pytorch

3e560239 - [Vulkan] Implement clone operator (#69551)

[Vulkan] Implement clone operator (#69551) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/69551 Implemented `clone` operator in the Vulkan backend: * Supports only <= 4D tensors. * Internal name is `aten::clone`. * Vulkan `clone` operator accepts only `c10::MemoryFormat::Preserve` and `c10::MemoryFormat::Contiguous` for the argument `c10::optional<c10::MemoryFormat> optional_memory_format`. * Throws an exception if the `optional_memory_format argument` is neither `MemoryFormat::Preserve` nor `MemoryFormat::Contiguous` * CPU implementation: [/aten/src/ATen/native/TensorFactories.cpp::clone()](https://github.com/pytorch/pytorch/blob/3e45739543fbce471fc4ed26ff079efe170de0f1/aten/src/ATen/native/TensorFactories.cpp#L1415) * MKL-DNN implementation: [/aten/src/ATen/native/mkldnn/TensorShape.cpp::mkldnn_clone()](https://github.com/pytorch/pytorch/blob/3e45739543fbce471fc4ed26ff079efe170de0f1/aten/src/ATen/native/mkldnn/TensorShape.cpp#L58) * `self.copy_(src)` calls `copy_()` for Vulkan to Vulkan copy operation ``` vTensor::copy_() vTensor::copy_() X -> Vulkan vTensor::copy_() CPU -> Vulkan vTensor::clone() vTensor::clone() -> MemoryFormat::Preserve vTensor::clone() -> MemoryFormat::Preserve -> self = at::empty_like(src) vTensor::clone() self.copy_(src); -> BEFORE vTensor::copy_() vTensor::copy_() X -> Vulkan vTensor::copy_() Vulkan -> Vulkan vTensor::clone() self.copy_(src); -> AFTER vTensor::copy_() vTensor::copy_() Vulkan -> X vTensor::copy_() Vulkan -> CPU ``` * References: * Function `torch.clone` in PyTorch documentation: https://pytorch.org/docs/stable/generated/torch.clone.html * Pytorch preferred way to copy a tensor: https://stackoverflow.com/questions/55266154/pytorch-preferred-way-to-copy-a-tensor * `torch.memory_format`: https://pytorch.org/docs/stable/tensor_attributes.html?highlight=memory_format#torch.torch.memory_format * `c10::MemoryFormat` definition in [/c10/core/MemoryFormat.h](https://github.com/pytorch/pytorch/blob/3e45739543fbce471fc4ed26ff079efe170de0f1/c10/core/MemoryFormat.h#L28) Test Plan: Build & test on Android: ``` cd ~/fbsource buck build -c ndk.custom_libcxx=false -c pt.enable_qpl=0 //xplat/caffe2:pt_vulkan_api_test_binAndroid\#android-arm64 --show-output adb push buck-out/gen/xplat/caffe2/pt_vulkan_api_test_binAndroid\#android-arm64 /data/local/tmp/vulkan_api_test adb shell "/data/local/tmp/vulkan_api_test" ``` Build & test on MacOS: ``` cd ~/fbsource buck build //xplat/caffe2:pt_vulkan_api_test_binAppleMac ./buck-out/gen/xplat/caffe2/pt_vulkan_api_test_binAppleMac\#macosx-x86_64 ``` Test result on Android (Google Pixel 5): ``` [ RUN ] VulkanAPITest.clone_success [ OK ] VulkanAPITest.clone_success (5 ms) [ RUN ] VulkanAPITest.clone_invalidinputs_exceptions [ OK ] VulkanAPITest.clone_invalidinputs_exceptions (1 ms) ``` Test result on MacOS: ``` [ RUN ] VulkanAPITest.clone_success [ OK ] VulkanAPITest.clone_success (19 ms) [ RUN ] VulkanAPITest.clone_invalidinputs_exceptions [ OK ] VulkanAPITest.clone_invalidinputs_exceptions (2 ms) ``` Reviewed By: SS-JIA Differential Revision: D32923535 fbshipit-source-id: ea29792e1b0080cbbc1c8c7e8bf2beffad9b5c0d