path: root/matrix/generic/cukernel.cu

#ifdef NERV_GENERIC_CUKERNEL
#include <assert.h>
#include <stdio.h>
#include "matrix.h"
#include "cuda.h"
#include "float.h"
#define CUDA_THREADS_N 16
#define CUDA_THREADS_NN ((CUDA_THREADS_N) * (CUDA_THREADS_N))
#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
__global__ void cudak_(