12 lines · cpp
1#include <gpuintrin.h>2#include <stdint.h>3 4extern __gpu_local uint32_t shared_mem[];5 6extern "C" __gpu_kernel void localmem(uint32_t *out) {7 shared_mem[__gpu_thread_id(0)] = __gpu_thread_id(0);8 shared_mem[__gpu_thread_id(0)] *= 2;9 out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =10 shared_mem[__gpu_thread_id(0)];11}12