pengzhao-intel commented on a change in pull request #13530: Integrate MKLDNN
Conv1d and support 3d layout
URL: https://github.com/apache/incubator-mxnet/pull/13530#discussion_r243874808
##########
File path: src/ndarray/ndarray.cc
##########
@@ -454,24 +454,18 @@ void NDArray::Chunk::SetMKLMem(const TShape &shape, int
dtype) {
mkldnn::memory::dims dims;
// These are shapes supprted by MKLDNN.
- if (shape.ndim() == 1 || shape.ndim() == 2 || shape.ndim() == 4
- || shape.ndim() == 5) {
+ if (shape.ndim() >= 1 && shape.ndim() <= 5) {
dims.resize(shape.ndim());
for (size_t i = 0; i < dims.size(); i++)
dims[i] = shape[i];
- } else if (shape.ndim() == 3) {
- // If there are 3 dimensions, we'll force it to 4 dimensions.
- dims.resize(shape.ndim() + 1);
- dims[0] = 1;
- for (size_t i = 0; i < shape.ndim(); i++)
- dims[i + 1] = shape[i];
Review comment:
Is there a performance difference between 3D and 4D implementation?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
