int sync_init() {
int i, arraysize;
nreceivers = conf.nreceivers;
corrlen = conf.sync_len;
if(nreceivers > NRECEIVERS_MAX) return -1;
/* half of fft input will be zero-padded */
fft1n = corrlen*2;
fft2n = fft1n * CORRELATION_OVERSAMPLE;
arraysize = nreceivers * fft1n;
fft1in = fftwf_malloc(arraysize * sizeof(*fft1in));
fft1out = fftwf_malloc(arraysize * sizeof(*fft1out));
for(i = 0; i < arraysize; i++)
fft1in[i] = fft1out[i] = 0;
arraysize = (nreceivers-1) * fft2n;
fft2in = fftwf_malloc(arraysize * sizeof(*fft2in));
fft2out = fftwf_malloc(arraysize * sizeof(*fft2out));
for(i = 0; i < arraysize; i++)
fft2in[i] = fft2out[i] = 0;
fft1plan = fftwf_plan_many_dft(
1, &fft1n, nreceivers,
fft1in, NULL, 1, fft1n,
fft1out, NULL, 1, fft1n,
FFTW_FORWARD, FFTW_ESTIMATE);
fft2plan = fftwf_plan_many_dft(
1, &fft2n, nreceivers-1,
fft2in, NULL, 1, fft2n,
fft2out, NULL, 1, fft2n,
FFTW_BACKWARD, FFTW_ESTIMATE);
return 0;
}
Fourier::Fourier(int n)
{
const char* fname = "void Fourier::Initialize()";
VRB.Debug(fname, "Allocating memory and creating plans for FFTW.");
batch_size = n;
#ifdef USE_SINGLE
b = (fftComplex*) fftwf_malloc(batch_size*GJP.Vol()*sizeof(fftComplex));
#endif
#ifdef USE_DOUBLE
b = (fftComplex*) fftw_malloc(batch_size*GJP.Vol()*sizeof(fftComplex));
#endif
#ifdef USE_LONG_DOUBLE
b = (fftComplex*) fftwl_malloc(batch_size*GJP.Vol()*sizeof(fftComplex));
#endif
// Below needs to be adjusted for batch ffts; double check in place
int vol = GJP.Vol();
int dims[3] = { GJP.Xsites(), GJP.Ysites(), GJP.Zsites() };
#ifdef USE_SINGLE
p1 = fftwf_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_BACKWARD, FFTW_EXHAUSTIVE);
p2 = fftwf_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_FORWARD, FFTW_EXHAUSTIVE);
#endif
#ifdef USE_DOUBLE
p1 = fftw_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_BACKWARD, FFTW_EXHAUSTIVE);
p2 = fftw_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_FORWARD, FFTW_EXHAUSTIVE);
#endif
#ifdef USE_LONG_DOUBLE
p1 = fftwl_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_BACKWARD, FFTW_EXHAUSTIVE);
p2 = fftwl_plan_many_dft(3, dims, batch_size, b, NULL, 1, vol, b, NULL, 1, vol, FFTW_FORWARD, FFTW_EXHAUSTIVE);
#endif
}
void Qsquaref(fftwf_complex const *FQ,fftwf_complex *FQ2,int nx,int nz,int k){
fftwf_complex *FQ1;
FQ1 = (fftwf_complex *) fftwf_malloc((2*k-1)*nx*nz*sizeof(fftwf_complex));
int m,n,l;
// printf(" k = %d \n",k);
for (l=0;l<k;l++)
{//printf("l= %d\n",l);
for (m=0;m<nx;m++)
{
for(n=0;n<nz;n++)
{//printf("%d ",l+(k)*(m*nz+n));
FQ1[l+(2*k-1)*(m*nz+n)] = FQ[l+k*(m*nz+n)];
}
}
}
// printf("padded array set in Qsquare \n");
int rank =1;
int *dims;
dims = (int*) malloc(rank*sizeof(int));
dims[0]=(2*k-1); //dims[1]=nz; dims[2]=2*k-1;
int howmany = nx*nz;
//inembed=onembed=dims
int istride = 1;
int ostride = 1;
int idist = (2*k-1);
int odist = idist;
// printf("variables set ...\n");
fftwf_plan plan_forward,plan_backward;
plan_forward = fftwf_plan_many_dft(rank,dims,howmany,FQ1,NULL,istride,idist,FQ1,NULL,ostride,odist,FFTW_FORWARD,FFTW_ESTIMATE);
plan_backward = fftwf_plan_many_dft(rank,dims,howmany,FQ1,NULL,istride,idist,FQ2,NULL,ostride,odist,FFTW_BACKWARD,FFTW_ESTIMATE);
fftwf_execute(plan_forward);
//printf("plan executed \n");
for (l=0;l<2*k-1;l++)
{//printf("l= %d\n",l);
for (m=0;m<nx;m++)
{
for(n=0;n<nz;n++)
{//printf("%d ",l+(k)*(m*nz+n));
FQ1[l+(2*k-1)*(m*nz+n)] = FQ1[l+(2*k-1)*(m*nz+n)]*FQ1[l+(2*k-1)*(m*nz+n)]/(2*k-1); //printf("set \n");
}
}
}
fftwf_execute(plan_backward);
//cleaning
fftwf_destroy_plan(plan_forward);
fftwf_destroy_plan(plan_backward);
fftwf_free(FQ1);
}
int main (int argc, char **argv)
{
int n1, nx, n3, dim, n[SF_MAX_DIM]; /* dimensions */
int i1, ix, i3, j; /* loop counters */
int nk; /* number of wavenumbers */
int npad; /* padding */
float dx; /* space sampling interval */
float dk; /* wavenumber sampling interval */
float x0; /* staring space */
float k0; /* starting wavenumber */
float wt; /* Fourier scaling */
kiss_fft_cpx **cp; /* frequency-wavenumber */
bool inv; /* forward or inverse */
bool sym; /* symmetric scaling */
bool opt; /* optimal padding */
int sign; /* transform sign */
int axis; /* transform axis */
char varname[12]; /* variable name */
char *label; /* transformed axis label */
#ifdef SF_HAS_FFTW
fftwf_plan cfg;
#else
kiss_fft_cpx *ctrace;
kiss_fft_cfg cfg;
#endif
sf_file in=NULL, out=NULL;
sf_init(argc,argv);
in = sf_input ( "in");
out = sf_output("out");
if (SF_COMPLEX != sf_gettype(in)) sf_error ("Need complex input");
if (!sf_getbool("inv",&inv)) inv = false;
/* if y, perform inverse transform */
if (!sf_getbool("sym",&sym)) sym=false;
/* if y, apply symmetric scaling to make the FFT operator Hermitian */
if (!sf_getint("sign",&sign)) sign = inv? 1: 0;
/* transform sign (0 or 1) */
if (!sf_getbool("opt",&opt)) opt=true;
/* if y, determine optimal size for efficiency */
if (!sf_getint("axis",&axis)) axis=2;
/* Axis to transform */
dim = sf_filedims(in,n);
n1=n3=1;
for (j=0; j < dim; j++) {
if (j < axis-1) n1 *= n[j];
else if (j > axis-1) n3 *= n[j];
}
if (inv) {
sprintf(varname,"n%d",axis);
if (!sf_histint (in,varname,&nk)) sf_error("No %s= in input",varname);
sprintf(varname,"d%d",axis);
if (!sf_histfloat(in,varname,&dk)) sf_error("No %s= in input",varname);
sprintf(varname,"fft3_n%d",axis);
if (!sf_histint (in,varname,&nx)) nx=nk;
sprintf(varname,"fft3_o%d",axis);
if (!sf_histfloat(in,varname,&x0)) x0 = 0.;
sprintf(varname,"fft3_label%d",axis);
label = sf_histstring(in,varname);
dx = 1./(nk*dk);
sprintf(varname,"n%d",axis);
sf_putint (out,varname,nx);
sprintf(varname,"d%d",axis);
sf_putfloat (out,varname,dx);
sprintf(varname,"o%d",axis);
sf_putfloat (out,varname,x0);
sprintf(varname,"label%d",axis);
if (NULL != label) {
sf_putstring(out,varname,label);
} else if (NULL != (label = sf_histstring(in,varname))) {
(void) sf_fft_label(axis,label,out);
}
} else {
sprintf(varname,"n%d",axis);
if (!sf_histint (in,varname,&nx)) sf_error("No %s= in input",varname);
sprintf(varname,"d%d",axis);
if (!sf_histfloat(in,varname,&dx)) sf_error("No %s= in input",varname);
sprintf(varname,"o%d",axis);
if (!sf_histfloat(in,varname,&x0)) x0 = 0.;
sprintf(varname,"label%d",axis);
label = sf_histstring(in,varname);
sprintf(varname,"fft3_n%d",axis);
sf_putint(out,varname,nx);
sprintf(varname,"fft3_o%d",axis);
sf_putfloat(out,varname,x0);
if (NULL != label) {
sprintf(varname,"fft3_label%d",axis);
sf_putstring(out,varname,label);
}
if (!sf_getint("pad",&npad)) npad=2;
/* padding factor */
/* determine wavenumber sampling */
nk = opt? kiss_fft_next_fast_size(nx*npad): nx*npad;
if (nk != nx) sf_warning("padded to %d",nk);
dk = 1./(nk*dx);
k0 = -0.5/dx;
sprintf(varname,"n%d",axis);
sf_putint (out,varname,nk);
sprintf(varname,"d%d",axis);
sf_putfloat (out,varname,dk);
sprintf(varname,"o%d",axis);
sf_putfloat (out,varname,k0);
if (NULL != label && !sf_fft_label(axis,label,out)) {
sprintf(varname,"label%d",axis);
sf_putstring(out,varname,"Wavenumber");
}
}
sprintf(varname,"unit%d",axis);
sf_fft_unit(axis,sf_histstring(in,varname),out);
cp = (kiss_fft_cpx**) sf_complexalloc2(n1,nk);
#ifdef SF_HAS_FFTW
ix = nk;
cfg = fftwf_plan_many_dft(1, &ix, n1,
(fftwf_complex*) cp[0], NULL, n1, 1,
(fftwf_complex*) cp[0], NULL, n1, 1,
sign? FFTW_BACKWARD: FFTW_FORWARD,
FFTW_ESTIMATE);
if (NULL == cfg) sf_error("FFTW failure.");
#else
ctrace = (kiss_fft_cpx*) sf_complexalloc(nk);
cfg = kiss_fft_alloc(nk,sign,NULL,NULL);
#endif
/* FFT scaling */
wt = sym? 1./sqrtf((float) nk): 1./nk;
for (i3=0; i3<n3; i3++) {
if (inv) {
sf_floatread((float*) cp[0],n1*nk*2,in);
#ifdef SF_HAS_FFTW
fftwf_execute(cfg);
for (ix=0; ix<nx; ix++) {
for (i1=0; i1 < n1; i1++) {
cp[ix][i1] = sf_crmul(cp[ix][i1],ix%2? -wt: wt);
}
}
#else
for (i1=0; i1 < n1; i1++) {
/* Fourier transform k to x */
kiss_fft_stride(cfg,cp[0]+i1,ctrace,n1);
for (ix=0; ix<nx; ix++) {
cp[ix][i1] = sf_crmul(ctrace[ix],ix%2? -wt: wt);
}
}
#endif
sf_floatwrite((float*) cp[0],n1*nx*2,out);
} else {
sf_floatread((float*) cp[0],n1*nx*2,in);
/* FFT centering */
for (ix=1; ix<nx; ix+=2) {
for (i1=0; i1<n1; i1++) {
cp[ix][i1] = sf_cneg(cp[ix][i1]);
}
}
if (sym) {
for (ix=0; ix<nx; ix++) {
for (i1=0; i1 < n1; i1++) {
cp[ix][i1] = sf_crmul(cp[ix][i1],wt);
}
}
}
/* pad with zeros */
for (ix=nx; ix<nk; ix++) {
for (i1=0; i1<n1; i1++) {
cp[ix][i1].r = 0.;
cp[ix][i1].i = 0.;
}
}
#ifdef SF_HAS_FFTW
fftwf_execute(cfg);
#else
for (i1=0; i1 < n1; i1++) {
/* Fourier transform x to k */
kiss_fft_stride(cfg,cp[0]+i1,ctrace,n1);
/* Transpose */
for (ix=0; ix<nk; ix++) {
cp[ix][i1] = ctrace[ix];
}
}
#endif
sf_floatwrite((float*) cp[0],n1*nk*2,out);
}
}
exit (0);
}
int main(int argc, char* argv[])
{
bool verb, pow2;
char key[7], *mode;;
int n1, n2, n1padded, n2padded, num, dim, n[SF_MAX_DIM], npadded[SF_MAX_DIM], ii[SF_MAX_DIM];
int i, j, i1, i2, index, nw, iter, niter, nthr, *pad;
float thr, pclip, normp;
float *dobs_t, *thresh, *mask;
fftwf_complex *mm, *dd, *dobs;
fftwf_plan fft1, ifft1, fftrem, ifftrem;/* execute plan for FFT and IFFT */
sf_file in, out, Fmask; /* mask and I/O files*/
sf_init(argc,argv); /* Madagascar initialization */
in=sf_input("in"); /* read the data to be interpolated */
out=sf_output("out"); /* output the reconstructed data */
Fmask=sf_input("mask"); /* read the (n-1)-D mask for n-D data */
if(!sf_getbool("verb",&verb)) verb=false;
/* verbosity */
if(!sf_getbool("pow2",&pow2)) pow2=false;
/* round up the length of each axis to be power of 2 */
if (!sf_getint("niter",&niter)) niter=100;
/* total number of iterations */
if (!sf_getfloat("pclip",&pclip)) pclip=10.;
/* starting data clip percentile (default is 10)*/
if ( !(mode=sf_getstring("mode")) ) mode = "exp";
/* thresholding mode: 'hard', 'soft','pthresh','exp';
'hard', hard thresholding; 'soft', soft thresholding;
'pthresh', generalized quasi-p; 'exp', exponential shrinkage */
if (pclip <=0. || pclip > 100.) sf_error("pclip=%g should be > 0 and <= 100",pclip);
if (!sf_getfloat("normp",&normp)) normp=1.;
/* quasi-norm: normp<2 */
for (i=0; i < SF_MAX_DIM; i++) {/* dimensions */
snprintf(key,3,"n%d",i+1);
if (!sf_getint(key,n+i) &&
(NULL == in || !sf_histint(in,key,n+i))) break;
/*( n# size of #-th axis )*/
sf_putint(out,key,n[i]);
}
if (0==i) sf_error("Need n1=");
dim=i;
pad=sf_intalloc (dim);
for (i=0; i<dim; i++) pad[i]=0;
sf_getints("pad",pad,dim); /* number of zeros to be padded for each axis */
n1=n[0];
n2=sf_leftsize(in,1);
for (i=0; i<SF_MAX_DIM; i++) npadded[i]=1;
npadded[0]=n1+pad[0];
n1padded=npadded[0];
n2padded=1;
for (i=1; i<dim; i++){
npadded[i]=n[i]+pad[i];
if (pow2) {/* zero-padding to be power of 2 */
npadded[i]=nextpower2(n[i]);
fprintf(stderr,"n%d=%d n%dpadded=%d\n",i,n[i],i,npadded[i]);
}
n2padded*=npadded[i];
}
nw=npadded[0]/2+1;
num=nw*n2padded;/* data: total number of elements in frequency domain */
/* allocate data and mask arrays */
thresh=(float*) malloc(nw*n2padded*sizeof(float));
dobs_t=(float*) fftwf_malloc(n1padded*n2padded*sizeof(float)); /* time domain observation */
dobs=(fftwf_complex*)fftwf_malloc(nw*n2padded*sizeof(fftwf_complex));/* freq-domain observation */
dd=(fftwf_complex*) fftwf_malloc(nw*n2padded*sizeof(fftwf_complex));
mm=(fftwf_complex*) fftwf_malloc(nw*n2padded*sizeof(fftwf_complex));
if (NULL != sf_getstring("mask")){
mask=sf_floatalloc(n2padded);
} else sf_error("mask needed!");
/* initialize the input data and mask arrays */
memset(dobs_t,0,n1padded*n2padded*sizeof(float));
memset(mask,0,n2padded*sizeof(float));
for (i=0; i<n1*n2; i+=n1){
sf_line2cart(dim,n,i,ii);
j=sf_cart2line(dim,npadded,ii);
sf_floatread(&dobs_t[j], n1, in);
sf_floatread(&mask[j/n1padded], 1, Fmask);
}
/* FFT for the 1st dimension and the remaining dimensions */
fft1=fftwf_plan_many_dft_r2c(1, &n1padded, n2padded, dobs_t, &n1padded, 1, n1padded, dobs, &n1padded, 1, nw, FFTW_MEASURE);
ifft1=fftwf_plan_many_dft_c2r(1, &n1padded, n2padded, dobs, &n1padded, 1, nw, dobs_t, &n1padded, 1, n1padded, FFTW_MEASURE);
fftrem=fftwf_plan_many_dft(dim-1, &npadded[1], nw, dd, &npadded[1], nw, 1, dd, &npadded[1], nw, 1, FFTW_FORWARD, FFTW_MEASURE);
ifftrem=fftwf_plan_many_dft(dim-1, &npadded[1], nw, dd, &npadded[1], nw, 1, dd, &npadded[1], nw, 1, FFTW_BACKWARD, FFTW_MEASURE);
/* transform the data from time domain to frequency domain: dobs_t-->dobs */
fftwf_execute(fft1);
for(i=0; i<num; i++) dobs[i]/=sqrtf(n1padded);
memset(mm,0,num*sizeof(fftwf_complex));
/* Iterative Shrinkage-Thresholding (IST) Algorithm:
mm^{k+1}=T[mm^k+A^* M^* (dobs-M A mm^k)] (M^*=M; Mdobs=dobs)
=T[mm^k+A^*(dobs-M A mm^k)]; (k=0,1,...niter-1)
dd^=A mm^;
*/
for(iter=0; iter<niter; iter++)
{
/* dd<-- A mm^k */
memcpy(dd, mm, num*sizeof(fftwf_complex));
fftwf_execute(ifftrem);
for(i=0; i<num; i++) dd[i]/=sqrtf(n2padded);
/* apply mask: dd<--dobs-M A mm^k=dobs-M dd */
for(i2=0; i2<n2padded; i2++)
for(i1=0; i1<nw; i1++)
{
index=i1+nw*i2;
dd[index]=dobs[index]-mask[i2]*dd[index];
}
/* mm^k += A^*(dobs-M A mm^k); dd=dobs-M A mm^k */
fftwf_execute(fftrem);
for(i=0; i<num; i++) mm[i]+=dd[i]/sqrtf(n2padded);
/* perform thresholding */
for(i=0; i<num; i++) thresh[i]=cabsf(mm[i]);
nthr = 0.5+num*(1.-0.01*pclip); /* round off */
if (nthr < 0) nthr=0;
if (nthr >= num) nthr=num-1;
thr=sf_quantile(nthr, num, thresh);
sf_cpthresh(mm, num, thr, normp, mode);
if (verb) sf_warning("iteration %d;",iter+1);
}
/* frequency--> time domain: dobs-->dobs_t */
memcpy(dd, mm, num*sizeof(fftwf_complex));
fftwf_execute(ifftrem);
for(i=0; i<num; i++) dd[i]/=sqrtf(n2padded);
memcpy(dobs, dd, num*sizeof(fftwf_complex));
fftwf_execute(ifft1);
for(i=0; i<n1padded*n2padded; i++) dobs_t[i]/=sqrtf(n1padded);
for (i=0; i<n1*n2; i+=n1){
sf_line2cart(dim,n,i,ii);
j=sf_cart2line(dim,npadded,ii);
sf_floatwrite(&dobs_t[j],n1,out);
}
free(thresh);
fftwf_free(dobs_t);
fftwf_free(dobs);
fftwf_free(dd);
fftwf_free(mm);
exit(0);
}
void FFT::apply(Window im, bool transformX, bool transformY, bool transformT, bool inverse) {
assert(im.channels % 2 == 0, "-fft requires an image with an even number of channels\n");
if (im.width == 1) transformX = false;
if (im.height == 1) transformY = false;
if (im.frames == 1) transformT = false;
// rank 0
if (!transformX && !transformY && !transformT) return;
if (transformX && transformY && transformT) { // rank 3
int n[] = {im.frames, im.height, im.width};
int nembed[] = {im.frames, im.tstride/im.ystride, im.ystride/im.xstride};
fftwf_plan plan = fftwf_plan_many_dft(3, n, im.channels/2,
(fftwf_complex *)im(0, 0, 0), nembed, im.channels/2, 1,
(fftwf_complex *)im(0, 0, 0), nembed, im.channels/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
} else if (transformX && transformY) { // rank 2
int n[] = {im.height, im.width};
int nembed[] = {im.tstride/im.ystride, im.ystride/im.xstride};
for (int t = 0; t < im.frames; t++) {
fftwf_plan plan = fftwf_plan_many_dft(2, n, im.channels/2,
(fftwf_complex *)im(0, 0, t), nembed, im.channels/2, 1,
(fftwf_complex *)im(0, 0, t), nembed, im.channels/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
} else if (transformT && transformY) { // rank 2
int n[] = {im.frames, im.height};
int nembed[] = {im.frames, im.tstride/im.ystride};
fftwf_plan plan = fftwf_plan_many_dft(2, n, im.width*im.channels/2,
(fftwf_complex *)im(0, 0, 0), nembed, im.width*im.channels/2, 1,
(fftwf_complex *)im(0, 0, 0), nembed, im.width*im.channels/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
} else if (transformT && transformX) { // rank 2
int n[] = {im.frames, im.width};
int nembed[] = {im.frames, im.tstride/im.xstride};
for (int y = 0; y < im.height; y++) {
fftwf_plan plan = fftwf_plan_many_dft(2, n, im.channels/2,
(fftwf_complex *)im(0, y, 0), nembed, im.channels/2, 1,
(fftwf_complex *)im(0, y, 0), nembed, im.channels/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
} else if (transformX) { // rank 1
int n[] = {im.width};
int nembed[] = {im.width};
for (int t = 0; t < im.frames; t++) {
for (int y = 0; y < im.height; y++) {
fftwf_plan plan = fftwf_plan_many_dft(1, n, im.channels/2,
(fftwf_complex *)im(0, y, t), nembed, im.channels/2, 1,
(fftwf_complex *)im(0, y, t), nembed, im.channels/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
}
} else if (transformY) { // rank 1
int n[] = {im.height};
int nembed[] = {im.height};
for (int t = 0; t < im.frames; t++) {
fftwf_plan plan = fftwf_plan_many_dft(1, n, im.width*im.channels/2,
(fftwf_complex *)im(0, 0, t), nembed, im.ystride/2, 1,
(fftwf_complex *)im(0, 0, t), nembed, im.ystride/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
} else if (transformT) { // rank 1
int n[] = {im.frames};
int nembed[] = {im.frames};
for (int y = 0; y < im.height; y++) {
fftwf_plan plan = fftwf_plan_many_dft(1, n, im.width*im.channels/2,
(fftwf_complex *)im(0, y, 0), nembed, im.tstride/2, 1,
(fftwf_complex *)im(0, y, 0), nembed, im.tstride/2, 1,
inverse ? FFTW_BACKWARD : FFTW_FORWARD, FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
}
if (inverse) {
float m = 1.0;
if (transformX) m *= im.width;
if (transformY) m *= im.height;
if (transformT) m *= im.frames;
Scale::apply(im, 1.0f/m);
}
}
Array<T> fftconvolve(Array<T> const& signal, Array<T> const& filter,
const bool expand, ConvolveBatchKind kind)
{
const af::dim4 sd = signal.dims();
const af::dim4 fd = filter.dims();
dim_t fftScale = 1;
af::dim4 packed_dims;
int fft_dims[baseDim];
af::dim4 sig_tmp_dims, sig_tmp_strides;
af::dim4 filter_tmp_dims, filter_tmp_strides;
// Pack both signal and filter on same memory array, this will ensure
// better use of batched cuFFT capabilities
for (dim_t k = 0; k < 4; k++) {
if (k < baseDim)
packed_dims[k] = nextpow2((unsigned)(sd[k] + fd[k] - 1));
else if (k == baseDim)
packed_dims[k] = sd[k] + fd[k];
else
packed_dims[k] = 1;
if (k < baseDim) {
fft_dims[baseDim-k-1] = (k == 0) ? packed_dims[k] / 2 : packed_dims[k];
fftScale *= fft_dims[baseDim-k-1];
}
}
Array<convT> packed = createEmptyArray<convT>(packed_dims);
convT *packed_ptr = packed.get();
const af::dim4 packed_strides = packed.strides();
sig_tmp_dims[0] = filter_tmp_dims[0] = packed_dims[0];
sig_tmp_strides[0] = filter_tmp_strides[0] = 1;
for (dim_t k = 1; k < 4; k++) {
if (k < baseDim) {
sig_tmp_dims[k] = packed_dims[k];
filter_tmp_dims[k] = packed_dims[k];
}
else {
sig_tmp_dims[k] = sd[k];
filter_tmp_dims[k] = fd[k];
}
sig_tmp_strides[k] = sig_tmp_strides[k - 1] * sig_tmp_dims[k - 1];
filter_tmp_strides[k] = filter_tmp_strides[k - 1] * filter_tmp_dims[k - 1];
}
// Calculate memory offsets for packed signal and filter
convT *sig_tmp_ptr = packed_ptr;
convT *filter_tmp_ptr = packed_ptr + sig_tmp_strides[3] * sig_tmp_dims[3];
// Number of packed complex elements in dimension 0
dim_t sig_half_d0 = divup(sd[0], 2);
// Pack signal in a complex matrix where first dimension is half the input
// (allows faster FFT computation) and pad array to a power of 2 with 0s
packData<convT, T>(sig_tmp_ptr, sig_tmp_dims, sig_tmp_strides, signal);
// Pad filter array with 0s
padArray<convT, T>(filter_tmp_ptr, filter_tmp_dims, filter_tmp_strides, filter);
// Compute forward FFT
if (isDouble) {
fftw_plan plan = fftw_plan_many_dft(baseDim,
fft_dims,
packed_dims[baseDim],
(fftw_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
(fftw_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
FFTW_FORWARD,
FFTW_ESTIMATE);
fftw_execute(plan);
fftw_destroy_plan(plan);
}
else {
fftwf_plan plan = fftwf_plan_many_dft(baseDim,
fft_dims,
packed_dims[baseDim],
(fftwf_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
(fftwf_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
FFTW_FORWARD,
FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
// Multiply filter and signal FFT arrays
if (kind == ONE2MANY)
complexMultiply<convT>(filter_tmp_ptr, filter_tmp_dims, filter_tmp_strides,
sig_tmp_ptr, sig_tmp_dims, sig_tmp_strides,
filter_tmp_ptr, filter_tmp_dims, filter_tmp_strides,
kind);
else
complexMultiply<convT>(sig_tmp_ptr, sig_tmp_dims, sig_tmp_strides,
sig_tmp_ptr, sig_tmp_dims, sig_tmp_strides,
filter_tmp_ptr, filter_tmp_dims, filter_tmp_strides,
kind);
// Compute inverse FFT
if (isDouble) {
fftw_plan plan = fftw_plan_many_dft(baseDim,
fft_dims,
packed_dims[baseDim],
(fftw_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
(fftw_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
FFTW_BACKWARD,
FFTW_ESTIMATE);
fftw_execute(plan);
fftw_destroy_plan(plan);
}
else {
fftwf_plan plan = fftwf_plan_many_dft(baseDim,
fft_dims,
packed_dims[baseDim],
(fftwf_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
(fftwf_complex*)packed.get(),
NULL,
packed_strides[0],
packed_strides[baseDim] / 2,
FFTW_BACKWARD,
FFTW_ESTIMATE);
fftwf_execute(plan);
fftwf_destroy_plan(plan);
}
// Compute output dimensions
dim4 oDims(1);
if (expand) {
for(dim_t d=0; d<4; ++d) {
if (kind==ONE2ONE || kind==ONE2MANY) {
oDims[d] = sd[d]+fd[d]-1;
} else {
oDims[d] = (d<baseDim ? sd[d]+fd[d]-1 : sd[d]);
}
}
} else {
oDims = sd;
if (kind==ONE2MANY) {
for (dim_t i=baseDim; i<4; ++i)
oDims[i] = fd[i];
}
}
Array<T> out = createEmptyArray<T>(oDims);
T* out_ptr = out.get();
const af::dim4 out_dims = out.dims();
const af::dim4 out_strides = out.strides();
const af::dim4 filter_dims = filter.dims();
// Reorder the output
if (kind == ONE2MANY) {
reorderOutput<T, convT, roundOut>
(out_ptr, out_dims, out_strides,
filter_tmp_ptr, filter_tmp_dims, filter_tmp_strides,
filter_dims, sig_half_d0, baseDim, fftScale, expand);
}
else {
reorderOutput<T, convT, roundOut>
(out_ptr, out_dims, out_strides,
sig_tmp_ptr, sig_tmp_dims, sig_tmp_strides,
filter_dims, sig_half_d0, baseDim, fftScale, expand);
}
return out;
}
/* Set up fft plans. Need to have npol, nphase, nchan
* already filled in struct */
int cyclic_init_ffts(struct cyclic_work *w) {
/* Infer lag, harmonic sizes from chan/phase */
w->nlag = w->nchan; // Total number of lags including + and -
w->nharm = w->nphase/2 + 1; // Only DC and positive harmonics
/* Alloc temp arrays for planning */
PS ps;
CS cs;
CC cc;
PC pc;
struct filter_time ft;
struct filter_freq ff;
struct profile_phase pp;
struct profile_harm ph;
ps.npol = cs.npol = cc.npol = pc.npol = w->npol;
ps.nphase = pc.nphase = w->nphase;
ps.nchan = cs.nchan = w->nchan;
cs.nharm = cc.nharm = w->nharm;
cc.nlag = pc.nlag = w->nlag;
cyclic_alloc_ps(&ps);
cyclic_alloc_cs(&cs);
cyclic_alloc_cc(&cc);
cyclic_alloc_pc(&pc);
ft.nlag = w->nlag;
ff.nchan = w->nchan;
pp.nphase = w->nphase;
ph.nharm = w->nharm;
filter_alloc_time(&ft);
filter_alloc_freq(&ff);
profile_alloc_phase(&pp);
profile_alloc_harm(&ph);
/* FFT plans */
int rv=0;
/* ps2cs - r2c fft along phase (fastest) axis */
w->ps2cs = fftwf_plan_many_dft_r2c(1, &w->nphase, w->npol*w->nchan,
ps.data, NULL, 1, w->nphase,
cs.data, NULL, 1, w->nharm,
FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->ps2cs == NULL) rv++;
/* cs2cc - c2c ifft along channel axis */
w->cs2cc = fftwf_plan_many_dft(1, &w->nchan, w->npol*w->nharm,
cs.data, NULL, w->nharm*w->npol, 1,
cc.data, NULL, w->nharm*w->npol, 1,
FFTW_BACKWARD, FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->cs2cc == NULL) rv++;
/* cc2cs - c2c fft along lag axis */
w->cc2cs = fftwf_plan_many_dft(1, &w->nlag, w->npol*w->nharm,
cc.data, NULL, w->nharm*w->npol, 1,
cs.data, NULL, w->nharm*w->npol, 1,
FFTW_FORWARD, FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->cc2cs == NULL) rv++;
/* time2freq, freq2time for filters */
w->time2freq = fftwf_plan_dft_1d(w->nlag, ft.data, ff.data,
FFTW_FORWARD, FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->time2freq == NULL) rv++;
w->freq2time = fftwf_plan_dft_1d(w->nchan, ff.data, ft.data,
FFTW_BACKWARD, FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->freq2time == NULL) rv++;
/* phase2harm, harm2phase for profiles */
w->phase2harm = fftwf_plan_dft_r2c_1d(w->nphase, pp.data, ph.data,
FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->phase2harm == NULL) rv++;
w->harm2phase = fftwf_plan_dft_c2r_1d(w->nphase, ph.data, pp.data,
FFTW_MEASURE | FFTW_PRESERVE_INPUT);
if (w->harm2phase == NULL) rv++;
cyclic_free_ps(&ps);
cyclic_free_cs(&cs);
cyclic_free_cc(&cc);
cyclic_free_pc(&pc);
filter_free_time(&ft);
filter_free_freq(&ff);
profile_free_phase(&pp);
profile_free_harm(&ph);
return(rv);
}
