diff --git a/norch/csrc/cuda.cu b/norch/csrc/cuda.cu index bf413e3..071828f 100644 --- a/norch/csrc/cuda.cu +++ b/norch/csrc/cuda.cu @@ -5,7 +5,7 @@ #define THREADS_PER_BLOCK 128 #define TILE_SIZE 32 -#define SHMEM_SIZE THREADS_PER_BLOCK_SUM * sizeof(float) +#define SHMEM_SIZE THREADS_PER_BLOCK * sizeof(float) __host__ void cpu_to_cuda(Tensor* tensor) {