int main(int argc, char *argv[])
{
int rows = SIZE;
int cols = SIZE;
float density = DENSITY;
BenchTimer timer;
#if 1
EigenSparseTriMatrix sm1(rows,cols);
typedef Matrix<Scalar,Dynamic,1> DenseVector;
DenseVector b = DenseVector::Random(cols);
DenseVector x = DenseVector::Random(cols);
bool densedone = false;
for (float density = DENSITY; density>=MINDENSITY; density*=0.5)
{
EigenSparseTriMatrix sm1(rows, cols);
fillMatrix(density, rows, cols, sm1);
// dense matrices
#ifdef DENSEMATRIX
if (!densedone)
{
densedone = true;
std::cout << "Eigen Dense\t" << density*100 << "%\n";
DenseMatrix m1(rows,cols);
Matrix<Scalar,Dynamic,Dynamic,Dynamic,Dynamic,RowMajorBit> m2(rows,cols);
eiToDense(sm1, m1);
m2 = m1;
BENCH(x = m1.marked<UpperTriangular>().solveTriangular(b);)
std::cout << " colmajor^-1 * b:\t" << timer.value() << endl;
// std::cerr << x.transpose() << "\n";
BENCH(x = m2.marked<UpperTriangular>().solveTriangular(b);)
static void bench_record(SkPicture* src, const char* name, SkBBHFactory* bbhFactory) {
BenchTimer timer;
timer.start();
const int width = src ? src->width() : FLAGS_nullSize;
const int height = src ? src->height() : FLAGS_nullSize;
for (int i = 0; i < FLAGS_loops; i++) {
if (FLAGS_skr) {
EXPERIMENTAL::SkRecording recording(width, height);
if (NULL != src) {
src->draw(recording.canvas());
}
// Release and delete the SkPlayback so that recording optimizes its SkRecord.
SkDELETE(recording.releasePlayback());
} else {
SkPictureRecorder recorder;
SkCanvas* canvas = recorder.beginRecording(width, height, bbhFactory, FLAGS_flags);
if (NULL != src) {
src->draw(canvas);
}
if (FLAGS_endRecording) {
SkAutoTUnref<SkPicture> dst(recorder.endRecording());
}
}
}
timer.end();
const double msPerLoop = timer.fCpu / (double)FLAGS_loops;
printf("%f\t%s\n", scale_time(msPerLoop), name);
}
int main()
{
BenchTimer t;
int tries = 10;
int rep = 400000;
typedef Matrix3f Mat;
typedef Vector3f Vec;
Mat A = Mat::Random(3,3);
A = A.adjoint() * A;
SelfAdjointEigenSolver<Mat> eig(A);
BENCH(t, tries, rep, eig.compute(A));
std::cout << "Eigen: " << t.best() << "s\n";
Mat evecs;
Vec evals;
BENCH(t, tries, rep, eigen33(A,evecs,evals));
std::cout << "Direct: " << t.best() << "s\n\n";
std::cerr << "Eigenvalue/eigenvector diffs:\n";
std::cerr << (evals - eig.eigenvalues()).transpose() << "\n";
for(int k=0;k<3;++k)
if(evecs.col(k).dot(eig.eigenvectors().col(k))<0)
evecs.col(k) = -evecs.col(k);
std::cerr << evecs - eig.eigenvectors() << "\n\n";
}
int main(int argc, char *argv[])
{
int rows = SIZE;
int cols = SIZE;
float density = DENSITY;
EigenSparseMatrix sm1(rows,cols), sm3(rows,cols);
BenchTimer timer;
for (float density = DENSITY; density>=MINDENSITY; density*=0.5)
{
fillMatrix(density, rows, cols, sm1);
// dense matrices
#ifdef DENSEMATRIX
{
DenseMatrix m1(rows,cols), m3(rows,cols);
eiToDense(sm1, m1);
BENCH(for (int k=0; k<REPEAT; ++k) m3 = m1.transpose();)
std::cout << " Eigen dense:\t" << timer.value() << endl;
}
#endif
std::cout << "Non zeros: " << sm1.nonZeros()/float(sm1.rows()*sm1.cols())*100 << "%\n";
// eigen sparse matrices
{
BENCH(for (int k=0; k<REPEAT; ++k) sm3 = sm1.transpose();)
std::cout << " Eigen:\t" << timer.value() << endl;
}
int main(int argc, char *argv[])
{
int rows = SIZE;
int cols = SIZE;
bool fullyrand = true;
BenchTimer timer;
Coordinates coords;
Values values;
if(fullyrand)
{
Coordinates pool;
pool.reserve(cols*NBPERROW);
std::cerr << "fill pool" << "\n";
for (int i=0; i<cols*NBPERROW; )
{
// DynamicSparseMatrix<int> stencil(SIZE,SIZE);
Vector2i ij(ei_random<int>(0,rows-1),ei_random<int>(0,cols-1));
// if(stencil.coeffRef(ij.x(), ij.y())==0)
{
// stencil.coeffRef(ij.x(), ij.y()) = 1;
pool.push_back(ij);
}
++i;
}
std::cerr << "pool ok" << "\n";
int n = cols*NBPERROW*KK;
coords.reserve(n);
values.reserve(n);
for (int i=0; i<n; ++i)
{
int i = ei_random<int>(0,pool.size());
coords.push_back(pool[i]);
values.push_back(ei_random<Scalar>());
}
}
else
{
for (int j=0; j<cols; ++j)
for (int i=0; i<NBPERROW; ++i)
{
coords.push_back(Vector2i(ei_random<int>(0,rows-1),j));
values.push_back(ei_random<Scalar>());
}
}
std::cout << "nnz = " << coords.size() << "\n";
CHECK_MEM
// dense matrices
#ifdef DENSEMATRIX
{
BENCH(setrand_eigen_dense(coords,values);)
std::cout << "Eigen Dense\t" << timer.value() << "\n";
}
// Copy samples from archive with index_name
// to new index copy_name.
// Uses all samples in source archive or [start ... end[ .
void copy(const stdString &index_name, const stdString ©_name,
int RTreeM, const epicsTime *start, const epicsTime *end,
const stdString &single_name)
{
IndexFile index(RTreeM), new_index(RTreeM);
IndexFile::NameIterator names;
size_t channel_count = 0, value_count = 0, back_count = 0;
BenchTimer timer;
stdString dir1, dir2;
Filename::getDirname(index_name, dir1);
Filename::getDirname(copy_name, dir2);
if (dir1 == dir2)
{
printf("You have to assert that the new index (%s)\n"
"is in a directory different from the old index\n"
"(%s)\n", copy_name.c_str(), index_name.c_str());
return;
}
index.open(index_name, true);
new_index.open(copy_name, false);
if (verbose)
printf("Copying values from '%s' to '%s'\n",
index_name.c_str(), copy_name.c_str());
RawDataReader reader(index);
if (single_name.empty())
{
bool ok = index.getFirstChannel(names);
while (ok)
{
copy_channel(names.getName(), start, end, index, reader,
new_index, channel_count, value_count, back_count);
ok = index.getNextChannel(names);
}
}
else
copy_channel(single_name, start, end, index, reader,
new_index, channel_count, value_count, back_count);
new_index.close();
index.close();
timer.stop();
if (verbose)
{
printf("Total: %lu channels, %lu values\n",
(unsigned long) channel_count, (unsigned long) value_count);
printf("Skipped %lu back-in-time values\n",
(unsigned long) back_count);
printf("Runtime: %s\n", timer.toString().c_str());
}
}
template<typename Quat> void bench(const std::string& label)
{
int tries = 10;
int rep = 1000000;
BenchTimer t;
Quat a(4, 1, 2, 3);
Quat b(2, 3, 4, 5);
Quat c;
std::cout.precision(3);
BENCH(t, tries, rep, quatmul_default(a,b,c));
std::cout << label << " default " << 1e3*t.best(CPU_TIMER) << "ms \t" << 1e-6*double(rep)/(t.best(CPU_TIMER)) << " M mul/s\n";
BENCH(t, tries, rep, quatmul_novec(a,b,c));
std::cout << label << " novec " << 1e3*t.best(CPU_TIMER) << "ms \t" << 1e-6*double(rep)/(t.best(CPU_TIMER)) << " M mul/s\n";
}
static void run()
{
arg1 a1;
a1.setIdentity();
arg2 a2;
a2.setIdentity();
BenchTimer timer;
timer.reset();
for (int k=0; k<10; ++k)
{
timer.start();
for (int k=0; k<REPEAT; ++k)
a2 = func::run( a1, a2 );
timer.stop();
}
cout << setprecision(4) << fixed << timer.value() << "s " << endl;;
}
EIGEN_DONT_INLINE void bench_prod()
{
typedef Matrix<Scalar,M,K> Lhs; Lhs a; a.setRandom();
typedef Matrix<Scalar,K,N> Rhs; Rhs b; b.setRandom();
typedef Matrix<Scalar,M,N> Res; Res c; c.setRandom();
BenchTimer t;
double n = 2.*double(M)*double(N)*double(K);
int rep = 100000./n;
rep /= 2;
if(rep<1) rep = 1;
do {
rep *= 2;
t.reset();
BENCH(t,1,rep,prod<CoeffBasedProductMode>(a,b,c));
} while(t.best()<0.1);
t.reset();
BENCH(t,5,rep,prod<Mode>(a,b,c));
print_mode(Mode);
std::cout << int(1e-6*n*rep/t.best()) << "\t";
}
void test(Test &t) {
build_list(&t.noobs, 100);
std::vector<std::thread> producers;
std::vector<std::thread> consumers;
// XXX: we should probably synchronize thread starting work and
// just time the actual work.
BenchTimer timer;
for (int i = 0; i < t.producers; i++) {
producers.push_back(std::thread(producer, &t, i));
}
for (int i = 0; i < t.consumers; i++) {
consumers.push_back(std::thread(consumer, &t, i));
}
joinAll(consumers);
timer.stop();
t.consumersDone = true;
joinAll(producers);
timer.report(t.count * t.consumers, !BenchMode);
}
void test(Test &t) {
std::vector<std::thread> producers;
std::vector<std::thread> consumers;
// XXX: we should probably synchronize thread starting work and
// just time the actual work.
BenchTimer timer;
for (int i = 0; i < t.producers; i++) {
producers.push_back(std::thread(producer, &t));
}
for (int i = 0; i < t.consumers; i++) {
consumers.push_back(std::thread(consumer, &t, i));
}
joinAll(consumers);
timer.stop();
t.consumersDone = true;
joinAll(producers);
//printf("Max thing: %ld\n", t.totalSum.load());
timer.report(t.count * t.consumers, !BenchMode);
}
void bench(int nfft,bool fwd,bool unscaled=false, bool halfspec=false)
{
typedef typename NumTraits<T>::Real Scalar;
typedef typename std::complex<Scalar> Complex;
int nits = NDATA/nfft;
vector<T> inbuf(nfft);
vector<Complex > outbuf(nfft);
FFT< Scalar > fft;
if (unscaled) {
fft.SetFlag(fft.Unscaled);
cout << "unscaled ";
}
if (halfspec) {
fft.SetFlag(fft.HalfSpectrum);
cout << "halfspec ";
}
std::fill(inbuf.begin(),inbuf.end(),0);
fft.fwd( outbuf , inbuf);
BenchTimer timer;
timer.reset();
for (int k=0;k<8;++k) {
timer.start();
if (fwd)
for(int i = 0; i < nits; i++)
fft.fwd( outbuf , inbuf);
else
for(int i = 0; i < nits; i++)
fft.inv(inbuf,outbuf);
timer.stop();
}
cout << nameof<Scalar>() << " ";
double mflops = 5.*nfft*log2((double)nfft) / (1e6 * timer.value() / (double)nits );
if ( NumTraits<T>::IsComplex ) {
cout << "complex";
}else{
cout << "real ";
mflops /= 2;
}
if (fwd)
cout << " fwd";
else
cout << " inv";
cout << " NFFT=" << nfft << " " << (double(1e-6*nfft*nits)/timer.value()) << " MS/s " << mflops << "MFLOPS\n";
}
int main(int argc, char *argv[])
{
// bench_sort();
int rows = SIZE;
int cols = SIZE;
float density = DENSITY;
EigenSparseMatrix sm1(rows,cols), sm2(rows,cols), sm3(rows,cols), sm4(rows,cols);
BenchTimer timer;
for (int nnzPerCol = NNZPERCOL; nnzPerCol>1; nnzPerCol/=1.1)
{
sm1.setZero();
sm2.setZero();
fillMatrix2(nnzPerCol, rows, cols, sm1);
fillMatrix2(nnzPerCol, rows, cols, sm2);
// std::cerr << "filling OK\n";
// dense matrices
#ifdef DENSEMATRIX
{
std::cout << "Eigen Dense\t" << nnzPerCol << "%\n";
DenseMatrix m1(rows,cols), m2(rows,cols), m3(rows,cols);
eiToDense(sm1, m1);
eiToDense(sm2, m2);
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
m3 = m1 * m2;
timer.stop();
std::cout << " a * b:\t" << timer.value() << endl;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
m3 = m1.transpose() * m2;
timer.stop();
std::cout << " a' * b:\t" << timer.value() << endl;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
m3 = m1.transpose() * m2.transpose();
timer.stop();
std::cout << " a' * b':\t" << timer.value() << endl;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
m3 = m1 * m2.transpose();
timer.stop();
std::cout << " a * b':\t" << timer.value() << endl;
}
#endif
// eigen sparse matrices
{
std::cout << "Eigen sparse\t" << sm1.nonZeros()/(float(sm1.rows())*float(sm1.cols()))*100 << "% * "
<< sm2.nonZeros()/(float(sm2.rows())*float(sm2.cols()))*100 << "%\n";
BENCH(sm3 = sm1 * sm2; )
std::cout << " a * b:\t" << timer.value() << endl;
// BENCH(sm3 = sm1.transpose() * sm2; )
// std::cout << " a' * b:\t" << timer.value() << endl;
// //
// BENCH(sm3 = sm1.transpose() * sm2.transpose(); )
// std::cout << " a' * b':\t" << timer.value() << endl;
// //
// BENCH(sm3 = sm1 * sm2.transpose(); )
// std::cout << " a * b' :\t" << timer.value() << endl;
// std::cout << "\n";
//
// BENCH( sm3._experimentalNewProduct(sm1, sm2); )
// std::cout << " a * b:\t" << timer.value() << endl;
//
// BENCH(sm3._experimentalNewProduct(sm1.transpose(),sm2); )
// std::cout << " a' * b:\t" << timer.value() << endl;
// //
// BENCH(sm3._experimentalNewProduct(sm1.transpose(),sm2.transpose()); )
// std::cout << " a' * b':\t" << timer.value() << endl;
// //
// BENCH(sm3._experimentalNewProduct(sm1, sm2.transpose());)
// std::cout << " a * b' :\t" << timer.value() << endl;
}
// eigen dyn-sparse matrices
/*{
DynamicSparseMatrix<Scalar> m1(sm1), m2(sm2), m3(sm3);
std::cout << "Eigen dyn-sparse\t" << m1.nonZeros()/(float(m1.rows())*float(m1.cols()))*100 << "% * "
<< m2.nonZeros()/(float(m2.rows())*float(m2.cols()))*100 << "%\n";
// timer.reset();
// timer.start();
BENCH(for (int k=0; k<REPEAT; ++k) m3 = m1 * m2;)
// timer.stop();
std::cout << " a * b:\t" << timer.value() << endl;
// std::cout << sm3 << "\n";
timer.reset();
timer.start();
// std::cerr << "transpose...\n";
// EigenSparseMatrix sm4 = sm1.transpose();
// std::cout << sm4.nonZeros() << " == " << sm1.nonZeros() << "\n";
// exit(1);
// std::cerr << "transpose OK\n";
// std::cout << sm1 << "\n\n" << sm1.transpose() << "\n\n" << sm4.transpose() << "\n\n";
BENCH(for (int k=0; k<REPEAT; ++k) m3 = m1.transpose() * m2;)
// timer.stop();
std::cout << " a' * b:\t" << timer.value() << endl;
// timer.reset();
// timer.start();
BENCH( for (int k=0; k<REPEAT; ++k) m3 = m1.transpose() * m2.transpose(); )
// timer.stop();
std::cout << " a' * b':\t" << timer.value() << endl;
// timer.reset();
// timer.start();
BENCH( for (int k=0; k<REPEAT; ++k) m3 = m1 * m2.transpose(); )
// timer.stop();
std::cout << " a * b' :\t" << timer.value() << endl;
}*/
// CSparse
#ifdef CSPARSE
{
std::cout << "CSparse \t" << nnzPerCol << "%\n";
cs *m1, *m2, *m3;
eiToCSparse(sm1, m1);
eiToCSparse(sm2, m2);
// timer.reset();
// timer.start();
// for (int k=0; k<REPEAT; ++k)
BENCH(
{
m3 = cs_sorted_multiply(m1, m2);
if (!m3)
{
std::cerr << "cs_multiply failed\n";
// break;
}
// cs_print(m3, 0);
cs_spfree(m3);
}
);
// timer.stop();
std::cout << " a * b:\t" << timer.value() << endl;
// BENCH( { m3 = cs_sorted_multiply2(m1, m2); cs_spfree(m3); } );
// std::cout << " a * b:\t" << timer.value() << endl;
}
int main(int argc, char *argv[])
{
int rows = SIZE;
int cols = SIZE;
float density = DENSITY;
EigenSparseMatrix sm1(rows,cols);
DenseVector v1(cols), v2(cols);
v1.setRandom();
BenchTimer timer;
for (float density = DENSITY; density>=MINDENSITY; density*=0.5)
{
fillMatrix(density, rows, cols, sm1);
// dense matrices
#ifdef DENSEMATRIX
{
std::cout << "Eigen Dense\t" << density*100 << "%\n";
DenseMatrix m1(rows,cols);
eiToDense(sm1, m1);
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
v2 = m1 * v1;
timer.stop();
std::cout << " a * v:\t" << timer.value() << endl;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
v2 = m1.transpose() * v1;
timer.stop();
std::cout << " a' * v:\t" << timer.value() << endl;
}
#endif
// eigen sparse matrices
{
std::cout << "Eigen sparse\t" << sm1.nonZeros()/float(sm1.rows()*sm1.cols())*100 << "%\n";
BENCH(for (int k=0; k<REPEAT; ++k) v2 = sm1 * v1;)
std::cout << " a * v:\t" << timer.value() << endl;
BENCH(for (int k=0; k<REPEAT; ++k) { asm("#mya"); v2 = sm1.transpose() * v1; asm("#myb"); })
std::cout << " a' * v:\t" << timer.value() << endl;
}
int tool_main(int argc, char** argv) {
SetupCrashHandler();
SkCommandLineFlags::Parse(argc, argv);
#if SK_ENABLE_INST_COUNT
if (FLAGS_leaks) {
gPrintInstCount = true;
}
#endif
SkAutoGraphics ag;
// First, parse some flags.
BenchLogger logger;
if (FLAGS_logFile.count()) {
logger.SetLogFile(FLAGS_logFile[0]);
}
LoggerResultsWriter logWriter(logger, FLAGS_timeFormat[0]);
MultiResultsWriter writer;
writer.add(&logWriter);
SkAutoTDelete<JSONResultsWriter> jsonWriter;
if (FLAGS_outResultsFile.count()) {
jsonWriter.reset(SkNEW(JSONResultsWriter(FLAGS_outResultsFile[0])));
writer.add(jsonWriter.get());
}
// Instantiate after all the writers have been added to writer so that we
// call close() before their destructors are called on the way out.
CallEnd<MultiResultsWriter> ender(writer);
const uint8_t alpha = FLAGS_forceBlend ? 0x80 : 0xFF;
SkTriState::State dither = SkTriState::kDefault;
for (size_t i = 0; i < 3; i++) {
if (strcmp(SkTriState::Name[i], FLAGS_forceDither[0]) == 0) {
dither = static_cast<SkTriState::State>(i);
}
}
BenchMode benchMode = kNormal_BenchMode;
for (size_t i = 0; i < SK_ARRAY_COUNT(BenchMode_Name); i++) {
if (strcmp(FLAGS_mode[0], BenchMode_Name[i]) == 0) {
benchMode = static_cast<BenchMode>(i);
}
}
SkTDArray<int> configs;
bool runDefaultConfigs = false;
// Try user-given configs first.
for (int i = 0; i < FLAGS_config.count(); i++) {
for (int j = 0; j < static_cast<int>(SK_ARRAY_COUNT(gConfigs)); ++j) {
if (0 == strcmp(FLAGS_config[i], gConfigs[j].name)) {
*configs.append() = j;
} else if (0 == strcmp(FLAGS_config[i], kDefaultsConfigStr)) {
runDefaultConfigs = true;
}
}
}
// If there weren't any, fill in with defaults.
if (runDefaultConfigs) {
for (int i = 0; i < static_cast<int>(SK_ARRAY_COUNT(gConfigs)); ++i) {
if (gConfigs[i].runByDefault) {
*configs.append() = i;
}
}
}
// Filter out things we can't run.
if (kNormal_BenchMode != benchMode) {
// Non-rendering configs only run in normal mode
for (int i = 0; i < configs.count(); ++i) {
const Config& config = gConfigs[configs[i]];
if (Benchmark::kNonRendering_Backend == config.backend) {
configs.remove(i, 1);
--i;
}
}
}
#if SK_SUPPORT_GPU
for (int i = 0; i < configs.count(); ++i) {
const Config& config = gConfigs[configs[i]];
if (Benchmark::kGPU_Backend == config.backend) {
GrContext* context = gContextFactory.get(config.contextType);
if (NULL == context) {
SkDebugf("GrContext could not be created for config %s. Config will be skipped.\n",
config.name);
configs.remove(i);
--i;
continue;
}
if (config.sampleCount > context->getMaxSampleCount()){
SkDebugf(
"Sample count (%d) for config %s is not supported. Config will be skipped.\n",
config.sampleCount, config.name);
configs.remove(i);
--i;
continue;
}
}
}
#endif
// All flags should be parsed now. Report our settings.
if (FLAGS_runOnce) {
logger.logError("bench was run with --runOnce, so we're going to hide the times."
" It's for your own good!\n");
}
writer.option("mode", FLAGS_mode[0]);
writer.option("alpha", SkStringPrintf("0x%02X", alpha).c_str());
writer.option("antialias", SkStringPrintf("%d", FLAGS_forceAA).c_str());
writer.option("filter", SkStringPrintf("%d", FLAGS_forceFilter).c_str());
writer.option("dither", SkTriState::Name[dither]);
writer.option("rotate", SkStringPrintf("%d", FLAGS_rotate).c_str());
writer.option("scale", SkStringPrintf("%d", FLAGS_scale).c_str());
writer.option("clip", SkStringPrintf("%d", FLAGS_clip).c_str());
#if defined(SK_BUILD_FOR_WIN32)
writer.option("system", "WIN32");
#elif defined(SK_BUILD_FOR_MAC)
writer.option("system", "MAC");
#elif defined(SK_BUILD_FOR_ANDROID)
writer.option("system", "ANDROID");
#elif defined(SK_BUILD_FOR_UNIX)
writer.option("system", "UNIX");
#else
writer.option("system", "other");
#endif
#if defined(SK_DEBUG)
writer.option("build", "DEBUG");
#else
writer.option("build", "RELEASE");
#endif
// Set texture cache limits if non-default.
for (size_t i = 0; i < SK_ARRAY_COUNT(gConfigs); ++i) {
#if SK_SUPPORT_GPU
const Config& config = gConfigs[i];
if (Benchmark::kGPU_Backend != config.backend) {
continue;
}
GrContext* context = gContextFactory.get(config.contextType);
if (NULL == context) {
continue;
}
size_t bytes;
int count;
context->getResourceCacheLimits(&count, &bytes);
if (-1 != FLAGS_gpuCacheBytes) {
bytes = static_cast<size_t>(FLAGS_gpuCacheBytes);
}
if (-1 != FLAGS_gpuCacheCount) {
count = FLAGS_gpuCacheCount;
}
context->setResourceCacheLimits(count, bytes);
#endif
}
// Run each bench in each configuration it supports and we asked for.
Iter iter;
Benchmark* bench;
while ((bench = iter.next()) != NULL) {
SkAutoTUnref<Benchmark> benchUnref(bench);
if (SkCommandLineFlags::ShouldSkip(FLAGS_match, bench->getName())) {
continue;
}
bench->setForceAlpha(alpha);
bench->setForceAA(FLAGS_forceAA);
bench->setForceFilter(FLAGS_forceFilter);
bench->setDither(dither);
bench->preDraw();
bool loggedBenchName = false;
for (int i = 0; i < configs.count(); ++i) {
const int configIndex = configs[i];
const Config& config = gConfigs[configIndex];
if (!bench->isSuitableFor(config.backend)) {
continue;
}
GrContext* context = NULL;
#if SK_SUPPORT_GPU
SkGLContextHelper* glContext = NULL;
if (Benchmark::kGPU_Backend == config.backend) {
context = gContextFactory.get(config.contextType);
if (NULL == context) {
continue;
}
glContext = gContextFactory.getGLContext(config.contextType);
}
#endif
SkAutoTUnref<SkCanvas> canvas;
SkAutoTUnref<SkPicture> recordFrom;
SkPictureRecorder recorderTo;
const SkIPoint dim = bench->getSize();
SkAutoTUnref<SkSurface> surface;
if (Benchmark::kNonRendering_Backend != config.backend) {
surface.reset(make_surface(config.fColorType,
dim,
config.backend,
config.sampleCount,
context));
if (!surface.get()) {
logger.logError(SkStringPrintf(
"Device creation failure for config %s. Will skip.\n", config.name));
continue;
}
switch(benchMode) {
case kDeferredSilent_BenchMode:
case kDeferred_BenchMode:
canvas.reset(SkDeferredCanvas::Create(surface.get()));
break;
case kRecord_BenchMode:
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
break;
case kPictureRecord_BenchMode: {
SkPictureRecorder recorderFrom;
bench->draw(1, recorderFrom.beginRecording(dim.fX, dim.fY));
recordFrom.reset(recorderFrom.endRecording());
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
break;
}
case kNormal_BenchMode:
canvas.reset(SkRef(surface->getCanvas()));
break;
default:
SkASSERT(false);
}
}
if (NULL != canvas) {
canvas->clear(SK_ColorWHITE);
if (FLAGS_clip) {
perform_clip(canvas, dim.fX, dim.fY);
}
if (FLAGS_scale) {
perform_scale(canvas, dim.fX, dim.fY);
}
if (FLAGS_rotate) {
perform_rotate(canvas, dim.fX, dim.fY);
}
}
if (!loggedBenchName) {
loggedBenchName = true;
writer.bench(bench->getName(), dim.fX, dim.fY);
}
#if SK_SUPPORT_GPU
SkGLContextHelper* contextHelper = NULL;
if (Benchmark::kGPU_Backend == config.backend) {
contextHelper = gContextFactory.getGLContext(config.contextType);
}
BenchTimer timer(contextHelper);
#else
BenchTimer timer;
#endif
double previous = std::numeric_limits<double>::infinity();
bool converged = false;
// variables used to compute loopsPerFrame
double frameIntervalTime = 0.0f;
int frameIntervalTotalLoops = 0;
bool frameIntervalComputed = false;
int loopsPerFrame = 0;
int loopsPerIter = 0;
if (FLAGS_verbose) { SkDebugf("%s %s: ", bench->getName(), config.name); }
if (!FLAGS_dryRun) {
do {
// Ramp up 1 -> 2 -> 4 -> 8 -> 16 -> ... -> ~1 billion.
loopsPerIter = (loopsPerIter == 0) ? 1 : loopsPerIter * 2;
if (loopsPerIter >= (1<<30) || timer.fWall > FLAGS_maxMs) {
// If you find it takes more than a billion loops to get up to 20ms of runtime,
// you've got a computer clocked at several THz or have a broken benchmark. ;)
// "1B ought to be enough for anybody."
logger.logError(SkStringPrintf(
"\nCan't get %s %s to converge in %dms (%d loops)",
bench->getName(), config.name, FLAGS_maxMs, loopsPerIter));
break;
}
if ((benchMode == kRecord_BenchMode || benchMode == kPictureRecord_BenchMode)) {
// Clear the recorded commands so that they do not accumulate.
canvas.reset(SkRef(recorderTo.beginRecording(dim.fX, dim.fY)));
}
timer.start();
// Inner loop that allows us to break the run into smaller
// chunks (e.g. frames). This is especially useful for the GPU
// as we can flush and/or swap buffers to keep the GPU from
// queuing up too much work.
for (int loopCount = loopsPerIter; loopCount > 0; ) {
// Save and restore around each call to draw() to guarantee a pristine canvas.
SkAutoCanvasRestore saveRestore(canvas, true/*also save*/);
int loops;
if (frameIntervalComputed && loopCount > loopsPerFrame) {
loops = loopsPerFrame;
loopCount -= loopsPerFrame;
} else {
loops = loopCount;
loopCount = 0;
}
if (benchMode == kPictureRecord_BenchMode) {
recordFrom->draw(canvas);
} else {
bench->draw(loops, canvas);
}
if (kDeferredSilent_BenchMode == benchMode) {
static_cast<SkDeferredCanvas*>(canvas.get())->silentFlush();
} else if (NULL != canvas) {
canvas->flush();
}
#if SK_SUPPORT_GPU
// swap drawing buffers on each frame to prevent the GPU
// from queuing up too much work
if (NULL != glContext) {
glContext->swapBuffers();
}
#endif
}
// Stop truncated timers before GL calls complete, and stop the full timers after.
timer.truncatedEnd();
#if SK_SUPPORT_GPU
if (NULL != glContext) {
context->flush();
SK_GL(*glContext, Finish());
}
#endif
timer.end();
// setup the frame interval for subsequent iterations
if (!frameIntervalComputed) {
frameIntervalTime += timer.fWall;
frameIntervalTotalLoops += loopsPerIter;
if (frameIntervalTime >= FLAGS_minMs) {
frameIntervalComputed = true;
loopsPerFrame =
(int)(((double)frameIntervalTotalLoops / frameIntervalTime) * FLAGS_minMs);
if (loopsPerFrame < 1) {
loopsPerFrame = 1;
}
// SkDebugf(" %s has %d loops in %f ms (normalized to %d)\n",
// bench->getName(), frameIntervalTotalLoops,
// timer.fWall, loopsPerFrame);
}
}
const double current = timer.fWall / loopsPerIter;
if (FLAGS_verbose && current > previous) { SkDebugf("↑"); }
if (FLAGS_verbose) { SkDebugf("%.3g ", current); }
converged = HasConverged(previous, current, timer.fWall);
previous = current;
} while (!FLAGS_runOnce && !converged);
}
if (FLAGS_verbose) { SkDebugf("\n"); }
if (!FLAGS_dryRun && FLAGS_outDir.count() && Benchmark::kNonRendering_Backend != config.backend) {
SkAutoTUnref<SkImage> image(surface->newImageSnapshot());
if (image.get()) {
saveFile(bench->getName(), config.name, FLAGS_outDir[0],
image);
}
}
if (FLAGS_runOnce) {
// Let's not mislead ourselves by looking at Debug build or single iteration bench times!
continue;
}
// Normalize to ms per 1000 iterations.
const double normalize = 1000.0 / loopsPerIter;
const struct { char shortName; const char* longName; double ms; } times[] = {
{'w', "msecs", normalize * timer.fWall},
{'W', "Wmsecs", normalize * timer.fTruncatedWall},
{'c', "cmsecs", normalize * timer.fCpu},
{'C', "Cmsecs", normalize * timer.fTruncatedCpu},
{'g', "gmsecs", normalize * timer.fGpu},
};
writer.config(config.name);
for (size_t i = 0; i < SK_ARRAY_COUNT(times); i++) {
if (strchr(FLAGS_timers[0], times[i].shortName) && times[i].ms > 0) {
writer.timer(times[i].longName, times[i].ms);
}
}
}
}
#if SK_SUPPORT_GPU
gContextFactory.destroyContexts();
#endif
return 0;
}
int main(int argc, char* argv[])
{
int size = SIZE * 8;
int size2 = size * size;
Scalar* a = internal::aligned_new<Scalar>(size2);
Scalar* b = internal::aligned_new<Scalar>(size2+4)+1;
Scalar* c = internal::aligned_new<Scalar>(size2);
for (int i=0; i<size; ++i)
{
a[i] = b[i] = c[i] = 0;
}
BenchTimer timer;
timer.reset();
for (int k=0; k<10; ++k)
{
timer.start();
benchVec(a, b, c, size2);
timer.stop();
}
std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
return 0;
for (int innersize = size; innersize>2 ; --innersize)
{
if (size2%innersize==0)
{
int outersize = size2/innersize;
MatrixXf ma = Map<MatrixXf>(a, innersize, outersize );
MatrixXf mb = Map<MatrixXf>(b, innersize, outersize );
MatrixXf mc = Map<MatrixXf>(c, innersize, outersize );
timer.reset();
for (int k=0; k<3; ++k)
{
timer.start();
benchVec(ma, mb, mc);
timer.stop();
}
std::cout << innersize << " x " << outersize << " " << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
}
}
VectorXf va = Map<VectorXf>(a, size2);
VectorXf vb = Map<VectorXf>(b, size2);
VectorXf vc = Map<VectorXf>(c, size2);
timer.reset();
for (int k=0; k<3; ++k)
{
timer.start();
benchVec(va, vb, vc);
timer.stop();
}
std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
return 0;
}
int main(int argc, char ** argv)
{
std::ptrdiff_t l1 = internal::queryL1CacheSize();
std::ptrdiff_t l2 = internal::queryTopLevelCacheSize();
std::cout << "L1 cache size = " << (l1>0 ? l1/1024 : -1) << " KB\n";
std::cout << "L2/L3 cache size = " << (l2>0 ? l2/1024 : -1) << " KB\n";
typedef internal::gebp_traits<Scalar,Scalar> Traits;
std::cout << "Register blocking = " << Traits::mr << " x " << Traits::nr << "\n";
int rep = 1; // number of repetitions per try
int tries = 2; // number of tries, we keep the best
int s = 2048;
int cache_size = -1;
bool need_help = false;
for (int i=1; i<argc; ++i)
{
if(argv[i][0]=='s')
s = atoi(argv[i]+1);
else if(argv[i][0]=='c')
cache_size = atoi(argv[i]+1);
else if(argv[i][0]=='t')
tries = atoi(argv[i]+1);
else if(argv[i][0]=='p')
rep = atoi(argv[i]+1);
else
need_help = true;
}
if(need_help)
{
std::cout << argv[0] << " s<matrix size> c<cache size> t<nb tries> p<nb repeats>\n";
return 1;
}
if(cache_size>0)
setCpuCacheSizes(cache_size,96*cache_size);
int m = s;
int n = s;
int p = s;
A a(m,p); a.setRandom();
B b(p,n); b.setRandom();
C c(m,n); c.setOnes();
C rc = c;
std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
std::ptrdiff_t mc(m), nc(n), kc(p);
internal::computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
std::cout << "blocking size (mc x kc) = " << mc << " x " << kc << "\n";
C r = c;
// check the parallel product is correct
#if defined EIGEN_HAS_OPENMP
int procs = omp_get_max_threads();
if(procs>1)
{
#ifdef HAVE_BLAS
blas_gemm(a,b,r);
#else
omp_set_num_threads(1);
r.noalias() += a * b;
omp_set_num_threads(procs);
#endif
c.noalias() += a * b;
if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
}
#elif defined HAVE_BLAS
blas_gemm(a,b,r);
c.noalias() += a * b;
if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
#else
gemm(a,b,c);
r.noalias() += a.cast<Scalar>() * b.cast<Scalar>();
if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n";
#endif
#ifdef HAVE_BLAS
BenchTimer tblas;
c = rc;
BENCH(tblas, tries, rep, blas_gemm(a,b,c));
std::cout << "blas cpu " << tblas.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(CPU_TIMER) << "s)\n";
std::cout << "blas real " << tblas.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(REAL_TIMER) << "s)\n";
#endif
BenchTimer tmt;
c = rc;
BENCH(tmt, tries, rep, gemm(a,b,c));
std::cout << "eigen cpu " << tmt.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(CPU_TIMER) << "s)\n";
std::cout << "eigen real " << tmt.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(REAL_TIMER) << "s)\n";
#ifdef EIGEN_HAS_OPENMP
if(procs>1)
{
BenchTimer tmono;
omp_set_num_threads(1);
Eigen::internal::setNbThreads(1);
c = rc;
BENCH(tmono, tries, rep, gemm(a,b,c));
std::cout << "eigen mono cpu " << tmono.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(CPU_TIMER) << "s)\n";
std::cout << "eigen mono real " << tmono.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(REAL_TIMER) << "s)\n";
std::cout << "mt speed up x" << tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER) << " => " << (100.0*tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER))/procs << "%\n";
}
#endif
#ifdef DECOUPLED
if((NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
{
M ar(m,p); ar.setRandom();
M ai(m,p); ai.setRandom();
M br(p,n); br.setRandom();
M bi(p,n); bi.setRandom();
M cr(m,n); cr.setRandom();
M ci(m,n); ci.setRandom();
BenchTimer t;
BENCH(t, tries, rep, matlab_cplx_cplx(ar,ai,br,bi,cr,ci));
std::cout << "\"matlab\" cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cout << "\"matlab\" real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
}
if((!NumTraits<A::Scalar>::IsComplex) && (NumTraits<B::Scalar>::IsComplex))
{
M a(m,p); a.setRandom();
M br(p,n); br.setRandom();
M bi(p,n); bi.setRandom();
M cr(m,n); cr.setRandom();
M ci(m,n); ci.setRandom();
BenchTimer t;
BENCH(t, tries, rep, matlab_real_cplx(a,br,bi,cr,ci));
std::cout << "\"matlab\" cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cout << "\"matlab\" real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
}
if((NumTraits<A::Scalar>::IsComplex) && (!NumTraits<B::Scalar>::IsComplex))
{
M ar(m,p); ar.setRandom();
M ai(m,p); ai.setRandom();
M b(p,n); b.setRandom();
M cr(m,n); cr.setRandom();
M ci(m,n); ci.setRandom();
BenchTimer t;
BENCH(t, tries, rep, matlab_cplx_real(ar,ai,b,cr,ci));
std::cout << "\"matlab\" cpu " << t.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << t.total(CPU_TIMER) << "s)\n";
std::cout << "\"matlab\" real " << t.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/t.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << t.total(REAL_TIMER) << "s)\n";
}
#endif
return 0;
}
