| 1 | // RUN: %libomptarget-compile-generic && %libomptarget-run-generic |
| 2 | // RUN: %libomptarget-compileopt-generic && %libomptarget-run-generic |
| 3 | |
| 4 | #include <stdio.h> |
| 5 | #include <stdlib.h> |
| 6 | |
| 7 | int main() { |
| 8 | long unsigned *DP = 0; |
| 9 | int N = 32; |
| 10 | int Threads = 64; |
| 11 | int Teams = 10; |
| 12 | |
| 13 | // Allocate ~55MB on the device. |
| 14 | #pragma omp target map(from : DP) |
| 15 | DP = (long unsigned *)malloc(size: sizeof(long unsigned) * N * Threads * Teams); |
| 16 | |
| 17 | #pragma omp target teams distribute parallel for num_teams(Teams) \ |
| 18 | thread_limit(Threads) is_device_ptr(DP) |
| 19 | for (int i = 0; i < Threads * Teams; ++i) { |
| 20 | for (int j = 0; j < N; ++j) { |
| 21 | DP[i * N + j] = i + j; |
| 22 | } |
| 23 | } |
| 24 | |
| 25 | long unsigned s = 0; |
| 26 | #pragma omp target teams distribute parallel for num_teams(Teams) \ |
| 27 | thread_limit(Threads) reduction(+ : s) |
| 28 | for (int i = 0; i < Threads * Teams; ++i) { |
| 29 | for (int j = 0; j < N; ++j) { |
| 30 | s += DP[i * N + j]; |
| 31 | } |
| 32 | } |
| 33 | |
| 34 | // CHECK: Sum: 6860800 |
| 35 | printf(format: "Sum: %li\n" , s); |
| 36 | return 0; |
| 37 | } |
| 38 | |