Intel Open Questions

Intel Open Questions

1. Is it possible to create PR to PaddlePaddle repo without using a fork? Yes, but only by contributors who have the write permission. According to Github's permission management rules, if a user could create a PR without forking, it assumes that the user has the write permission to the main repo, thus s/he could git push his/her feature branch into the main repo. However, as the number of contributors of PaddlePaddle grows, we have to formalize the permission management process -- new contributors start from the read permission only, after several rounds of PRs, they could get the write permission if a number of committers endorse his/her work.

2. What is kAnyLayout. How should we treat tensor in such layout when we want to transform it to some other layout, let say the MKLDNN one?

   https://github.com/PaddlePaddle/Paddle/issues/8298

3. Why only convolution and pooling have “use_cudnn” flag both in paddle v2 and fluid? Does it make sense to allow user to choose if cudnn has to be used only for these two OPs?

   cuDNN was designed to accelerate image processing related operators, in particular, convolution and pooling. For these operators, we do prefer that our users could benefit from cuDNN; however, not all users have NVIDIA hardware required by cuDNN. Therefore, we want that (1) our image related operators could work with and without cuDNN, and (2) our unit tests can verify that these operators work with and without cuDNN. The operator attributes use_cudnn currently serves the purpose of specifying explicitly if the operators should use cuDNN during testing.

   As long as we are sure that any other operator could benefit from cuDNN, we should add the use_cudnn attribute to them too.

4. How does Python executes C++ code? Particularly, how C++ operations are registered to be available in Python

   • We use pybind to register C++ OPs.

5. We're going to add new DataLayout enum like kMKLDNN in paddle to avoid unnecessary transformations to MKLDNN format, which is required for this library to work efficiently. Are you OK with that?

   • Yes, you can add new layout type.
   • Fluid will provide InferLayout/ShareLayout to set/pass layout alone with Tensor.
   • Fluid will add a registry for registering layout transformation functions to the framework. Intel can use it to register data layout transformation function, and the framework will call it automatically.

6. Should we use both global use_mkldnn flag and per-OP use_ mkldnn flag as it is in paddle v2?

   • In fluid, we should use MKLDNNPlace instead of use_mkldnn flag. For examples , place = fluid.CUDAPlace(0).
   • Fluid will provide Priority mechanism and a better configuration interface.

7. In paddle v2 there are c++UTs for operators. In fluid it seems that there are only python UTs. Should we apply to that and create UTs for MKLDNN OPs in python?

   • Yes, and you should modify the if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): places.append(core.CUDAPlace(0)), adding MKLDNNPlace to ensure the UTs.
   • Python UT actually call c++ core to do computation.

8. Can you define “Intel” label and use it with issues you find we (Intel) team should look at?

   • Yes, we can. But how about "MKLDNN" label instead of "Intel" label?