Neon Depthwise Convolution Transpose Function #3792
Conversation
There was a problem hiding this comment.
I see that the NeonDepthwiseConvTranspose support the situation that stride == 2. But why not change here https://github.com/hedaoyuan/Paddle/blob/90bf4f60aea012a3eeb819fe4655069d66dbe6e6/paddle/function/neon/NeonDepthwiseConvTranspose.cpp#L102
| @@ -474,6 +475,97 @@ struct DepthwiseConvKernel<4, 2> { | |||
| } | |||
| }; | |||
|
|
|||
| template <class T> | |||
There was a problem hiding this comment.
Should we put the padding function into the neon_util file or something else. In case there are neon extensions of other convolution which use padding method.
There was a problem hiding this comment.
I think it can be moved into neon_util.h after it is really needed.
In conv_transpose stride only used to convert input data(some operator like padding), after converted, the convolution process can be considered to be a stride size of 1. |
|
LGTM |
Add a depthwise convolution transpose function based on ARM-NEON optimization.