int copy_fftgrid_to_pmegrid(struct gmx_pme_t *pme, const real *fftgrid, real *pmegrid, int grid_index,
int nthread, int thread)
{
ivec local_fft_ndata, local_fft_offset, local_fft_size;
ivec local_pme_size;
int ixy0, ixy1, ixy, ix, iy, iz;
int pmeidx, fftidx;
#ifdef PME_TIME_THREADS
gmx_cycles_t c1;
static double cs1 = 0;
static int cnt = 0;
#endif
#ifdef PME_TIME_THREADS
c1 = omp_cyc_start();
#endif
/* Dimensions should be identical for A/B grid, so we just use A here */
gmx_parallel_3dfft_real_limits(pme->pfft_setup[grid_index],
local_fft_ndata,
local_fft_offset,
local_fft_size);
local_pme_size[0] = pme->pmegrid_nx;
local_pme_size[1] = pme->pmegrid_ny;
local_pme_size[2] = pme->pmegrid_nz;
/* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
the offset is identical, and the PME grid always has more data (due to overlap)
*/
ixy0 = ((thread )*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
ixy1 = ((thread+1)*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
for (ixy = ixy0; ixy < ixy1; ixy++)
{
ix = ixy/local_fft_ndata[YY];
iy = ixy - ix*local_fft_ndata[YY];
pmeidx = (ix*local_pme_size[YY] + iy)*local_pme_size[ZZ];
fftidx = (ix*local_fft_size[YY] + iy)*local_fft_size[ZZ];
for (iz = 0; iz < local_fft_ndata[ZZ]; iz++)
{
pmegrid[pmeidx+iz] = fftgrid[fftidx+iz];
}
}
#ifdef PME_TIME_THREADS
c1 = omp_cyc_end(c1);
cs1 += (double)c1;
cnt++;
if (cnt % 20 == 0)
{
printf("copy %.2f\n", cs1*1e-9);
}
#endif
return 0;
}
//check eriks mail for one thing left to do
int main(int argc,char** argv) {
int prank,i,j,k;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD,&prank);
debug=stderr;
if (argc!=6) {
if (prank==0) printf("Usage: gmx_test_3dfft N M K P1 P2\n");
return 1;
}
gmx_parallel_3dfft_t pfft_setup;
ivec ndata = {atoi(argv[1]),atoi(argv[2]),atoi(argv[3])};
real * real_data;
t_complex * complex_data;
MPI_Comm comm[2];
ivec local_ndata;
ivec local_offset;
ivec local_size;
int P[]= {atoi(argv[4]),atoi(argv[5])};
real* compare;
split_communicator(MPI_COMM_WORLD,comm,P);
gmx_parallel_3dfft_init (&pfft_setup, ndata, &real_data, &complex_data, comm, 0, 0); //last two: slab2index, bReprodusible
gmx_parallel_3dfft_real_limits(pfft_setup,local_ndata,local_offset,local_size);
snew(compare,ndata[0]*ndata[1]*ndata[2]);
srand(time(0)+prank);
if (debug) {
fprintf(debug,"local_ndata: %d %d %d\n",local_ndata[0],local_ndata[1],local_ndata[2]);
fprintf(debug,"local_size: %d %d %d\n",local_size[0],local_size[1],local_size[2]);
fprintf(debug,"local_offset: %d %d %d\n",local_offset[0],local_offset[1],local_offset[2]);
}
for (i=0; i<local_ndata[0]; i++) {
for (j=0; j<local_ndata[1]; j++) {
for (k=0; k<local_ndata[2]; k++) {
compare[i*local_ndata[1]*local_ndata[2]+j*local_ndata[2]+k]=real_data[i*local_size[1]*local_size[2]+j*local_size[2]+k]=((real)rand())/RAND_MAX;
}
}
}
gmx_parallel_3dfft_execute(pfft_setup,GMX_FFT_REAL_TO_COMPLEX,0,0);
gmx_parallel_3dfft_execute(pfft_setup,GMX_FFT_COMPLEX_TO_REAL,0,0);
for (i=0; i<local_ndata[0]; i++) {
for (j=0; j<local_ndata[1]; j++) {
for (k=0; k<local_ndata[2]; k++) {
if (fabs(compare[i*local_ndata[1]*local_ndata[2]+j*local_ndata[2]+k] -
real_data[i*local_size[1]*local_size[2]+j*local_size[2]+k]/(ndata[0]*ndata[1]*ndata[2]))>2*ndata[0]*ndata[1]*ndata[2]*GMX_REAL_EPS) {
printf("error: %d %d %d: %f %f\n",i,j,k,compare[i*local_ndata[1]*local_ndata[2]+j*local_ndata[2]+k],
real_data[i*local_size[1]*local_size[2]+j*local_size[2]+k]/(ndata[0]*ndata[1]*ndata[2]));
}
}
}
}
gmx_parallel_3dfft_destroy(pfft_setup);
MPI_Finalize();
return 0;
}
/* This function is called from gmx_pme_do() only from debugging code
that is commented out. */
void dump_local_fftgrid(struct gmx_pme_t *pme, const real *fftgrid)
{
ivec local_fft_ndata, local_fft_offset, local_fft_size;
gmx_parallel_3dfft_real_limits(pme->pfft_setup[PME_GRID_QA],
local_fft_ndata,
local_fft_offset,
local_fft_size);
dump_grid(stderr,
pme->pmegrid_start_ix,
pme->pmegrid_start_iy,
pme->pmegrid_start_iz,
pme->pmegrid_nx-pme->pme_order+1,
pme->pmegrid_ny-pme->pme_order+1,
pme->pmegrid_nz-pme->pme_order+1,
local_fft_size[YY],
local_fft_size[ZZ],
fftgrid);
}
int copy_pmegrid_to_fftgrid(struct gmx_pme_t *pme, real *pmegrid, real *fftgrid, int grid_index)
{
ivec local_fft_ndata, local_fft_offset, local_fft_size;
ivec local_pme_size;
int ix, iy, iz;
int pmeidx, fftidx;
/* Dimensions should be identical for A/B grid, so we just use A here */
gmx_parallel_3dfft_real_limits(pme->pfft_setup[grid_index],
local_fft_ndata,
local_fft_offset,
local_fft_size);
local_pme_size[0] = pme->pmegrid_nx;
local_pme_size[1] = pme->pmegrid_ny;
local_pme_size[2] = pme->pmegrid_nz;
/* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
the offset is identical, and the PME grid always has more data (due to overlap)
*/
{
#ifdef DEBUG_PME
FILE *fp, *fp2;
char fn[STRLEN];
real val;
sprintf(fn, "pmegrid%d.pdb", pme->nodeid);
fp = gmx_ffopen(fn, "w");
sprintf(fn, "pmegrid%d.txt", pme->nodeid);
fp2 = gmx_ffopen(fn, "w");
#endif
for (ix = 0; ix < local_fft_ndata[XX]; ix++)
{
for (iy = 0; iy < local_fft_ndata[YY]; iy++)
{
for (iz = 0; iz < local_fft_ndata[ZZ]; iz++)
{
pmeidx = ix*(local_pme_size[YY]*local_pme_size[ZZ])+iy*(local_pme_size[ZZ])+iz;
fftidx = ix*(local_fft_size[YY]*local_fft_size[ZZ])+iy*(local_fft_size[ZZ])+iz;
fftgrid[fftidx] = pmegrid[pmeidx];
#ifdef DEBUG_PME
val = 100*pmegrid[pmeidx];
if (pmegrid[pmeidx] != 0)
{
gmx_fprintf_pdb_atomline(fp, epdbATOM, pmeidx, "CA", ' ', "GLY", ' ', pmeidx, ' ',
5.0*ix, 5.0*iy, 5.0*iz, 1.0, val, "");
}
if (pmegrid[pmeidx] != 0)
{
fprintf(fp2, "%-12s %5d %5d %5d %12.5e\n",
"qgrid",
pme->pmegrid_start_ix + ix,
pme->pmegrid_start_iy + iy,
pme->pmegrid_start_iz + iz,
pmegrid[pmeidx]);
}
#endif
}
}
}
#ifdef DEBUG_PME
gmx_ffclose(fp);
gmx_ffclose(fp2);
#endif
}
return 0;
}
