Use half_float16 instead of float32 if defined both CUDNN and CUDNN_HALF. Use Tensor Cores.

This commit is contained in:
AlexeyAB
2018-02-22 22:54:40 +03:00
parent 033e934ce8
commit dda993f3dd
3 changed files with 117 additions and 11 deletions

View File

@ -242,6 +242,8 @@ struct layer{
float * weights_gpu;
float * weight_updates_gpu;
float * weights_gpu16;
float * biases_gpu;
float * bias_updates_gpu;