OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2014 Google Inc. | 2 * Copyright 2014 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #include <ctype.h> | 8 #include <ctype.h> |
9 | 9 |
10 #include "Benchmark.h" | 10 #include "Benchmark.h" |
(...skipping 36 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
47 | 47 |
48 DEFINE_bool(cpu, true, "Master switch for CPU-bound work."); | 48 DEFINE_bool(cpu, true, "Master switch for CPU-bound work."); |
49 DEFINE_bool(gpu, true, "Master switch for GPU-bound work."); | 49 DEFINE_bool(gpu, true, "Master switch for GPU-bound work."); |
50 | 50 |
51 DEFINE_string(outResultsFile, "", "If given, write results here as JSON."); | 51 DEFINE_string(outResultsFile, "", "If given, write results here as JSON."); |
52 DEFINE_bool(resetGpuContext, true, "Reset the GrContext before running each benc
h."); | 52 DEFINE_bool(resetGpuContext, true, "Reset the GrContext before running each benc
h."); |
53 DEFINE_int32(maxCalibrationAttempts, 3, | 53 DEFINE_int32(maxCalibrationAttempts, 3, |
54 "Try up to this many times to guess loops for a bench, or skip the
bench."); | 54 "Try up to this many times to guess loops for a bench, or skip the
bench."); |
55 DEFINE_int32(maxLoops, 1000000, "Never run a bench more times than this."); | 55 DEFINE_int32(maxLoops, 1000000, "Never run a bench more times than this."); |
56 | 56 |
| 57 DEFINE_bool2(veryVerbose, V, false, "Print lots for debugging bots."); |
57 | 58 |
58 static SkString humanize(double ms) { | 59 static SkString humanize(double ms) { |
59 if (ms > 1e+3) return SkStringPrintf("%.3gs", ms/1e3); | 60 if (ms > 1e+3) return SkStringPrintf("%.3gs", ms/1e3); |
60 if (ms < 1e-3) return SkStringPrintf("%.3gns", ms*1e6); | 61 if (ms < 1e-3) return SkStringPrintf("%.3gns", ms*1e6); |
61 #ifdef SK_BUILD_FOR_WIN | 62 #ifdef SK_BUILD_FOR_WIN |
62 if (ms < 1) return SkStringPrintf("%.3gus", ms*1e3); | 63 if (ms < 1) return SkStringPrintf("%.3gus", ms*1e3); |
63 #else | 64 #else |
64 if (ms < 1) return SkStringPrintf("%.3gµs", ms*1e3); | 65 if (ms < 1) return SkStringPrintf("%.3gµs", ms*1e3); |
65 #endif | 66 #endif |
66 return SkStringPrintf("%.3gms", ms); | 67 return SkStringPrintf("%.3gms", ms); |
67 } | 68 } |
68 #define HUMANIZE(ms) humanize(ms).c_str() | 69 #define HUMANIZE(ms) humanize(ms).c_str() |
69 | 70 |
70 static double time(int loops, Benchmark* bench, SkCanvas* canvas, SkGLContextHel
per* gl) { | 71 static double time(int loops, Benchmark* bench, SkCanvas* canvas, SkGLContextHel
per* gl) { |
71 WallTimer timer; | 72 WallTimer timer; |
72 timer.start(); | 73 timer.start(); |
73 if (bench) { | 74 if (bench) { |
| 75 if (FLAGS_veryVerbose) { SkDebugf("Timing %s for %d loops.\n", bench->ge
tName(), loops); } |
74 bench->draw(loops, canvas); | 76 bench->draw(loops, canvas); |
75 } | 77 } |
76 if (canvas) { | 78 if (canvas) { |
77 canvas->flush(); | 79 canvas->flush(); |
78 } | 80 } |
79 #if SK_SUPPORT_GPU | 81 #if SK_SUPPORT_GPU |
80 if (gl) { | 82 if (gl) { |
81 SK_GL(*gl, Flush()); | 83 SK_GL(*gl, Flush()); |
82 gl->swapBuffers(); | 84 gl->swapBuffers(); |
83 } | 85 } |
(...skipping 50 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
134 // ------------------------------------------ < N | 136 // ------------------------------------------ < N |
135 // bench_plus_overhead - overhead) | 137 // bench_plus_overhead - overhead) |
136 // | 138 // |
137 // Luckily, this also works well in practice. :) | 139 // Luckily, this also works well in practice. :) |
138 const double numer = overhead / FLAGS_overheadGoal - overhead; | 140 const double numer = overhead / FLAGS_overheadGoal - overhead; |
139 const double denom = bench_plus_overhead - overhead; | 141 const double denom = bench_plus_overhead - overhead; |
140 const int loops = clamp_loops(FLAGS_runOnce ? 1 : (int)ceil(numer / denom)); | 142 const int loops = clamp_loops(FLAGS_runOnce ? 1 : (int)ceil(numer / denom)); |
141 | 143 |
142 for (int i = 0; i < FLAGS_samples; i++) { | 144 for (int i = 0; i < FLAGS_samples; i++) { |
143 samples[i] = time(loops, bench, canvas, NULL) / loops; | 145 samples[i] = time(loops, bench, canvas, NULL) / loops; |
| 146 if (FLAGS_veryVerbose) { SkDebugf(" %s\n", HUMANIZE(samples[i])); } |
144 } | 147 } |
145 return loops; | 148 return loops; |
146 } | 149 } |
147 | 150 |
148 #if SK_SUPPORT_GPU | 151 #if SK_SUPPORT_GPU |
149 static int gpu_bench(SkGLContextHelper* gl, | 152 static int gpu_bench(SkGLContextHelper* gl, |
150 Benchmark* bench, | 153 Benchmark* bench, |
151 SkCanvas* canvas, | 154 SkCanvas* canvas, |
152 double* samples) { | 155 double* samples) { |
153 // Make sure we're done with whatever came before. | 156 // Make sure we're done with whatever came before. |
154 SK_GL(*gl, Finish()); | 157 SK_GL(*gl, Finish()); |
155 | 158 |
156 // First, figure out how many loops it'll take to get a frame up to FLAGS_gp
uMs. | 159 // First, figure out how many loops it'll take to get a frame up to FLAGS_gp
uMs. |
157 int loops = 1; | 160 int loops = 1; |
158 if (!FLAGS_runOnce) { | 161 if (!FLAGS_runOnce) { |
159 double elapsed = 0; | 162 double elapsed = 0; |
160 do { | 163 do { |
161 loops *= 2; | 164 loops *= 2; |
162 // If the GPU lets frames lag at all, we need to make sure we're tim
ing | 165 // If the GPU lets frames lag at all, we need to make sure we're tim
ing |
163 // _this_ round, not still timing last round. We force this by loop
ing | 166 // _this_ round, not still timing last round. We force this by loop
ing |
164 // more times than any reasonable GPU will allow frames to lag. | 167 // more times than any reasonable GPU will allow frames to lag. |
165 for (int i = 0; i < FLAGS_gpuFrameLag; i++) { | 168 for (int i = 0; i < FLAGS_gpuFrameLag; i++) { |
166 elapsed = time(loops, bench, canvas, gl); | 169 elapsed = time(loops, bench, canvas, gl); |
167 } | 170 } |
168 } while (elapsed < FLAGS_gpuMs); | 171 } while (elapsed < FLAGS_gpuMs); |
169 | 172 |
170 // We've overshot at least a little. Scale back linearly. | 173 // We've overshot at least a little. Scale back linearly. |
171 loops = (int)ceil(loops * FLAGS_gpuMs / elapsed); | 174 loops = (int)ceil(loops * FLAGS_gpuMs / elapsed); |
172 | 175 |
| 176 if (FLAGS_veryVerbose) { SkDebugf("elapsed %s, loops %d\n", HUMANIZE(ela
psed), loops); } |
| 177 |
173 // Might as well make sure we're not still timing our calibration. | 178 // Might as well make sure we're not still timing our calibration. |
174 SK_GL(*gl, Finish()); | 179 SK_GL(*gl, Finish()); |
175 } | 180 } |
176 loops = clamp_loops(loops); | 181 loops = clamp_loops(loops); |
177 | 182 |
178 // Pretty much the same deal as the calibration: do some warmup to make | 183 // Pretty much the same deal as the calibration: do some warmup to make |
179 // sure we're timing steady-state pipelined frames. | 184 // sure we're timing steady-state pipelined frames. |
180 for (int i = 0; i < FLAGS_gpuFrameLag; i++) { | 185 for (int i = 0; i < FLAGS_gpuFrameLag; i++) { |
181 time(loops, bench, canvas, gl); | 186 time(loops, bench, canvas, gl); |
182 } | 187 } |
183 | 188 |
184 // Now, actually do the timing! | 189 // Now, actually do the timing! |
185 for (int i = 0; i < FLAGS_samples; i++) { | 190 for (int i = 0; i < FLAGS_samples; i++) { |
186 samples[i] = time(loops, bench, canvas, gl) / loops; | 191 samples[i] = time(loops, bench, canvas, gl) / loops; |
| 192 if (FLAGS_veryVerbose) { SkDebugf(" %s\n", HUMANIZE(samples[i])); } |
187 } | 193 } |
188 return loops; | 194 return loops; |
189 } | 195 } |
190 #endif | 196 #endif |
191 | 197 |
192 static SkString to_lower(const char* str) { | 198 static SkString to_lower(const char* str) { |
193 SkString lower(str); | 199 SkString lower(str); |
194 for (size_t i = 0; i < lower.size(); i++) { | 200 for (size_t i = 0; i < lower.size(); i++) { |
195 lower[i] = tolower(lower[i]); | 201 lower[i] = tolower(lower[i]); |
196 } | 202 } |
(...skipping 138 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
335 | 341 |
336 const int loops = | 342 const int loops = |
337 #if SK_SUPPORT_GPU | 343 #if SK_SUPPORT_GPU |
338 Benchmark::kGPU_Backend == targets[j]->backend | 344 Benchmark::kGPU_Backend == targets[j]->backend |
339 ? gpu_bench(targets[j]->gl, bench.get(), canvas, samples.get()) | 345 ? gpu_bench(targets[j]->gl, bench.get(), canvas, samples.get()) |
340 : | 346 : |
341 #endif | 347 #endif |
342 cpu_bench( overhead, bench.get(), canvas, samples.get()); | 348 cpu_bench( overhead, bench.get(), canvas, samples.get()); |
343 | 349 |
344 if (loops == 0) { | 350 if (loops == 0) { |
345 SkDebugf("Unable to time %s\t%s (overhead %s)\n", | |
346 bench->getName(), config, HUMANIZE(overhead)); | |
347 continue; | 351 continue; |
348 } | 352 } |
349 | 353 |
350 Stats stats(samples.get(), FLAGS_samples); | 354 Stats stats(samples.get(), FLAGS_samples); |
351 log.config(config); | 355 log.config(config); |
352 log.timer("min_ms", stats.min); | 356 log.timer("min_ms", stats.min); |
353 log.timer("median_ms", stats.median); | 357 log.timer("median_ms", stats.median); |
354 log.timer("mean_ms", stats.mean); | 358 log.timer("mean_ms", stats.mean); |
355 log.timer("max_ms", stats.max); | 359 log.timer("max_ms", stats.max); |
356 log.timer("stddev_ms", sqrt(stats.var)); | 360 log.timer("stddev_ms", sqrt(stats.var)); |
(...skipping 38 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
395 } | 399 } |
396 | 400 |
397 return 0; | 401 return 0; |
398 } | 402 } |
399 | 403 |
400 #if !defined SK_BUILD_FOR_IOS | 404 #if !defined SK_BUILD_FOR_IOS |
401 int main(int argc, char * const argv[]) { | 405 int main(int argc, char * const argv[]) { |
402 return tool_main(argc, (char**) argv); | 406 return tool_main(argc, (char**) argv); |
403 } | 407 } |
404 #endif | 408 #endif |
OLD | NEW |