SMC::SMC ()
{
if (!initialized) {
init_runtime();
init_stencil();
init_chemistry();
init_variables();
initialized = true;
}
build_multifabs();
init_from_scratch();
wt_fb1 = wt_fb2 = wt_chem1 = wt_chem2 = wt_hypdiff = 0.0;
}
int main(int argc, char** argv)
{
int i,j,k;
struct timeval t0,t1;
double time;
int validate = 0;
int ni = NI, nj = NJ;
int niter = STEPS;
int di = EDIM, dj = EDIM;
// stencil weights
float w0 = 0.166666666f;
float w1 = 0.333333333f;
float w2 = 0.166666666f;
float w3 = 0.166666666f;
float w4 = 0.166666666f;
i = 1;
while (i < argc) {
if (!strcmp(argv[i],"-n")) { ni = atoi(argv[++i]); nj = atoi(argv[++i]); }
else if (!strcmp(argv[i],"-i")) niter = atoi(argv[++i]);
else if (!strcmp(argv[i],"-d")) { di = atoi(argv[++i]), dj = atoi(argv[++i]); }
else if (!strcmp(argv[i],"--validate")) validate = 1;
else if (!strcmp(argv[i],"--help") || !strcmp(argv[i],"-h")) goto help;
else {
fprintf(stderr,"unrecognized option: %s\n",argv[i]);
help:
ERROR("use -n [internal X size] [internal Y size] -i [iteration step count] -d [number of Epiphany cores/threads] --validate\n");
}
++i;
}
if(ni%di) ERROR("ni = %d is not divisible by di = %d\n", ni, di);
if(nj%dj) ERROR("nj = %d is not divisible by dj = %d\n", nj, dj);
printf("Using N = {%d x %d}, # iterations = %d, # cores = {%d x %d} = %d\n", ni, nj, niter, di, dj, di*dj);
// From here on, we're using padded ni and nj values
ni += 2;
nj += 2;
// allocate memory on host
float* A = (float*)malloc(ni*nj*sizeof(float));
float* B = (float*)malloc(ni*nj*sizeof(float));
// initialize
init_stencil(A, B, ni, nj);
gettimeofday(&t0,0);
update_stencil_epiphany(A, B, ni, nj, di, dj, niter, w0, w1, w2, w3, w4);
gettimeofday(&t1,0);
time = t1.tv_sec - t0.tv_sec + 1e-6*(t1.tv_usec - t0.tv_usec);
float gflops = __gflops(time, ni-2, nj-2, niter);
printf("Epiphany Performance.... : %f GFLOPS (includes overhead)\n",gflops);
printf("Execution Time.......... : %f seconds\n",time);
if (validate) {
float* A_validate = (float*)malloc(ni*nj*sizeof(float));
float* B_validate = (float*)malloc(ni*nj*sizeof(float));
init_stencil(A_validate, B_validate, ni, nj);
printf("Validating on CPU host....\n");
gettimeofday(&t0,0);
update_stencil_cpu(A_validate, B_validate, ni, nj, niter, w0, w1, w2, w3, w4);
gettimeofday(&t1,0);
int errors = validate_stencil(B, B_validate, ni, nj);
print_stencil(B_validate, ni, nj);
time = t1.tv_sec - t0.tv_sec + 1e-6*(t1.tv_usec - t0.tv_usec);
printf("CPU Execution time... : %f seconds\n",time);
printf("Errors............... : %d (%0.1f%%)\n", errors, 100.0f*errors/((ni-2)*(nj-2)));
}
return 0;
}
