blob: 9542e2cb1d64847dfdb58af48b17e0225377a9d9 (
plain)
1
2
3
4
5
6
7
8
9
10
11
|
#include <gpuintrin.h>
#include <stdint.h>
extern __gpu_local uint32_t shared_mem[];
extern "C" __gpu_kernel void localmem(uint32_t *out) {
shared_mem[__gpu_thread_id(0)] = __gpu_thread_id(0);
shared_mem[__gpu_thread_id(0)] *= 2;
out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =
shared_mem[__gpu_thread_id(0)];
}
|