Paddle
Paddle copied to clipboard
PaddlePaddle
Fix reorder bug in Conv MKLDNN
PR types
Bug fixes
PR changes
OPs
Describe
Fix reorder bug in conv MKLDNN
@yeliang2258 Please update your PR to latest codebase . We would like to evaluate it by our CI to check impact on all models we trace which will take around two days since conflicts are solved.
@yeliang2258 Please update your PR to latest codebase . We would like to evaluate it by our CI to check impact on all models we trace which will take around two days since conflicts are solved.
Done
@jczaja Hello, I want to know, how can I print out the specific data in the onednnl memory
Hi @yeliang2258,
currently fuses with elementwise_add require both elementwise_addends to have same dimensions. That's because append_sum combined with ShareDataWith accumulates both of these tensors in one memory buffer. So if they're not equal, probably some random data is treated as output from convolution.
I reproduced error offline, and it can be fixed by disabling fuse pass if dimensions don't match. Please see changes on my branch https://github.com/Silv3S/Paddle/tree/fix_elem_add_fuses.
Our new strategy for implementing fuse passes with elementwise_add is to use append_binary post-op, which supports broadcasting.
@Silv3S Thank you for your reply. The solution in your branch does work, but when the shape is different, what is the problem with the code I submitted? I create a new memory object and return the calculated data, I don't understand why the result is wrong in this case
@Silv3S Hello, I had a user problem, same as the one here, I solved it using the solution you provided, and I made a PR to fix conv + elementwise_add merge pass first. https://github.com/PaddlePaddle/Paddle/pull/45018