void
init_arrays(double target_time)
{
if (DEBUG) fprintf(stderr, "called init_arrays with target_time = %f \n",
(target_time * 1e6));
#ifdef _ENABLE_CUDA_KERNEL_
if (options.target == gpu || options.target == both) {
/* Setting size of arrays for Dummy Compute */
int N = options.device_array_size;
/* Device Arrays for Dummy Compute */
allocate_device_arrays(N);
double time_elapsed = 0.0;
double t1 = 0.0, t2 = 0.0;
while (1) {
t1 = MPI_Wtime();
if (options.target == gpu || options.target == both) {
cudaStreamCreate(&stream);
call_kernel(A, d_x, d_y, N, &stream);
cudaDeviceSynchronize();
cudaStreamDestroy(stream);
}
t2 = MPI_Wtime();
if ((t2-t1) < target_time)
{
N += 32;
/* Now allocate arrays of size N */
allocate_device_arrays(N);
}
else {
break;
}
}
/* we reach here with desired N so save it and pass it to options */
options.device_array_size = N;
if (DEBUG) fprintf(stderr, "correct N = %d\n", N);
}
#endif
}
/* Print code for initializing the device for execution of the transformed
* code. This includes declaring locally defined variables as well as
* declaring and allocating the required copies of arrays on the device.
*/
static __isl_give isl_printer *init_device(__isl_take isl_printer *p,
struct gpu_prog *prog)
{
p = print_cuda_macros(p);
p = gpu_print_local_declarations(p, prog);
p = declare_device_arrays(p, prog);
p = allocate_device_arrays(p, prog);
p = bind_device_textures_surfaces(p, prog);
return p;
}
void
init_arrays(double target_time)
{
if (DEBUG) fprintf(stderr, "called init_arrays with target_time = %f \n", (target_time * 1e6));
int i = 0, j = 0;
a = (float **)malloc(DIM * sizeof(float *));
for (i = 0; i < DIM; i++) {
a[i] = (float *)malloc(DIM * sizeof(float));
}
x = (float *)malloc(DIM * sizeof(float));
y = (float *)malloc(DIM * sizeof(float));
for (i = 0; i < DIM; i++) {
x[i] = y[i] = 1.0f;
for (j = 0; j < DIM; j++) {
a[i][j] = 2.0f;
}
}
#ifdef _ENABLE_CUDA_KERNEL_
if (options.target == gpu || options.target == both) {
/* Setting size of arrays for Dummy Compute */
int N = options.device_array_size;
/* Device Arrays for Dummy Compute */
allocate_device_arrays(N);
double time_elapsed = 0.0;
double t1 = 0.0, t2 = 0.0;
while (1) {
t1 = MPI_Wtime();
if (options.target == gpu || options.target == both) {
cudaStreamCreate(&stream);
call_kernel(A, d_x, d_y, N, &stream);
cudaDeviceSynchronize();
cudaStreamDestroy(stream);
}
t2 = MPI_Wtime();
if ((t2-t1) < target_time)
{
N += 32;
/* First free the old arrays */
free_device_arrays();
/* Now allocate arrays of size N */
allocate_device_arrays(N);
}
else {
break;
}
}
/* we reach here with desired N so save it and pass it to options */
options.device_array_size = N;
if (DEBUG) fprintf(stderr, "correct N = %d\n", N);
}
#endif
}
