void run_bench(const char *s, int (*fn)()) {
printf("benching %20s ..", s); fflush(stdout);
double t0 = uclock_sec(), t1, tmax = 0.3;
double niter = 0;
do {
niter += fn();
t1 = uclock_sec();
} while (t1 - t0 < tmax);
#define REF_FREQ_MHZ 1000.0
printf(" -> %6.1f millions of vector evaluations/second -> %3.0f cycles/value on a %gMHz computer\n", floor(niter/(t1-t0)/1e5)/10, (t1-t0)*REF_FREQ_MHZ*1e6/niter/4, REF_FREQ_MHZ);
}
void benchmark_ffts(int N, int cplx) {
int Nfloat = (cplx ? N*2 : N);
int Nbytes = Nfloat * sizeof(float);
float *X = pffft_aligned_malloc(Nbytes), *Y = pffft_aligned_malloc(Nbytes), *Z = pffft_aligned_malloc(Nbytes);
double t0, t1, flops;
int k;
int max_iter = 5120000/N*4;
#ifdef __arm__
max_iter /= 4;
#endif
int iter;
for (k = 0; k < Nfloat; ++k) {
X[k] = 0; //sqrtf(k+1);
}
// FFTPack benchmark
{
float *wrk = malloc(2*Nbytes + 15*sizeof(float));
int max_iter_ = max_iter/pffft_simd_size(); if (max_iter_ == 0) max_iter_ = 1;
if (cplx) cffti(N, wrk);
else rffti(N, wrk);
t0 = uclock_sec();
for (iter = 0; iter < max_iter_; ++iter) {
if (cplx) {
cfftf(N, X, wrk);
cfftb(N, X, wrk);
} else {
rfftf(N, X, wrk);
rfftb(N, X, wrk);
}
}
t1 = uclock_sec();
free(wrk);
flops = (max_iter_*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); // see http://www.fftw.org/speed/method.html
show_output("FFTPack", N, cplx, flops, t0, t1, max_iter_);
}
#ifdef HAVE_VECLIB
int log2N = (int)(log(N)/log(2) + 0.5f);
if (N == (1<<log2N)) {
FFTSetup setup;
setup = vDSP_create_fftsetup(log2N, FFT_RADIX2);
DSPSplitComplex zsamples;
zsamples.realp = &X[0];
zsamples.imagp = &X[Nfloat/2];
t0 = uclock_sec();
for (iter = 0; iter < max_iter; ++iter) {
if (cplx) {
vDSP_fft_zip(setup, &zsamples, 1, log2N, kFFTDirection_Forward);
vDSP_fft_zip(setup, &zsamples, 1, log2N, kFFTDirection_Inverse);
} else {
vDSP_fft_zrip(setup, &zsamples, 1, log2N, kFFTDirection_Forward);
vDSP_fft_zrip(setup, &zsamples, 1, log2N, kFFTDirection_Inverse);
}
}
t1 = uclock_sec();
vDSP_destroy_fftsetup(setup);
flops = (max_iter*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); // see http://www.fftw.org/speed/method.html
show_output("vDSP", N, cplx, flops, t0, t1, max_iter);
} else {
show_output("vDSP", N, cplx, -1, -1, -1, -1);
}
#endif
#ifdef HAVE_FFTW
{
fftwf_plan planf, planb;
fftw_complex *in = (fftw_complex*) fftwf_malloc(sizeof(fftw_complex) * N);
fftw_complex *out = (fftw_complex*) fftwf_malloc(sizeof(fftw_complex) * N);
memset(in, 0, sizeof(fftw_complex) * N);
int flags = (N < 40000 ? FFTW_MEASURE : FFTW_ESTIMATE); // measure takes a lot of time on largest ffts
//int flags = FFTW_ESTIMATE;
if (cplx) {
planf = fftwf_plan_dft_1d(N, (fftwf_complex*)in, (fftwf_complex*)out, FFTW_FORWARD, flags);
planb = fftwf_plan_dft_1d(N, (fftwf_complex*)in, (fftwf_complex*)out, FFTW_BACKWARD, flags);
} else {
planf = fftwf_plan_dft_r2c_1d(N, (float*)in, (fftwf_complex*)out, flags);
planb = fftwf_plan_dft_c2r_1d(N, (fftwf_complex*)in, (float*)out, flags);
}
t0 = uclock_sec();
for (iter = 0; iter < max_iter; ++iter) {
fftwf_execute(planf);
fftwf_execute(planb);
}
t1 = uclock_sec();
fftwf_destroy_plan(planf);
fftwf_destroy_plan(planb);
fftwf_free(in); fftwf_free(out);
flops = (max_iter*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); // see http://www.fftw.org/speed/method.html
show_output((flags == FFTW_MEASURE ? "FFTW (meas.)" : " FFTW (estim)"), N, cplx, flops, t0, t1, max_iter);
}
#endif
// PFFFT benchmark
{
PFFFT_Setup *s = pffft_new_setup(N, cplx ? PFFFT_COMPLEX : PFFFT_REAL);
if (s) {
t0 = uclock_sec();
for (iter = 0; iter < max_iter; ++iter) {
pffft_transform(s, X, Z, Y, PFFFT_FORWARD);
pffft_transform(s, X, Z, Y, PFFFT_BACKWARD);
}
t1 = uclock_sec();
pffft_destroy_setup(s);
flops = (max_iter*2) * ((cplx ? 5 : 2.5)*N*log((double)N)/M_LN2); // see http://www.fftw.org/speed/method.html
show_output("PFFFT", N, cplx, flops, t0, t1, max_iter);
}
}
if (!array_output_format) {
printf("--\n");
}
pffft_aligned_free(X);
pffft_aligned_free(Y);
pffft_aligned_free(Z);
}
