Fixed multi-GPU training for Tensor Cores

AlexeyAB

2018-03-09 880cf187d87c904f5fe574802ecff99118643f2d

 src/avgpool_layer_kernels.cu

@@ -1,3 +1,7 @@
#include "cuda_runtime.h"
#include "curand.h"
#include "cublas_v2.h"

extern "C" {
#include "avgpool_layer.h"
#include "cuda.h"