aboutsummaryrefslogtreecommitdiffhomepage
diff options
context:
space:
mode:
Diffstat
-rw-r--r--bench/nanobench.cpp82
-rw-r--r--tools/Stats.h17
-rw-r--r--tools/VisualBench.cpp2
3 files changed, 76 insertions, 25 deletions
diff --git a/bench/nanobench.cpp b/bench/nanobench.cpp
index 9f3887bc20..4aad61da72 100644
--- a/bench/nanobench.cpp
+++ b/bench/nanobench.cpp
@@ -55,6 +55,8 @@
__SK_FORCE_IMAGE_DECODER_LINKING;
+static const int kTimedSampling = 0;
+
static const int kAutoTuneLoops = 0;
static const int kDefaultLoops =
@@ -72,9 +74,17 @@ static SkString loops_help_txt() {
return help;
}
+static SkString to_string(int n) {
+ SkString str;
+ str.appendS32(n);
+ return str;
+}
+
DEFINE_int32(loops, kDefaultLoops, loops_help_txt().c_str());
DEFINE_int32(samples, 10, "Number of samples to measure for each bench.");
+DEFINE_string(samplingTime, "0", "Amount of time to run each bench. Takes precedence over samples."
+ "Must be \"0\", \"%%lfs\", or \"%%lfms\"");
DEFINE_int32(overheadLoops, 100000, "Loops to estimate timer overhead.");
DEFINE_double(overheadGoal, 0.0001,
"Loop until timer overhead is at most this fraction of our measurments.");
@@ -269,7 +279,7 @@ static bool write_canvas_png(Target* target, const SkString& filename) {
}
static int kFailedLoops = -2;
-static int cpu_bench(const double overhead, Target* target, Benchmark* bench, double* samples) {
+static int setup_cpu_bench(const double overhead, Target* target, Benchmark* bench) {
// First figure out approximately how many loops of bench it takes to make overhead negligible.
double bench_plus_overhead = 0.0;
int round = 0;
@@ -310,16 +320,10 @@ static int cpu_bench(const double overhead, Target* target, Benchmark* bench, do
loops = detect_forever_loops(loops);
}
- for (int i = 0; i < FLAGS_samples; i++) {
- samples[i] = time(loops, bench, target) / loops;
- }
return loops;
}
-static int gpu_bench(Target* target,
- Benchmark* bench,
- double* samples,
- int maxGpuFrameLag) {
+static int setup_gpu_bench(Target* target, Benchmark* bench, int maxGpuFrameLag) {
// First, figure out how many loops it'll take to get a frame up to FLAGS_gpuMs.
int loops = bench->calculateLoops(FLAGS_loops);
if (kAutoTuneLoops == loops) {
@@ -355,11 +359,6 @@ static int gpu_bench(Target* target,
time(loops, bench, target);
}
- // Now, actually do the timing!
- for (int i = 0; i < FLAGS_samples; i++) {
- samples[i] = time(loops, bench, target) / loops;
- }
-
return loops;
}
@@ -946,6 +945,24 @@ int nanobench_main() {
FLAGS_verbose = true;
}
+ double samplingTimeMs = 0;
+ if (0 != strcmp("0", FLAGS_samplingTime[0])) {
+ SkSTArray<8, char> timeUnit;
+ timeUnit.push_back_n(static_cast<int>(strlen(FLAGS_samplingTime[0])) + 1);
+ if (2 != sscanf(FLAGS_samplingTime[0], "%lf%s", &samplingTimeMs, timeUnit.begin()) ||
+ (0 != strcmp("s", timeUnit.begin()) && 0 != strcmp("ms", timeUnit.begin()))) {
+ SkDebugf("Invalid --samplingTime \"%s\". Must be \"0\", \"%%lfs\", or \"%%lfms\"\n",
+ FLAGS_samplingTime[0]);
+ exit(0);
+ }
+ if (0 == strcmp("s", timeUnit.begin())) {
+ samplingTimeMs *= 1000;
+ }
+ if (samplingTimeMs) {
+ FLAGS_samples = kTimedSampling;
+ }
+ }
+
if (kAutoTuneLoops != FLAGS_loops) {
FLAGS_samples = 1;
FLAGS_gpuFrameLag = 0;
@@ -983,7 +1000,7 @@ int nanobench_main() {
const double overhead = estimate_timer_overhead();
SkDebugf("Timer overhead: %s\n", HUMANIZE(overhead));
- SkAutoTMalloc<double> samples(FLAGS_samples);
+ SkTArray<double> samples;
if (kAutoTuneLoops != FLAGS_loops) {
SkDebugf("Fixed number of loops; times would only be misleading so we won't print them.\n");
@@ -991,6 +1008,8 @@ int nanobench_main() {
// No header.
} else if (FLAGS_quiet) {
SkDebugf("median\tbench\tconfig\n");
+ } else if (kTimedSampling == FLAGS_samples) {
+ SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\tsamples\tconfig\tbench\n");
} else {
SkDebugf("curr/maxrss\tloops\tmin\tmedian\tmean\tmax\tstddev\t%-*s\tconfig\tbench\n",
FLAGS_samples, "samples");
@@ -1022,11 +1041,29 @@ int nanobench_main() {
targets[j]->setup();
bench->perCanvasPreDraw(canvas);
- int frameLag;
- const int loops =
- targets[j]->needsFrameTiming(&frameLag)
- ? gpu_bench(targets[j], bench.get(), samples.get(), frameLag)
- : cpu_bench(overhead, targets[j], bench.get(), samples.get());
+ int maxFrameLag;
+ const int loops = targets[j]->needsFrameTiming(&maxFrameLag)
+ ? setup_gpu_bench(targets[j], bench.get(), maxFrameLag)
+ : setup_cpu_bench(overhead, targets[j], bench.get());
+
+ if (kTimedSampling != FLAGS_samples) {
+ samples.reset(FLAGS_samples);
+ for (int s = 0; s < FLAGS_samples; s++) {
+ samples[s] = time(loops, bench, targets[j]) / loops;
+ }
+ } else if (samplingTimeMs) {
+ samples.reset();
+ if (FLAGS_verbose) {
+ SkDebugf("Begin sampling %s for %ims\n",
+ bench->getUniqueName(), static_cast<int>(samplingTimeMs));
+ }
+ WallTimer timer;
+ timer.start();
+ do {
+ samples.push_back(time(loops, bench, targets[j]) / loops);
+ timer.end();
+ } while (timer.fWall < samplingTimeMs);
+ }
bench->perCanvasPostDraw(canvas);
@@ -1043,7 +1080,7 @@ int nanobench_main() {
continue;
}
- Stats stats(samples.get(), FLAGS_samples);
+ Stats stats(samples);
log->config(config);
log->configOption("name", bench->getName());
benchStream.fillCurrentOptions(log.get());
@@ -1063,7 +1100,7 @@ int nanobench_main() {
, bench->getUniqueName()
, config);
} else if (FLAGS_verbose) {
- for (int i = 0; i < FLAGS_samples; i++) {
+ for (int i = 0; i < samples.count(); i++) {
SkDebugf("%s ", HUMANIZE(samples[i]));
}
SkDebugf("%s\n", bench->getUniqueName());
@@ -1083,7 +1120,8 @@ int nanobench_main() {
, HUMANIZE(stats.mean)
, HUMANIZE(stats.max)
, stddev_percent
- , stats.plot.c_str()
+ , kTimedSampling != FLAGS_samples ? stats.plot.c_str()
+ : to_string(samples.count()).c_str()
, config
, bench->getUniqueName()
);
diff --git a/tools/Stats.h b/tools/Stats.h
index 8487a9497d..12c1d35e1f 100644
--- a/tools/Stats.h
+++ b/tools/Stats.h
@@ -1,3 +1,10 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
#ifndef Stats_DEFINED
#define Stats_DEFINED
@@ -11,7 +18,13 @@
#endif
struct Stats {
- Stats(const double samples[], int n) {
+ Stats(const SkTArray<double>& samples) {
+ int n = samples.count();
+ if (!n) {
+ min = max = mean = var = median = 0;
+ return;
+ }
+
min = samples[0];
max = samples[0];
for (int i = 0; i < n; i++) {
@@ -32,7 +45,7 @@ struct Stats {
var = err / (n-1);
SkAutoTMalloc<double> sorted(n);
- memcpy(sorted.get(), samples, n * sizeof(double));
+ memcpy(sorted.get(), samples.begin(), n * sizeof(double));
SkTQSort(sorted.get(), sorted.get() + n - 1);
median = sorted[n/2];
diff --git a/tools/VisualBench.cpp b/tools/VisualBench.cpp
index cbc8994592..ac53b43167 100644
--- a/tools/VisualBench.cpp
+++ b/tools/VisualBench.cpp
@@ -152,7 +152,7 @@ void VisualBench::printStats() {
SkDebugf("%s\n", shortName.c_str());
} else {
SkASSERT(measurements.count());
- Stats stats(measurements.begin(), measurements.count());
+ Stats stats(measurements);
const double stdDevPercent = 100 * sqrt(stats.var) / stats.mean;
SkDebugf("%4d/%-4dMB\t%d\t%d\t%s\t%s\t%s\t%s\t%.0f%%\t%s\n",
sk_tools::getCurrResidentSetSizeMB(),
generated by cgit v1.2.3 (git 2.39.1) at 2024-06-22 22:57:22 +0000