cyx-6 commented on code in PR #169: URL: https://github.com/apache/tvm-ffi/pull/169#discussion_r2543882240
########## docs/guides/kernel_library_guide.rst: ########## @@ -0,0 +1,167 @@ +.. Licensed to the Apache Software Foundation (ASF) under one +.. or more contributor license agreements. See the NOTICE file +.. distributed with this work for additional information +.. regarding copyright ownership. The ASF licenses this file +.. to you under the Apache License, Version 2.0 (the +.. "License"); you may not use this file except in compliance +.. with the License. You may obtain a copy of the License at +.. +.. http://www.apache.org/licenses/LICENSE-2.0 +.. +.. Unless required by applicable law or agreed to in writing, +.. software distributed under the License is distributed on an +.. "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +.. KIND, either express or implied. See the License for the +.. specific language governing permissions and limitations +.. under the License. + +==================== +Kernel Library Guide +==================== + +This guide serves as a quick start for shipping python version and machine learning(ML) framework agnostic kernel libraries with TVM FFI. With the help of TVM FFI, we can connect the kernel libraries to multiple ML framework, such as PyTorch, XLA, JAX, together with the minimal efforts. + +Tensor +====== + +Almost all kernel libraries are about tensor computation and manipulation. For better adaptation to different ML frameworks, TVM FFI provides a minimal set of data structures to represent tensors from ML frameworks, including the tensor basic attributes and storage pointer. +To be specific, in TVM FFI, two types of tensor constructs, :cpp:class:`~tvm::ffi::Tensor` and :cpp:class:`~tvm::ffi::TensorView`, can be used to represent a tensor from ML frameworks. + +Tensor and TensorView +--------------------- + +Though both :cpp:class:`~tvm::ffi::Tensor` and :cpp:class:`~tvm::ffi::TensorView` are designed to represent tensors from ML frameworks that interact with the TVM FFI ABI. They are backed by the `DLTensor` in DLPack in practice. The main difference is whether it is an owning tensor structure. + +:cpp:class:`tvm::ffi::Tensor` + :cpp:class:`~tvm::ffi::Tensor` is a completely owning tensor with reference counting. It can be created and passed between C++ and Python side safely. When the counting reference goes to zero, its underlying deleter function will be called to free the tensor storage. + +:cpp:class:`tvm::ffi::TensorView` + :cpp:class:`~tvm::ffi::TensorView` is a non-owning view of an existing tensor, pointing to an existing tensor (e.g., a tensor allocated by PyTorch). + +It is **recommended** to use :cpp:class:`~tvm::ffi::TensorView` when possible, that helps us to support more cases, including cases where only view but not strong reference are passed, like XLA buffer. +It is also more lightweight. However, since :cpp:class:`~tvm::ffi::TensorView` is a non-owning view, it is the user's responsibility to ensure the lifetime of underlying tensor data and attributes of the viewed tensor object. + +Tensor Attributes +----------------- + +For the sake of convenience, :cpp:class:`~tvm::ffi::TensorView` and :cpp:class:`~tvm::ffi::Tensor` align the following attributes retrieval mehtods to :cpp:class:`at::Tensor` interface, to obtain tensor basic attributes and storage pointer: +``dim``, ``dtype``, ``sizes``, ``size``, ``strides``, ``stride``, ``numel``, ``data_ptr``, ``device``, ``is_contiguous`` Review Comment: updated ########## docs/guides/kernel_library_guide.rst: ########## @@ -0,0 +1,167 @@ +.. Licensed to the Apache Software Foundation (ASF) under one +.. or more contributor license agreements. See the NOTICE file +.. distributed with this work for additional information +.. regarding copyright ownership. The ASF licenses this file +.. to you under the Apache License, Version 2.0 (the +.. "License"); you may not use this file except in compliance +.. with the License. You may obtain a copy of the License at +.. +.. http://www.apache.org/licenses/LICENSE-2.0 +.. +.. Unless required by applicable law or agreed to in writing, +.. software distributed under the License is distributed on an +.. "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +.. KIND, either express or implied. See the License for the +.. specific language governing permissions and limitations +.. under the License. + +==================== +Kernel Library Guide +==================== + +This guide serves as a quick start for shipping python version and machine learning(ML) framework agnostic kernel libraries with TVM FFI. With the help of TVM FFI, we can connect the kernel libraries to multiple ML framework, such as PyTorch, XLA, JAX, together with the minimal efforts. + +Tensor +====== + +Almost all kernel libraries are about tensor computation and manipulation. For better adaptation to different ML frameworks, TVM FFI provides a minimal set of data structures to represent tensors from ML frameworks, including the tensor basic attributes and storage pointer. +To be specific, in TVM FFI, two types of tensor constructs, :cpp:class:`~tvm::ffi::Tensor` and :cpp:class:`~tvm::ffi::TensorView`, can be used to represent a tensor from ML frameworks. + +Tensor and TensorView +--------------------- + +Though both :cpp:class:`~tvm::ffi::Tensor` and :cpp:class:`~tvm::ffi::TensorView` are designed to represent tensors from ML frameworks that interact with the TVM FFI ABI. They are backed by the `DLTensor` in DLPack in practice. The main difference is whether it is an owning tensor structure. + +:cpp:class:`tvm::ffi::Tensor` + :cpp:class:`~tvm::ffi::Tensor` is a completely owning tensor with reference counting. It can be created and passed between C++ and Python side safely. When the counting reference goes to zero, its underlying deleter function will be called to free the tensor storage. + +:cpp:class:`tvm::ffi::TensorView` + :cpp:class:`~tvm::ffi::TensorView` is a non-owning view of an existing tensor, pointing to an existing tensor (e.g., a tensor allocated by PyTorch). + +It is **recommended** to use :cpp:class:`~tvm::ffi::TensorView` when possible, that helps us to support more cases, including cases where only view but not strong reference are passed, like XLA buffer. +It is also more lightweight. However, since :cpp:class:`~tvm::ffi::TensorView` is a non-owning view, it is the user's responsibility to ensure the lifetime of underlying tensor data and attributes of the viewed tensor object. + +Tensor Attributes +----------------- + +For the sake of convenience, :cpp:class:`~tvm::ffi::TensorView` and :cpp:class:`~tvm::ffi::Tensor` align the following attributes retrieval mehtods to :cpp:class:`at::Tensor` interface, to obtain tensor basic attributes and storage pointer: +``dim``, ``dtype``, ``sizes``, ``size``, ``strides``, ``stride``, ``numel``, ``data_ptr``, ``device``, ``is_contiguous`` + +:c:struct:`DLDataType` + The ``dtype`` of the tensor. It's represented by a struct with three fields: code, bits, and lanes, defined by DLPack protocol. + +:c:struct:`DLDevice` + The ``device`` where the tensor is stored. It is represented by a struct with two fields: device_type and device_id, defined by DLPack protocol. + +:cpp:class:`tvm::ffi::ShapeView` + The ``sizes`` and ``strides`` attributes retrieval are returned as :cpp:class:`~tvm::ffi::ShapeView`. It is an iterate-able data structure storing the shapes or strides data as ``int64_t`` array. + +Tensor Allocation +----------------- + +TVM FFI provides several methods to create or allocate tensors at C++ runtime. Generally, there are two types of tensor creation methods: + +* Allocate a tensor with new storage from scratch, i.e. :cpp:func:`~tvm::ffi::Tensor::FromEnvAlloc` and :cpp:func:`~tvm::ffi::Tensor::FromNDAlloc`. By this types of methods, the shapes, strides, data types, devices and other attributes are required for the allocation. +* Create a tensor with existing storage following DLPack protocol, i.e. :cpp:func:`~tvm::ffi::Tensor::FromDLPack` and :cpp:func:`~tvm::ffi::Tensor::FromDLPackVersioned`. By this types of methods, the shapes, data types, devices and other attributes can be inferred from the DLPack attributes. + +FromEnvAlloc +^^^^^^^^^^^^ + +To better adapt to the ML framework, it is **recommended** to reuse the framework tensor allocator anyway, instead of directly allocating the tensors via CUDA runtime API, like ``cudaMalloc``. Since reusing the framework tensor allocator: + +* Benefit from the framework's native caching allocator or related allocation mechanism. +* Help framework tracking memory usage and planning globally. + +For this case, TVM FFI provides :cpp:func:`tvm::ffi::Tensor::FromEnvAlloc`. It internally calls the framework tensor allocator. To determine which framework tensor allocator, TVM FFI infers it from the passed-in framework tensors. For example, when calling the kernel library at Python side, there is an input framework tensor if of type ``torch.Tensor``, TVM FFI will automatically bind the :cpp:func:`at::empty` as the current framework tensor allocator by ``TVMFFIEnvTensorAlloc``. And then the :cpp:func:`~tvm::ffi::Tensor::FromEnvAlloc` is calling the :cpp:class:`at::empty` actually: Review Comment: refactored -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
