aboutsummaryrefslogtreecommitdiffhomepage
path: root/tools/skpbench/skpbench.cpp
blob: 5f5e96b477e97222c38bd460837b1c3e9e31fc2b (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
/*
 * Copyright 2016 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include <stdlib.h>
#include <algorithm>
#include <array>
#include <chrono>
#include <cmath>
#include <vector>
#include "DDLPromiseImageHelper.h"
#include "DDLTileHelper.h"
#include "GpuTimer.h"
#include "GrCaps.h"
#include "GrContextFactory.h"
#include "GrContextPriv.h"
#include "SkCanvas.h"
#include "SkCommonFlags.h"
#include "SkCommonFlagsGpu.h"
#include "SkDeferredDisplayList.h"
#include "SkGraphics.h"
#include "SkGr.h"
#include "SkOSFile.h"
#include "SkOSPath.h"
#include "SkPerlinNoiseShader.h"
#include "SkPicture.h"
#include "SkPictureRecorder.h"
#include "SkStream.h"
#include "SkSurface.h"
#include "SkSurfaceProps.h"
#include "SkTaskGroup.h"
#include "flags/SkCommandLineFlags.h"
#include "flags/SkCommonFlagsConfig.h"
#include "picture_utils.h"
#include "sk_tool_utils.h"

/**
 * This is a minimalist program whose sole purpose is to open an skp file, benchmark it on a single
 * config, and exit. It is intended to be used through skpbench.py rather than invoked directly.
 * Limiting the entire process to a single config/skp pair helps to keep the results repeatable.
 *
 * No tiling, looping, or other fanciness is used; it just draws the skp whole into a size-matched
 * render target and syncs the GPU after each draw.
 *
 * Currently, only GPU configs are supported.
 */

DEFINE_bool(ddl, false, "record the skp into DDLs before rendering");
DEFINE_int32(ddlNumAdditionalThreads, 0, "number of DDL recording threads in addition to main one");
DEFINE_int32(ddlTilingWidthHeight, 0, "number of tiles along one edge when in DDL mode");
DEFINE_bool(ddlRecordTime, false, "report just the cpu time spent recording DDLs");

DEFINE_int32(duration, 5000, "number of milliseconds to run the benchmark");
DEFINE_int32(sampleMs, 50, "minimum duration of a sample");
DEFINE_bool(gpuClock, false, "time on the gpu clock (gpu work only)");
DEFINE_bool(fps, false, "use fps instead of ms");
DEFINE_string(skp, "", "path to a single .skp file, or 'warmup' for a builtin warmup run");
DEFINE_string(png, "", "if set, save a .png proof to disk at this file location");
DEFINE_int32(verbosity, 4, "level of verbosity (0=none to 5=debug)");
DEFINE_bool(suppressHeader, false, "don't print a header row before the results");

static const char* header =
"   accum    median       max       min   stddev  samples  sample_ms  clock  metric  config    bench";

static const char* resultFormat =
"%8.4g  %8.4g  %8.4g  %8.4g  %6.3g%%  %7li  %9i  %-5s  %-6s  %-9s %s";

struct Sample {
    using duration = std::chrono::nanoseconds;

    Sample() : fFrames(0), fDuration(0) {}
    double seconds() const { return std::chrono::duration<double>(fDuration).count(); }
    double ms() const { return std::chrono::duration<double, std::milli>(fDuration).count(); }
    double value() const { return FLAGS_fps ? fFrames / this->seconds() : this->ms() / fFrames; }
    static const char* metric() { return FLAGS_fps ? "fps" : "ms"; }

    int        fFrames;
    duration   fDuration;
};

class GpuSync {
public:
    GpuSync(const sk_gpu_test::FenceSync* fenceSync);
    ~GpuSync();

    void syncToPreviousFrame();

private:
    void updateFence();

    const sk_gpu_test::FenceSync* const   fFenceSync;
    sk_gpu_test::PlatformFence            fFence;
};

enum class ExitErr {
    kOk           = 0,
    kUsage        = 64,
    kData         = 65,
    kUnavailable  = 69,
    kIO           = 74,
    kSoftware     = 70
};

static void draw_skp_and_flush(SkCanvas*, const SkPicture*);
static sk_sp<SkPicture> create_warmup_skp();
static bool mkdir_p(const SkString& name);
static SkString join(const SkCommandLineFlags::StringArray&);
static void exitf(ExitErr, const char* format, ...);

static void ddl_sample(GrContext* context, DDLTileHelper* tiles, GpuSync* gpuSync, Sample* sample,
                       std::chrono::high_resolution_clock::time_point* startStopTime) {
    using clock = std::chrono::high_resolution_clock;

    clock::time_point start = *startStopTime;

    tiles->createDDLsInParallel();

    if (!FLAGS_ddlRecordTime) {
        tiles->drawAllTilesAndFlush(context, true);
        if (gpuSync) {
            gpuSync->syncToPreviousFrame();
        }
    }

    *startStopTime = clock::now();

    tiles->resetAllTiles();

    if (sample) {
        SkASSERT(gpuSync);
        sample->fDuration += *startStopTime - start;
        sample->fFrames++;
    }
}

static void run_ddl_benchmark(const sk_gpu_test::FenceSync* fenceSync,
                              GrContext* context, SkCanvas* finalCanvas,
                              SkPicture* inputPicture, std::vector<Sample>* samples) {
    using clock = std::chrono::high_resolution_clock;
    const Sample::duration sampleDuration = std::chrono::milliseconds(FLAGS_sampleMs);
    const clock::duration benchDuration = std::chrono::milliseconds(FLAGS_duration);

    SkIRect viewport = finalCanvas->imageInfo().bounds();

    DDLPromiseImageHelper promiseImageHelper;
    sk_sp<SkData> compressedPictureData = promiseImageHelper.deflateSKP(inputPicture);
    if (!compressedPictureData) {
        exitf(ExitErr::kUnavailable, "DDL: conversion of skp failed");
    }

    promiseImageHelper.uploadAllToGPU(context);

    DDLTileHelper tiles(finalCanvas, viewport, FLAGS_ddlTilingWidthHeight);

    tiles.createSKPPerTile(compressedPictureData.get(), promiseImageHelper);

    clock::time_point startStopTime = clock::now();

    ddl_sample(context, &tiles, nullptr, nullptr, &startStopTime);
    GpuSync gpuSync(fenceSync);
    ddl_sample(context, &tiles, &gpuSync, nullptr, &startStopTime);

    clock::duration cumulativeDuration = std::chrono::milliseconds(0);

    do {
        samples->emplace_back();
        Sample& sample = samples->back();

        do {
            ddl_sample(context, &tiles, &gpuSync, &sample, &startStopTime);
        } while (sample.fDuration < sampleDuration);

        cumulativeDuration += sample.fDuration;
    } while (cumulativeDuration < benchDuration || 0 == samples->size() % 2);

    if (!FLAGS_png.isEmpty()) {
        // The user wants to see the final result
        tiles.composeAllTiles(finalCanvas);
    }
}

static void run_benchmark(const sk_gpu_test::FenceSync* fenceSync, SkCanvas* canvas,
                          const SkPicture* skp, std::vector<Sample>* samples) {
    using clock = std::chrono::high_resolution_clock;
    const Sample::duration sampleDuration = std::chrono::milliseconds(FLAGS_sampleMs);
    const clock::duration benchDuration = std::chrono::milliseconds(FLAGS_duration);

    draw_skp_and_flush(canvas, skp); // draw1
    GpuSync gpuSync(fenceSync);

    draw_skp_and_flush(canvas, skp); // draw2
    gpuSync.syncToPreviousFrame();   // waits for draw1 to finish (after draw2's cpu work is done).

    clock::time_point now = clock::now();
    const clock::time_point endTime = now + benchDuration;

    do {
        clock::time_point sampleStart = now;
        samples->emplace_back();
        Sample& sample = samples->back();

        do {
            draw_skp_and_flush(canvas, skp);
            gpuSync.syncToPreviousFrame();

            now = clock::now();
            sample.fDuration = now - sampleStart;
            ++sample.fFrames;
        } while (sample.fDuration < sampleDuration);
    } while (now < endTime || 0 == samples->size() % 2);
}

static void run_gpu_time_benchmark(sk_gpu_test::GpuTimer* gpuTimer,
                                   const sk_gpu_test::FenceSync* fenceSync, SkCanvas* canvas,
                                   const SkPicture* skp, std::vector<Sample>* samples) {
    using sk_gpu_test::PlatformTimerQuery;
    using clock = std::chrono::steady_clock;
    const clock::duration sampleDuration = std::chrono::milliseconds(FLAGS_sampleMs);
    const clock::duration benchDuration = std::chrono::milliseconds(FLAGS_duration);

    if (!gpuTimer->disjointSupport()) {
        fprintf(stderr, "WARNING: GPU timer cannot detect disjoint operations; "
                        "results may be unreliable\n");
    }

    draw_skp_and_flush(canvas, skp);
    GpuSync gpuSync(fenceSync);

    gpuTimer->queueStart();
    draw_skp_and_flush(canvas, skp);
    PlatformTimerQuery previousTime = gpuTimer->queueStop();
    gpuSync.syncToPreviousFrame();

    clock::time_point now = clock::now();
    const clock::time_point endTime = now + benchDuration;

    do {
        const clock::time_point sampleEndTime = now + sampleDuration;
        samples->emplace_back();
        Sample& sample = samples->back();

        do {
            gpuTimer->queueStart();
            draw_skp_and_flush(canvas, skp);
            PlatformTimerQuery time = gpuTimer->queueStop();
            gpuSync.syncToPreviousFrame();

            switch (gpuTimer->checkQueryStatus(previousTime)) {
                using QueryStatus = sk_gpu_test::GpuTimer::QueryStatus;
                case QueryStatus::kInvalid:
                    exitf(ExitErr::kUnavailable, "GPU timer failed");
                case QueryStatus::kPending:
                    exitf(ExitErr::kUnavailable, "timer query still not ready after fence sync");
                case QueryStatus::kDisjoint:
                    if (FLAGS_verbosity >= 4) {
                        fprintf(stderr, "discarding timer query due to disjoint operations.\n");
                    }
                    break;
                case QueryStatus::kAccurate:
                    sample.fDuration += gpuTimer->getTimeElapsed(previousTime);
                    ++sample.fFrames;
                    break;
            }
            gpuTimer->deleteQuery(previousTime);
            previousTime = time;
            now = clock::now();
        } while (now < sampleEndTime || 0 == sample.fFrames);
    } while (now < endTime || 0 == samples->size() % 2);

    gpuTimer->deleteQuery(previousTime);
}

void print_result(const std::vector<Sample>& samples, const char* config, const char* bench)  {
    if (0 == (samples.size() % 2)) {
        exitf(ExitErr::kSoftware, "attempted to gather stats on even number of samples");
    }

    Sample accum = Sample();
    std::vector<double> values;
    values.reserve(samples.size());
    for (const Sample& sample : samples) {
        accum.fFrames += sample.fFrames;
        accum.fDuration += sample.fDuration;
        values.push_back(sample.value());
    }
    std::sort(values.begin(), values.end());

    const double accumValue = accum.value();
    double variance = 0;
    for (double value : values) {
        const double delta = value - accumValue;
        variance += delta * delta;
    }
    variance /= values.size();
    // Technically, this is the relative standard deviation.
    const double stddev = 100/*%*/ * sqrt(variance) / accumValue;

    printf(resultFormat, accumValue, values[values.size() / 2], values.back(), values.front(),
           stddev, values.size(), FLAGS_sampleMs, FLAGS_gpuClock ? "gpu" : "cpu", Sample::metric(),
           config, bench);
    printf("\n");
    fflush(stdout);
}

int main(int argc, char** argv) {
    SkCommandLineFlags::SetUsage("Use skpbench.py instead. "
                                 "You usually don't want to use this program directly.");
    SkCommandLineFlags::Parse(argc, argv);

    if (!FLAGS_suppressHeader) {
        printf("%s\n", header);
    }
    if (FLAGS_duration <= 0) {
        exit(0); // This can be used to print the header and quit.
    }

    // Parse the config.
    const SkCommandLineConfigGpu* config = nullptr; // Initialize for spurious warning.
    SkCommandLineConfigArray configs;
    ParseConfigs(FLAGS_config, &configs);
    if (configs.count() != 1 || !(config = configs[0]->asConfigGpu())) {
        exitf(ExitErr::kUsage, "invalid config '%s': must specify one (and only one) GPU config",
                               join(FLAGS_config).c_str());
    }

    // Parse the skp.
    if (FLAGS_skp.count() != 1) {
        exitf(ExitErr::kUsage, "invalid skp '%s': must specify a single skp file, or 'warmup'",
                               join(FLAGS_skp).c_str());
    }

    SkGraphics::Init();
    SkTaskGroup::Enabler enabled(FLAGS_ddlNumAdditionalThreads);

    sk_sp<SkPicture> skp;
    SkString skpname;
    if (0 == strcmp(FLAGS_skp[0], "warmup")) {
        skp = create_warmup_skp();
        skpname = "warmup";
    } else {
        const char* skpfile = FLAGS_skp[0];
        std::unique_ptr<SkStream> skpstream(SkStream::MakeFromFile(skpfile));
        if (!skpstream) {
            exitf(ExitErr::kIO, "failed to open skp file %s", skpfile);
        }
        skp = SkPicture::MakeFromStream(skpstream.get());
        if (!skp) {
            exitf(ExitErr::kData, "failed to parse skp file %s", skpfile);
        }
        skpname = SkOSPath::Basename(skpfile);
    }
    int width = SkTMin(SkScalarCeilToInt(skp->cullRect().width()), 2048),
        height = SkTMin(SkScalarCeilToInt(skp->cullRect().height()), 2048);
    if (FLAGS_verbosity >= 3 &&
        (width != skp->cullRect().width() || height != skp->cullRect().height())) {
        fprintf(stderr, "%s is too large (%ix%i), cropping to %ix%i.\n",
                        skpname.c_str(), SkScalarCeilToInt(skp->cullRect().width()),
                        SkScalarCeilToInt(skp->cullRect().height()), width, height);
    }

    if (config->getSurfType() != SkCommandLineConfigGpu::SurfType::kDefault) {
        exitf(ExitErr::kUnavailable, "This tool only supports the default surface type. (%s)",
              config->getTag().c_str());
    }

    // Create a context.
    GrContextOptions ctxOptions;
    SetCtxOptionsFromCommonFlags(&ctxOptions);
    sk_gpu_test::GrContextFactory factory(ctxOptions);
    sk_gpu_test::ContextInfo ctxInfo =
        factory.getContextInfo(config->getContextType(), config->getContextOverrides());
    GrContext* ctx = ctxInfo.grContext();
    if (!ctx) {
        exitf(ExitErr::kUnavailable, "failed to create context for config %s",
                                     config->getTag().c_str());
    }
    if (ctx->maxRenderTargetSize() < SkTMax(width, height)) {
        exitf(ExitErr::kUnavailable, "render target size %ix%i not supported by platform (max: %i)",
              width, height, ctx->maxRenderTargetSize());
    }
    GrPixelConfig grPixConfig = SkColorType2GrPixelConfig(config->getColorType());
    if (kUnknown_GrPixelConfig == grPixConfig) {
        exitf(ExitErr::kUnavailable, "failed to get GrPixelConfig from SkColorType: %d",
                                     config->getColorType());
    }
    int supportedSampleCount = ctx->contextPriv().caps()->getRenderTargetSampleCount(
            config->getSamples(), grPixConfig);
    if (supportedSampleCount != config->getSamples()) {
        exitf(ExitErr::kUnavailable, "sample count %i not supported by platform",
                                     config->getSamples());
    }
    sk_gpu_test::TestContext* testCtx = ctxInfo.testContext();
    if (!testCtx) {
        exitf(ExitErr::kSoftware, "testContext is null");
    }
    if (!testCtx->fenceSyncSupport()) {
        exitf(ExitErr::kUnavailable, "GPU does not support fence sync");
    }

    // Create a render target.
    SkImageInfo info =
            SkImageInfo::Make(width, height, config->getColorType(), config->getAlphaType(),
                              sk_ref_sp(config->getColorSpace()));
    uint32_t flags = config->getUseDIText() ? SkSurfaceProps::kUseDeviceIndependentFonts_Flag : 0;
    SkSurfaceProps props(flags, SkSurfaceProps::kLegacyFontHost_InitType);
    sk_sp<SkSurface> surface =
        SkSurface::MakeRenderTarget(ctx, SkBudgeted::kNo, info, config->getSamples(), &props);
    if (!surface) {
        exitf(ExitErr::kUnavailable, "failed to create %ix%i render target for config %s",
                                     width, height, config->getTag().c_str());
    }

    // Run the benchmark.
    std::vector<Sample> samples;
    if (FLAGS_sampleMs > 0) {
        // +1 because we might take one more sample in order to have an odd number.
        samples.reserve(1 + (FLAGS_duration + FLAGS_sampleMs - 1) / FLAGS_sampleMs);
    } else {
        samples.reserve(2 * FLAGS_duration);
    }
    SkCanvas* canvas = surface->getCanvas();
    canvas->translate(-skp->cullRect().x(), -skp->cullRect().y());
    if (!FLAGS_gpuClock) {
        if (FLAGS_ddl) {
            run_ddl_benchmark(testCtx->fenceSync(), ctx, canvas, skp.get(), &samples);
        } else {
            run_benchmark(testCtx->fenceSync(), canvas, skp.get(), &samples);
        }
    } else {
        if (FLAGS_ddl) {
            exitf(ExitErr::kUnavailable, "DDL: GPU-only timing not supported");
        }
        if (!testCtx->gpuTimingSupport()) {
            exitf(ExitErr::kUnavailable, "GPU does not support timing");
        }
        run_gpu_time_benchmark(testCtx->gpuTimer(), testCtx->fenceSync(), canvas, skp.get(),
                               &samples);
    }
    print_result(samples, config->getTag().c_str(), skpname.c_str());

    // Save a proof (if one was requested).
    if (!FLAGS_png.isEmpty()) {
        SkBitmap bmp;
        bmp.allocPixels(info);
        if (!surface->getCanvas()->readPixels(bmp, 0, 0)) {
            exitf(ExitErr::kUnavailable, "failed to read canvas pixels for png");
        }
        const SkString &dirname = SkOSPath::Dirname(FLAGS_png[0]),
                       &basename = SkOSPath::Basename(FLAGS_png[0]);
        if (!mkdir_p(dirname)) {
            exitf(ExitErr::kIO, "failed to create directory \"%s\" for png", dirname.c_str());
        }
        if (!sk_tools::write_bitmap_to_disk(bmp, dirname, nullptr, basename)) {
            exitf(ExitErr::kIO, "failed to save png to \"%s\"", FLAGS_png[0]);
        }
    }

    exit(0);
}

static void draw_skp_and_flush(SkCanvas* canvas, const SkPicture* skp) {
    canvas->drawPicture(skp);
    canvas->flush();
}

static sk_sp<SkPicture> create_warmup_skp() {
    static constexpr SkRect bounds{0, 0, 500, 500};
    SkPictureRecorder recorder;
    SkCanvas* recording = recorder.beginRecording(bounds);

    recording->clear(SK_ColorWHITE);

    SkPaint stroke;
    stroke.setStyle(SkPaint::kStroke_Style);
    stroke.setStrokeWidth(2);

    // Use a big path to (theoretically) warmup the CPU.
    SkPath bigPath;
    sk_tool_utils::make_big_path(bigPath);
    recording->drawPath(bigPath, stroke);

    // Use a perlin shader to warmup the GPU.
    SkPaint perlin;
    perlin.setShader(SkPerlinNoiseShader::MakeTurbulence(0.1f, 0.1f, 1, 0, nullptr));
    recording->drawRect(bounds, perlin);

    return recorder.finishRecordingAsPicture();
}

bool mkdir_p(const SkString& dirname) {
    if (dirname.isEmpty()) {
        return true;
    }
    return mkdir_p(SkOSPath::Dirname(dirname.c_str())) && sk_mkdir(dirname.c_str());
}

static SkString join(const SkCommandLineFlags::StringArray& stringArray) {
    SkString joined;
    for (int i = 0; i < stringArray.count(); ++i) {
        joined.appendf(i ? " %s" : "%s", stringArray[i]);
    }
    return joined;
}

static void exitf(ExitErr err, const char* format, ...) {
    fprintf(stderr, ExitErr::kSoftware == err ? "INTERNAL ERROR: " : "ERROR: ");
    va_list args;
    va_start(args, format);
    vfprintf(stderr, format, args);
    va_end(args);
    fprintf(stderr, ExitErr::kSoftware == err ? "; this should never happen.\n": ".\n");
    exit((int)err);
}

GpuSync::GpuSync(const sk_gpu_test::FenceSync* fenceSync)
    : fFenceSync(fenceSync) {
    this->updateFence();
}

GpuSync::~GpuSync() {
    fFenceSync->deleteFence(fFence);
}

void GpuSync::syncToPreviousFrame() {
    if (sk_gpu_test::kInvalidFence == fFence) {
        exitf(ExitErr::kSoftware, "attempted to sync with invalid fence");
    }
    if (!fFenceSync->waitFence(fFence)) {
        exitf(ExitErr::kUnavailable, "failed to wait for fence");
    }
    fFenceSync->deleteFence(fFence);
    this->updateFence();
}

void GpuSync::updateFence() {
    fFence = fFenceSync->insertFence();
    if (sk_gpu_test::kInvalidFence == fFence) {
        exitf(ExitErr::kUnavailable, "failed to insert fence");
    }
}