18 lines · cpp
1#include <gpuintrin.h>2#include <stdint.h>3 4[[clang::loader_uninitialized]]5__gpu_local uint32_t shared_mem[64];6 7extern "C" __gpu_kernel void localmem_static(uint32_t *out) {8 shared_mem[__gpu_thread_id(0)] = 2;9 10 __gpu_sync_threads();11 12 if (__gpu_thread_id(0) == 0) {13 out[__gpu_block_id(0)] = 0;14 for (uint32_t i = 0; i < __gpu_num_threads(0); i++)15 out[__gpu_block_id(0)] += shared_mem[i];16 }17}18