VImage
VAniso3d(VImage src,VImage dest,VShort numiter,
VShort type,VFloat kappa,VFloat alpha)
{
VImage tmp1=NULL,tmp2=NULL;
int nbands,nrows,ncols;
int b,r,c,iter;
float delta;
float dx,dy,dz,d,u,v;
float ux1,ux2,uy1,uy2,uz1,uz2;
float b1,b2,r1,r2,c1,c2;
VDouble xmax,xmin;
VBoolean ignore = TRUE;
nbands = VImageNBands(src);
nrows = VImageNRows(src);
ncols = VImageNColumns(src);
if (nbands < 3) VError(" min number of slices is 3");
tmp1 = VConvertImageCopy(src,NULL,VAllBands,VFloatRepn);
tmp2 = VCreateImage(nbands,nrows,ncols,VFloatRepn);
VFillImage(tmp2,VAllBands,0);
xmax = VPixelMaxValue (tmp1);
xmin = VPixelMinValue (tmp1);
delta = 1.0 / 7.0;
dx = dy = dz = 0;
for (iter=0; iter < numiter; iter++) {
for (b=1; b<nbands-1; b++) {
for (r=1; r<nrows-1; r++) {
for (c=1; c<ncols-1; c++) {
u = VPixel(tmp1,b,r,c,VFloat);
if (ignore && ABS(u) < 1.0e-10) continue;
c1 = VPixel(tmp1,b,r,c+1,VFloat);
c2 = VPixel(tmp1,b,r,c-1,VFloat);
r1 = VPixel(tmp1,b,r+1,c,VFloat);
r2 = VPixel(tmp1,b,r-1,c,VFloat);
b1 = VPixel(tmp1,b+1,r,c,VFloat);
b2 = VPixel(tmp1,b-1,r,c,VFloat);
/* col-dir */
dx = c1-u;
dy = r1-r2;
dz = b1-b2;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
ux1 = d*(c1 - u);
dx = u-c2;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
ux2 = d*(u - c2);
/* row-dir */
dx = c1-c2;
dy = r1-u;
dz = b1-b2;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
uy1 = d*(r1 - u);
dy = u-r2;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
uy2 = d*(u - r2);
/* slice-dir */
dx = c1-c2;
dy = r1-r2;
dz = b1-u;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
uz1 = d*(b1 - u);
dz = u-b2;
d = diffusion3d(dx,dy,dz,type,kappa,alpha);
uz2 = d*(u - b2);
/* sum */
v = u + delta*(ux1 - ux2 + uy1 - uy2 + uz1 - uz2);
if (v > xmax) v = xmax;
if (v < xmin) v = xmin;
VPixel(tmp2,b,r,c,VFloat) = v;
}
}
}
tmp1 = VCopyImagePixels(tmp2,tmp1,VAllBands);
}
/*
** output
*/
dest = VCopyImage(src,dest,VAllBands);
xmax = VPixelMaxValue (dest);
xmin = VPixelMinValue (dest);
for (b=1; b<nbands-1; b++) {
for (r=1; r<nrows-1; r++) {
for (c=1; c<ncols-1; c++) {
v = VPixel(tmp2,b,r,c,VFloat);
if (v > xmax) v = xmax;
if (v < xmin) v = xmin;
VSetPixel(dest,b,r,c,(VDouble) v);
}
}
}
VDestroyImage(tmp1);
VDestroyImage(tmp2);
return dest;
}
void mainloop
// ====================================================================
//
// purpos : 2-dimensional diffusion equation solved by FDM
//
// date : 2012-5-10
// programmer : Michel Müller
// place : Tokyo Institute of Technology
//
(
FLOAT *f, /* dependent variable */
FLOAT *fn, /* updated dependent variable */
FLOAT kappa, /* diffusion coefficient */
FLOAT *time, /* time */
FLOAT dt, /* time step interval */
FLOAT dx, /* grid spacing in the x-direction */
FLOAT dy, /* grid spacing in the y-direction */
FLOAT dz /* grid spacing in the z-direction */
)
// --------------------------------------------------------------------
{
int icnt = 1;
double start_time, elapsed_time;
double start_time_total, start_computation_time, elapsed_time_total, elapsed_computation_time;
clock_t ctime_start_computation_time, ctime_start_total_time;
double ctime_elapsed_computation_time, ctime_elapsed_total_time;
long long int numOfStencilsComputed = 0;
long long int idealCacheModelBytesTransferred = 0;
long long int noCacheModelBytesTransferred = 0;
start_time = omp_get_wtime();
ctime_start_total_time = clock() / CLOCKS_PER_SEC;
printf("Starting Reference C Version of 3D Diffusion\n");
printf("kappa: %e, dt: %e, dx: %e\n", kappa, dt, dx);
#pragma omp parallel
#pragma omp master
{
printf("num threads: %d\n", omp_get_num_threads( ));
}
#pragma acc data copy(f[0:XYZ_SIZE]), create(fn[0:XYZ_SIZE])
{
#pragma omp master
{
start_computation_time = omp_get_wtime();
ctime_start_computation_time = clock() / CLOCKS_PER_SEC;
}
do { if(icnt % 100 == 0) fprintf(stderr,"time after iteration %4d:%7.5f\n",icnt+1,*time + dt);
diffusion3d(f,fn,kappa,dt,dx,dy,dz);
numOfStencilsComputed += DIM_X_INNER * DIM_Y_INNER * DIM_Z_INNER;
idealCacheModelBytesTransferred += DIM_X_INNER * DIM_Y_INNER * DIM_Z_INNER * FLOAT_BYTE_LENGTH * 2;
noCacheModelBytesTransferred += DIM_X_INNER * DIM_Y_INNER * DIM_Z_INNER * FLOAT_BYTE_LENGTH * 8;
swap(&f,&fn);
*time = *time + dt;
} while(icnt++ < 90000 && *time + 0.5*dt < 0.1);
#pragma acc wait
#pragma omp master
{
elapsed_computation_time = omp_get_wtime() - start_computation_time;
ctime_elapsed_computation_time = (clock() - ctime_start_computation_time) / (double) CLOCKS_PER_SEC;
}
}
elapsed_time_total = omp_get_wtime() - start_time;
ctime_elapsed_total_time = (clock() - ctime_start_total_time) / (double) CLOCKS_PER_SEC;
double elapsed_computation_time_combined = elapsed_computation_time;
if (elapsed_computation_time_combined <= 0.0) {
elapsed_computation_time_combined = ctime_elapsed_computation_time;
}
aprint("Calculated Time= %9.3e [sec]\n",*time);
aprint("Elapsed Total Time (OMP timer)= %9.3e [sec]\n",elapsed_time_total);
aprint("Elapsed Total Time (CTime)= %9.3e [sec]\n",ctime_elapsed_total_time);
aprint("Elapsed Computation Time (OMP timer)= %9.3e [sec]\n",elapsed_computation_time);
aprint("Elapsed Computation Time (CTime)= %9.3e [sec]\n",ctime_elapsed_computation_time);
aprint("Performance= %7.2f [million stencils/sec]\n",((double)numOfStencilsComputed)/elapsed_computation_time_combined*1.0e-06);
aprint("Bandwidth Ideal Cache Model= %7.2f [GB/s]\n",((double)idealCacheModelBytesTransferred)/elapsed_computation_time_combined*1.0e-09);
aprint("Bandwidth No Cache Model= %7.2f [GB/s]\n",((double)noCacheModelBytesTransferred)/elapsed_computation_time_combined*1.0e-09);
}