14 lines
472 B
Plaintext
14 lines
472 B
Plaintext
|
#include <stdint.h>
|
||
|
|
#include <cuda_fp16.h>
|
||
|
|
__global__ void vector_add(__half* A, const __half* B, int N) {
|
||
|
|
int idx=blockDim.x*blockIdx.x+threadIdx.x;
|
||
|
|
if(idx<N) A[idx]=__hadd(A[idx],B[idx]);
|
||
|
|
}
|
||
|
|
extern "C" void run_kernel(__half* A, const __half* B, int64_t numel){
|
||
|
|
int threadsPerBlock = 256;
|
||
|
|
int blocksPerGrid = (numel + threadsPerBlock - 1) / threadsPerBlock;
|
||
|
|
|
||
|
|
vector_add<<<blocksPerGrid, threadsPerBlock>>>(A, B, numel);
|
||
|
|
cudaDeviceSynchronize();
|
||
|
|
}
|