FFTHelper::~FFTHelper()
{
vDSP_destroy_fftsetup(mSpectrumAnalysis);
free (mDspSplitComplex.realp);
free (mDspSplitComplex.imagp);
}
DspRfft::~DspRfft() {
#if __APPLE__
vDSP_destroy_fftsetup(fftSetup);
#endif // __APPLE__
}
AccelerateFFT<float>::~AccelerateFFT()
{
free (complexSplit.realp);
free (complexSplit.imagp);
vDSP_destroy_fftsetup (fftSetupFloat);
}
void computeReferenceF(clFFT_SplitComplex *out, clFFT_Dim3 n,
unsigned int batchSize, clFFT_Dimension dim, clFFT_Direction dir)
{
FFTSetup plan_vdsp;
DSPSplitComplex out_vdsp;
FFTDirection dir_vdsp = dir == clFFT_Forward ? FFT_FORWARD : FFT_INVERSE;
unsigned int i, j, k;
unsigned int stride;
unsigned int log2Nx = (unsigned int) log2(n.x);
unsigned int log2Ny = (unsigned int) log2(n.y);
unsigned int log2Nz = (unsigned int) log2(n.z);
unsigned int log2N;
log2N = log2Nx;
log2N = log2N > log2Ny ? log2N : log2Ny;
log2N = log2N > log2Nz ? log2N : log2Nz;
plan_vdsp = vDSP_create_fftsetup(log2N, 2);
switch(dim)
{
case clFFT_1D:
for(i = 0; i < batchSize; i++)
{
stride = i * n.x;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, 1, log2Nx, dir_vdsp);
}
break;
case clFFT_2D:
for(i = 0; i < batchSize; i++)
{
for(j = 0; j < n.y; j++)
{
stride = j * n.x + i * n.x * n.y;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, 1, log2Nx, dir_vdsp);
}
}
for(i = 0; i < batchSize; i++)
{
for(j = 0; j < n.x; j++)
{
stride = j + i * n.x * n.y;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, n.x, log2Ny, dir_vdsp);
}
}
break;
case clFFT_3D:
for(i = 0; i < batchSize; i++)
{
for(j = 0; j < n.z; j++)
{
for(k = 0; k < n.y; k++)
{
stride = k * n.x + j * n.x * n.y + i * n.x * n.y * n.z;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, 1, log2Nx, dir_vdsp);
}
}
}
for(i = 0; i < batchSize; i++)
{
for(j = 0; j < n.z; j++)
{
for(k = 0; k < n.x; k++)
{
stride = k + j * n.x * n.y + i * n.x * n.y * n.z;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, n.x, log2Ny, dir_vdsp);
}
}
}
for(i = 0; i < batchSize; i++)
{
for(j = 0; j < n.y; j++)
{
for(k = 0; k < n.x; k++)
{
stride = k + j * n.x + i * n.x * n.y * n.z;
out_vdsp.realp = out->real + stride;
out_vdsp.imagp = out->imag + stride;
vDSP_fft_zip(plan_vdsp, &out_vdsp, n.x*n.y, log2Nz, dir_vdsp);
}
}
}
break;
}
vDSP_destroy_fftsetup(plan_vdsp);
}
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);
}
int main(int argc, char *argv[])
{
// Initialize the pseudo-random number generator.
InitializeRandom();
// Initialize FFT data.
FFTSetup Setup = vDSP_create_fftsetup(Log2SampleLength, FFT_RADIX2);
if (Setup == 0)
{
fprintf(stderr, "Error, unable to create FFT setup.\n");
exit(EXIT_FAILURE);
}
/* If there are no command-line arguments, prompt for keys
interactively.
*/
if (argc <= 1)
{
// Process keys for the user until they are done.
int c;
while (1)
{
// Prompt the user.
printf(
"Please enter a key (one of 0-9, *, #, or A-D): ");
// Skip whitespace except newlines.
do
c = getchar();
while (c != '\n' && isspace(c));
// When there is a blank line or an EOF, quit.
if (c == '\n' || c == EOF)
break;
// Look up the key in the table.
FrequencyPair F = ConvertKeyToFrequencies(toupper(c));
// If it is a valid key, demonstrate the FFT.
if (F.Frequency[0] != 0)
Demonstrate(Setup, F);
// Skip anything else on the line.
do
c = getchar();
while (c != EOF && c != '\n');
// When the input ends, quit. Otherwise, do more.
if (c == EOF)
break;
}
// Do not leave the cursor in the middle of a line.
if (c != '\n')
printf("\n");
}
// If there is one command line argument, process the keys in it.
else if (argc == 2)
{
for (char *p = argv[1]; *p; ++p)
{
// Look up the key in the table.
FrequencyPair F = ConvertKeyToFrequencies(toupper(*p));
// If it is a valid key, demonstrate the FFT.
if (F.Frequency[0] != 0)
{
printf("Simulating key %c.\n", *p);
Demonstrate(Setup, F);
}
else
fprintf(stderr,
"Error, key %c not recognized.\n", *p);
}
}
// If there are too many arguments, print a usage message.
else
{
fprintf(stderr,
"Usage: %s [telephone keys 0-9, #, *, or A-D]\n",
argv[0]);
exit(EXIT_FAILURE);
}
// Release resources.
vDSP_destroy_fftsetup(Setup);
return 0;
}
