Conv cudnn 3d #5783
Conv cudnn 3d #5783
Conversation
| int group_offset_out = | ||
| output_channels / groups * output_height * output_width; | ||
| output_channels / groups * output_height * output_width * output_depth; | ||
| int group_offset_filter = filter->numel() / groups; |
There was a problem hiding this comment.
Do you think it's simpler to write this ?
There was a problem hiding this comment.
According to http://www.cplusplus.com/reference/vector/vector/erase/
Because vectors use an array as their underlying storage, erasing elements in positions other than the vector end causes the container to relocate all the elements after the segment erased to their new positions.
Erasing first two elements will cause memory re-allocation, which is not efficient.
| int group_offset_out = | ||
| output_channels / groups * output_height * output_width; | ||
| output_channels / groups * output_height * output_width * output_depth; | ||
| int group_offset_filter = filter->numel() / groups; |
There was a problem hiding this comment.
group is supported in cudnn7.0 .
paddle/operators/conv_cudnn_op.cu.cc
Outdated
| cudnnConvolutionDescriptor_t cudnn_conv_desc = | ||
| conv_desc.descriptor<T>(paddings, strides, dilations); | ||
|
|
||
| #if CUDNN_VERSION > 6000 |
There was a problem hiding this comment.
#if CUDNN_VERSION > 6000 - > #if CUDNN_VERSION >= 7000 or #if CUDNN_VERSION_MIN(7,0,0)
This place needs to be changed too.
paddle/operators/conv_cudnn_op.cu.cc
Outdated
| @@ -155,19 +200,34 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> { | |||
| cudnnTensorDescriptor_t cudnn_input_grad_desc = nullptr; | |||
There was a problem hiding this comment.
cudnn_input_grad_desc and cudnn_input_desc are the same, you can replace cudnn_input_grad_desc with cudnn_input_desc. Just like this.
Fix #5784