void bspfft1d_init(int n1, int N, int s, int t, double *w0, double *w, double *tw,
int *rho_np, int *rho_p){
/* This parallel function initializes all the tables used in the FFT. */
int nlc, k1, ntw, c;
double alpha;
nlc= nloc(N,t,n1);
bitrev_init(nlc,rho_np);
bitrev_init(N,rho_p);
k1= k1_init(n1,N,nlc);
ufft_init(k1,w0);
ufft_init(nlc,w);
ntw= 0;
for (c=k1; c<=N; c *=nlc){
alpha= (t%c) / (double)(c);
twiddle_init(nlc,alpha,rho_np,&tw[2*ntw*nlc]);
ntw++;
}
} /* end bspfft_init */
int dfft_create_plan_common(dfft_plan *p,
int ndim, int *gdim,
int *inembed, int *oembed,
int *pdim, int *pidx, int row_m,
int input_cyclic, int output_cyclic,
MPI_Comm comm,
int *proc_map,
int device)
{
int nump;
p->comm = comm;
MPI_Comm_size(comm,&nump);
/* number of processor must be power of two */
if (nump & (nump-1)) return 4;
/* Allocate memory for processor map and copy over */
p->proc_map = malloc(sizeof(int)*nump);
memcpy(p->proc_map, proc_map, sizeof(int)*nump);
p->pdim = malloc(ndim*sizeof(int));
p->gdim = malloc(ndim*sizeof(int));
p->pidx = malloc(ndim*sizeof(int));
p->inembed = malloc(ndim*sizeof(int));
p->oembed = malloc(ndim*sizeof(int));
p->ndim = ndim;
int i;
for (i = 0; i < ndim; i++)
{
p->gdim[i] = gdim[i];
/* Every dimension must be a power of two */
if (gdim[i] & (gdim[i]-1)) return 5;
p->pdim[i] = pdim[i];
}
if (inembed != NULL)
{
for (i = 0; i < ndim; i++)
p->inembed[i] = inembed[i];
}
else
{
for (i = 0; i < ndim; i++)
p->inembed[i] = p->gdim[i]/p->pdim[i];
}
if (oembed != NULL)
{
for (i = 0; i < ndim; i++)
p->oembed[i] = oembed[i];
}
else
{
for (i = 0; i < ndim; i++)
p->oembed[i] = p->gdim[i]/p->pdim[i];
}
p->offset_send = (int *)malloc(sizeof(int)*nump);
p->offset_recv = (int *)malloc(sizeof(int)*nump);
p->nsend = (int *)malloc(sizeof(int)*nump);
p->nrecv = (int *)malloc(sizeof(int)*nump);
if (!device)
{
#ifdef ENABLE_HOST
p->plans_short_forward = malloc(sizeof(plan_t)*ndim);
p->plans_long_forward = malloc(sizeof(plan_t)*ndim);
p->plans_short_inverse = malloc(sizeof(plan_t)*ndim);
p->plans_long_inverse = malloc(sizeof(plan_t)*ndim);
#else
return 3;
#endif
}
/* local problem size */
int size_in = 1;
int size_out = 1;
/* since we expect column-major input, the leading dimension
has no embedding */
p->inembed[0] = gdim[0]/pdim[0];
p->oembed[0] = gdim[0]/pdim[0];
for (i = 0; i < ndim ; ++i)
{
size_in *= p->inembed[i];
size_out *= p->oembed[i];
}
p->size_in = size_in;
p->size_out = size_out;
/* find length k0 of last stage of butterflies */
p->k0 = malloc(sizeof(int)*ndim);
for (i = 0; i< ndim; ++i)
{
int length = gdim[i]/pdim[i];
if (length > 1)
{
int c;
for (c=gdim[i]; c>length; c /= length)
;
p->k0[i] = c;
}
else
{
p->k0[i] = 1;
}
}
p->rho_L = (int **)malloc(ndim*sizeof(int *));
p->rho_pk0= (int **)malloc(ndim*sizeof(int *));
p->rho_Lk0 = (int **)malloc(ndim*sizeof(int *));
for (i = 0; i < ndim; ++i)
{
int length = gdim[i]/pdim[i];
p->rho_L[i] = (int *) malloc(sizeof(int)*length);
p->rho_pk0[i] = (int *) malloc(sizeof(int)*pdim[i]/(p->k0[i]));
p->rho_Lk0[i] = (int *) malloc(sizeof(int)*length/(p->k0[i]));
bitrev_init(length, p->rho_L[i]);
bitrev_init(pdim[i]/(p->k0[i]),p->rho_pk0[i]);
bitrev_init(length/(p->k0[i]),p->rho_Lk0[i]);
}
/* processor coordinates */
for (i = 0; i < ndim; ++i)
{
p->pidx[i] = pidx[i];
}
/* init local FFT library */
int res;
if (!device)
{
#ifdef ENABLE_HOST
res = dfft_init_local_fft();
#else
return 3;
#endif
}
else
{
#ifdef ENABLE_CUDA
res = dfft_cuda_init_local_fft();
#else
return 2;
#endif
}
if (res) return 1;
int size = size_in;
p->dfft_multi = 0;
p->device = device;
if (device)
{
/* use multidimensional local transforms */
dfft_create_execution_flow(p);
/* allocate storage for variables */
int dmax = p->max_depth + 2;
p->rev_j1 = (int **) malloc(sizeof(int *)*dmax);
p->rev_global = (int **) malloc(sizeof(int *)*dmax);
p->rev_partial = (int **) malloc(sizeof(int *)*dmax);
p->c0 = (int **) malloc(sizeof(int *)*dmax);
p->c1 = (int **) malloc(sizeof(int *)*dmax);
int d;
for (d = 0; d < dmax; ++d)
{
p->rev_j1[d] = (int *) malloc(sizeof(int)*ndim);
p->rev_global[d] = (int *) malloc(sizeof(int)*ndim);
p->rev_partial[d] = (int *) malloc(sizeof(int)*ndim);
p->c0[d] = (int *) malloc(sizeof(int)*ndim);
p->c1[d] = (int *) malloc(sizeof(int)*ndim);
}
p->dfft_multi = 1;
}
else
{
for (i = 0; i < ndim; ++i)
{
/* plan for short-distance butterflies */
int st = size/p->inembed[i]*(gdim[i]/pdim[i]);
#ifdef ENABLE_HOST
int howmany = 1;
#ifdef FFT1D_SUPPORTS_THREADS
howmany = st/(p->k0[i]);
#endif
dfft_create_1d_plan(&(p->plans_short_forward[i]),p->k0[i],
howmany, st/(p->k0[i]), 1, st/(p->k0[i]), 1, 0);
dfft_create_1d_plan(&(p->plans_short_inverse[i]),p->k0[i],
howmany, st/(p->k0[i]), 1, st/(p->k0[i]), 1, 1);
/* plan for long-distance butterflies */
int length = gdim[i]/pdim[i];
#ifdef FFT1D_SUPPORTS_THREADS
howmany = st/length;
#endif
dfft_create_1d_plan(&(p->plans_long_forward[i]), length,
howmany, st/length,1, st/length,1, 0);
dfft_create_1d_plan(&(p->plans_long_inverse[i]), length,
howmany, st/length,1, st/length,1, 1);
#else
return 3;
#endif
size /= p->inembed[i];
size *= p->oembed[i];
}
}
/* Allocate scratch space */
int scratch_size = 1;
for (i = 0; i < ndim; ++i)
scratch_size *= ((p->inembed[i] > p->oembed[i]) ? p->inembed[i] : p->oembed[i]);
p->scratch_size = scratch_size;
if (!device)
{
#ifdef ENABLE_HOST
dfft_allocate_aligned_memory(&(p->scratch),sizeof(cpx_t)*scratch_size);
dfft_allocate_aligned_memory(&(p->scratch_2),sizeof(cpx_t)*scratch_size);
dfft_allocate_aligned_memory(&(p->scratch_3),sizeof(cpx_t)*scratch_size);
#else
return 3;
#endif
}
else
{
#ifdef ENABLE_CUDA
dfft_cuda_allocate_aligned_memory(&(p->d_scratch),sizeof(cuda_cpx_t)*scratch_size);
dfft_cuda_allocate_aligned_memory(&(p->d_scratch_2),sizeof(cuda_cpx_t)*scratch_size);
dfft_cuda_allocate_aligned_memory(&(p->d_scratch_3),sizeof(cuda_cpx_t)*scratch_size);
#else
return 2;
#endif
}
p->input_cyclic = input_cyclic;
p->output_cyclic = output_cyclic;
#ifdef ENABLE_CUDA
#ifndef NDEBUG
p->check_cuda_errors = 1;
#else
p->check_cuda_errors = 0;
#endif
#endif
p->row_m = row_m;
/* before plan creation is complete, an initialization run will
* be performed */
p->init = 1;
return 0;
}