diff --git a/csrc/cuda.h b/csrc/cuda.h index 1e83602..567b36b 100644 --- a/csrc/cuda.h +++ b/csrc/cuda.h @@ -13,10 +13,10 @@ __global__ void elementwise_mul_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int size); __host__ void elementwise_mul_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_data); + __global__ void matmul_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int rows1, int cols1, int cols2); + __host__ void matmul_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_data); + __global__ void pow_tensor_cuda_kernel(float* data, float power, float* result_data, int size); __host__ void pow_tensor_cuda(Tensor* tensor, float power, float* result_data); - __global__ void matmul_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int rows1, int cols1, int cols2) - __host__ void matmul_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_data); - #endif /* CUDA_KERNEL_H_ */