static void * setup(int len) {return pffft_new_setup(len, PFFFT_REAL);}
void TimeOnePfFFT(int count, int fft_log_size, float signal_value,
int signal_type) {
struct AlignedPtr* x_aligned;
struct AlignedPtr* y_aligned;
struct AlignedPtr* z_aligned;
struct AlignedPtr* y_true_aligned;
struct ComplexFloat* x;
struct ComplexFloat* y;
OMX_FC32* z;
struct ComplexFloat* y_true;
int n;
int fft_size;
struct timeval start_time;
struct timeval end_time;
double elapsed_time;
PFFFT_Setup *s;
struct SnrResult snr_forward;
struct SnrResult snr_inverse;
fft_size = 1 << fft_log_size;
x_aligned = AllocAlignedPointer(32, sizeof(*x) * fft_size);
y_aligned = AllocAlignedPointer(32, sizeof(*y) * (fft_size + 2));
z_aligned = AllocAlignedPointer(32, sizeof(*z) * fft_size);
y_true_aligned = AllocAlignedPointer(32, sizeof(*y_true) * fft_size);
x = x_aligned->aligned_pointer_;
y = y_aligned->aligned_pointer_;
z = z_aligned->aligned_pointer_;
y_true = y_true_aligned->aligned_pointer_;
s = pffft_new_setup(fft_size, PFFFT_COMPLEX);
if (!s) {
fprintf(stderr, "TimeOnePfFFT: Could not initialize structure for order %d\n",
fft_log_size);
}
GenerateTestSignalAndFFT(x, y_true, fft_size, signal_type, signal_value, 0);
if (do_forward_test) {
GetUserTime(&start_time);
for (n = 0; n < count; ++n) {
pffft_transform_ordered(s, (float*)x, (float*)y, NULL, PFFFT_FORWARD);
}
GetUserTime(&end_time);
elapsed_time = TimeDifference(&start_time, &end_time);
CompareComplexFloat(&snr_forward, (OMX_FC32*) y, (OMX_FC32*) y_true, fft_size);
PrintResult("Forward PFFFT FFT", fft_log_size, elapsed_time, count, snr_forward.complex_snr_);
if (verbose >= 255) {
printf("FFT Actual:\n");
DumpArrayComplexFloat("y", fft_size, (OMX_FC32*) y);
printf("FFT Expected:\n");
DumpArrayComplexFloat("true", fft_size, (OMX_FC32*) y_true);
}
}
if (do_inverse_test) {
float scale = 1.0 / fft_size;
memcpy(y, y_true, sizeof(*y) * (fft_size + 2));
GetUserTime(&start_time);
for (n = 0; n < count; ++n) {
int m;
pffft_transform_ordered(s, (float*)y_true, (float*)z, NULL, PFFFT_BACKWARD);
/*
* Need to include cost of scaling the inverse
*/
ScaleVector((OMX_F32*) z, 2 * fft_size, fft_size);
}
GetUserTime(&end_time);
elapsed_time = TimeDifference(&start_time, &end_time);
CompareComplexFloat(&snr_inverse, (OMX_FC32*) z, (OMX_FC32*) x, fft_size);
PrintResult("Inverse PFFFT FFT", fft_log_size, elapsed_time, count, snr_inverse.complex_snr_);
if (verbose >= 255) {
printf("IFFT Actual:\n");
DumpArrayComplexFloat("z", fft_size, z);
printf("IFFT Expected:\n");
DumpArrayComplexFloat("x", fft_size, (OMX_FC32*) x);
}
}
FreeAlignedPointer(x_aligned);
FreeAlignedPointer(y_aligned);
FreeAlignedPointer(z_aligned);
FreeAlignedPointer(y_true_aligned);
pffft_destroy_setup(s);
}
/* compare results with the regular fftpack */
void pffft_validate_N(int N, int cplx) {
int Nfloat = N*(cplx?2:1);
int Nbytes = Nfloat * sizeof(float);
float *ref, *in, *out, *tmp, *tmp2;
PFFFT_Setup *s = pffft_new_setup(N, cplx ? PFFFT_COMPLEX : PFFFT_REAL);
int pass;
if (!s) { printf("Skipping N=%d, not supported\n", N); return; }
ref = pffft_aligned_malloc(Nbytes);
in = pffft_aligned_malloc(Nbytes);
out = pffft_aligned_malloc(Nbytes);
tmp = pffft_aligned_malloc(Nbytes);
tmp2 = pffft_aligned_malloc(Nbytes);
for (pass=0; pass < 2; ++pass) {
float ref_max = 0;
int k;
//printf("N=%d pass=%d cplx=%d\n", N, pass, cplx);
// compute reference solution with FFTPACK
if (pass == 0) {
float *wrk = malloc(2*Nbytes+15*sizeof(float));
for (k=0; k < Nfloat; ++k) {
ref[k] = in[k] = frand()*2-1;
out[k] = 1e30;
}
if (!cplx) {
rffti(N, wrk);
rfftf(N, ref, wrk);
// use our ordering for real ffts instead of the one of fftpack
{
float refN=ref[N-1];
for (k=N-2; k >= 1; --k) ref[k+1] = ref[k];
ref[1] = refN;
}
} else {
cffti(N, wrk);
cfftf(N, ref, wrk);
}
free(wrk);
}
for (k = 0; k < Nfloat; ++k) ref_max = MAX(ref_max, fabs(ref[k]));
// pass 0 : non canonical ordering of transform coefficients
if (pass == 0) {
// test forward transform, with different input / output
pffft_transform(s, in, tmp, 0, PFFFT_FORWARD);
memcpy(tmp2, tmp, Nbytes);
memcpy(tmp, in, Nbytes);
pffft_transform(s, tmp, tmp, 0, PFFFT_FORWARD);
for (k = 0; k < Nfloat; ++k) {
assert(tmp2[k] == tmp[k]);
}
// test reordering
pffft_zreorder(s, tmp, out, PFFFT_FORWARD);
pffft_zreorder(s, out, tmp, PFFFT_BACKWARD);
for (k = 0; k < Nfloat; ++k) {
assert(tmp2[k] == tmp[k]);
}
pffft_zreorder(s, tmp, out, PFFFT_FORWARD);
} else {
// pass 1 : canonical ordering of transform coeffs.
pffft_transform_ordered(s, in, tmp, 0, PFFFT_FORWARD);
memcpy(tmp2, tmp, Nbytes);
memcpy(tmp, in, Nbytes);
pffft_transform_ordered(s, tmp, tmp, 0, PFFFT_FORWARD);
for (k = 0; k < Nfloat; ++k) {
assert(tmp2[k] == tmp[k]);
}
memcpy(out, tmp, Nbytes);
}
{
for (k=0; k < Nfloat; ++k) {
if (!(fabs(ref[k] - out[k]) < 1e-3*ref_max)) {
printf("%s forward PFFFT mismatch found for N=%d\n", (cplx?"CPLX":"REAL"), N);
exit(1);
}
}
if (pass == 0) pffft_transform(s, tmp, out, 0, PFFFT_BACKWARD);
else pffft_transform_ordered(s, tmp, out, 0, PFFFT_BACKWARD);
memcpy(tmp2, out, Nbytes);
memcpy(out, tmp, Nbytes);
if (pass == 0) pffft_transform(s, out, out, 0, PFFFT_BACKWARD);
else pffft_transform_ordered(s, out, out, 0, PFFFT_BACKWARD);
for (k = 0; k < Nfloat; ++k) {
assert(tmp2[k] == out[k]);
out[k] *= 1.f/N;
}
for (k = 0; k < Nfloat; ++k) {
if (fabs(in[k] - out[k]) > 1e-3 * ref_max) {
printf("pass=%d, %s IFFFT does not match for N=%d\n", pass, (cplx?"CPLX":"REAL"), N); break;
exit(1);
}
}
}
// quick test of the circular convolution in fft domain
{
float conv_err = 0, conv_max = 0;
pffft_zreorder(s, ref, tmp, PFFFT_FORWARD);
memset(out, 0, Nbytes);
pffft_zconvolve_accumulate(s, ref, ref, out, 1.0);
pffft_zreorder(s, out, tmp2, PFFFT_FORWARD);
for (k=0; k < Nfloat; k += 2) {
float ar = tmp[k], ai=tmp[k+1];
if (cplx || k > 0) {
tmp[k] = ar*ar - ai*ai;
tmp[k+1] = 2*ar*ai;
} else {
tmp[0] = ar*ar;
tmp[1] = ai*ai;
}
}
for (k=0; k < Nfloat; ++k) {
float d = fabs(tmp[k] - tmp2[k]), e = fabs(tmp[k]);
if (d > conv_err) conv_err = d;
if (e > conv_max) conv_max = e;
}
if (conv_err > 1e-5*conv_max) {
printf("zconvolve error ? %g %g\n", conv_err, conv_max); exit(1);
}
}
}
printf("%s PFFFT is OK for N=%d\n", (cplx?"CPLX":"REAL"), N); fflush(stdout);
pffft_destroy_setup(s);
pffft_aligned_free(ref);
pffft_aligned_free(in);
pffft_aligned_free(out);
pffft_aligned_free(tmp);
pffft_aligned_free(tmp2);
}
void TimeOnePfRFFT(int count, int fft_log_size, float signal_value,
int signal_type) {
struct AlignedPtr* x_aligned;
struct AlignedPtr* y_aligned;
struct AlignedPtr* z_aligned;
struct AlignedPtr* y_tmp_aligned;
float* x;
struct ComplexFloat* y;
OMX_F32* z;
float* y_true;
float* y_tmp;
int n;
int fft_size;
struct timeval start_time;
struct timeval end_time;
double elapsed_time;
PFFFT_Setup *s;
struct SnrResult snr_forward;
struct SnrResult snr_inverse;
fft_size = 1 << fft_log_size;
x_aligned = AllocAlignedPointer(32, sizeof(*x) * fft_size);
y_aligned = AllocAlignedPointer(32, sizeof(*y) * (fft_size + 2));
z_aligned = AllocAlignedPointer(32, sizeof(*z) * fft_size);
y_tmp_aligned = AllocAlignedPointer(32, sizeof(*y_tmp) * (fft_size + 2));
y_true = (float*) malloc(sizeof(*y_true) * 2 * fft_size);
x = x_aligned->aligned_pointer_;
y = y_aligned->aligned_pointer_;
z = z_aligned->aligned_pointer_;
y_tmp = y_tmp_aligned->aligned_pointer_;
s = pffft_new_setup(fft_size, PFFFT_REAL);
if (!s) {
fprintf(stderr, "TimeOnePfRFFT: Could not initialize structure for order %d\n",
fft_log_size);
}
GenerateRealFloatSignal(x, (struct ComplexFloat*) y_true, fft_size, signal_type, signal_value);
if (do_forward_test) {
GetUserTime(&start_time);
for (n = 0; n < count; ++n) {
pffft_transform_ordered(s, (float*)x, (float*)y, NULL, PFFFT_FORWARD);
}
GetUserTime(&end_time);
elapsed_time = TimeDifference(&start_time, &end_time);
/*
* Arrange the output of the FFT to match the expected output.
*/
y[fft_size / 2].Re = y[0].Im;
y[fft_size / 2].Im = 0;
y[0].Im = 0;
CompareComplexFloat(&snr_forward, (OMX_FC32*) y, (OMX_FC32*) y_true, fft_size / 2 + 1);
PrintResult("Forward PFFFT RFFT", fft_log_size, elapsed_time, count, snr_forward.complex_snr_);
if (verbose >= 255) {
printf("FFT Actual:\n");
DumpArrayComplexFloat("y", fft_size / 2 + 1, (OMX_FC32*) y);
printf("FFT Expected:\n");
DumpArrayComplexFloat("true", fft_size / 2 + 1, (OMX_FC32*) y_true);
}
}
if (do_inverse_test) {
float scale = 1.0 / fft_size;
/* Copy y_true to true, but arrange the values according to what rdft wants. */
memcpy(y_tmp, y_true, sizeof(y_tmp[0]) * fft_size);
y_tmp[1] = y_true[fft_size / 2];
GetUserTime(&start_time);
for (n = 0; n < count; ++n) {
int m;
pffft_transform_ordered(s, (float*)y_tmp, (float*)z, NULL, PFFFT_BACKWARD);
/*
* Need to include cost of scaling the inverse
*/
ScaleVector(z, fft_size, fft_size);
}
GetUserTime(&end_time);
elapsed_time = TimeDifference(&start_time, &end_time);
CompareFloat(&snr_inverse, (OMX_F32*) z, (OMX_F32*) x, fft_size);
PrintResult("Inverse PFFFT RFFT", fft_log_size, elapsed_time, count, snr_inverse.complex_snr_);
if (verbose >= 255) {
printf("IFFT Actual:\n");
DumpArrayFloat("z", fft_size, z);
printf("IFFT Expected:\n");
DumpArrayFloat("x", fft_size, x);
}
}
FreeAlignedPointer(x_aligned);
FreeAlignedPointer(y_aligned);
FreeAlignedPointer(z_aligned);
FreeAlignedPointer(y_tmp_aligned);
pffft_destroy_setup(s);
free(y_true);
}
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);
}