1653                         , HUMANIZE(stats.median)
1654                         , HUMANIZE(stats.mean)
1655                         , HUMANIZE(stats.max)
1656                         , stddev_percent
1657                         , FLAGS_ms ? to_string(samples.size()).c_str() : stats.plot.c_str()
1658                         , config
1659                         , bench->getUniqueName()
1660                         );
1661             }
1662 
1663             if (FLAGS_gpuStats && Benchmark::Backend::kGanesh == configs[i].backend) {
1664                 target->dumpStats();
1665             }
1666 
1667             if (FLAGS_verbose) {
1668                 SkDebugf("Samples:  ");
1669                 for (int j = 0; j < samples.size(); j++) {
1670                     SkDebugf("%s  ", HUMANIZE(samples[j]));
1671                 }
1672                 SkDebugf("%s\n", bench->getUniqueName());
1673             }
1674             cleanup_run(target);
1675             pool.drain();
1676         }
1677         if (!configs.empty()) {
1678             log.endBench();
1679         }
1680     }
1681 
1682     if (FLAGS_dmsaaStatsDump) {
1683         SkDebugf("<<Total Combined DMSAA Stats>>\n");
1684         combinedDMSAAStats.dump();
1685     }
1686 
1687     SkGraphics::PurgeAllCaches();
1688 
1689     log.beginBench("memory_usage", 0, 0);
1690     log.beginObject("meta"); // config
1691     log.appendS32("max_rss_mb", sk_tools::getMaxResidentSetSizeMB());
1692     log.endObject(); // config
1693     log.endBench();
1694 
1695     log.endObject(); // results
1696     log.endObject(); // root
1697     log.flush();
1698 
1699     return 0;
1700 }
1701