31b3041e88
This patch adds a pytorch interface to npcomp. This interface is modeled
after pytorch_xla and exposes the MLIR-based flow as a virtual device (similar
to a gpu device or the xla backend). Usage is intended to be something like:
dev = torch_mlir.mlir_device()
t0 = torch.randn((4,4), device=dev)
t1 = torch.randn((4,4), device=dev)
t2 = t0 + t1
t2_mlir = torch_mlir.get_mlir( t2 )
t2_cpu = t2.to('cpu')
In this case t2_cpu would contain the result of the computation, and t2_mlir
contains the mlir description of the computation. Note that this also
properly returns backward paths synthesized by pytorch. There are several
parts of this:
1) A tensor type (implemented by tensor.* and tensor_impl.*)
2) The device modeling (aten_mlir_bridge.*, aten_mlir_device.*, aten_mlir_type*)
3) a temporary IR (implemented by ir.cpp)
There is also a reference lowering directly from the ATen dialect to C
function calls consisting of two parts:
1) The driver that uses the IR to generate MLIR, run Passes and compile the
result using mlir::ExecutionEngine (implemented by jit.cpp and
mlir_gen.cpp)
2) A runtime library implemented by lib/aten_ops.cpp. Most of the operations
are implemented by callbacks into the torch C++ libraries.
Some aspects of this are known to be less than optimal, in particular:
1) There's some function definitions that don't live in the file corresponding
to their declaration.
2) More aspects of this (e.g. the IR) seem like they should be automatically
generated.
3) It's unclear to me how much of the 'IR' is actually necessary, or whether
MLIR could be created on the fly.
Note that this code is licensed in a way similar to pytorch, with the
intention that eventually (when npcomp reaches some maturity) it should be
pushed there. (see frontends/pytorch/LICENSE) The code is also structured
much closer to the pytorch coding style than the LLVM coding style.
|
||
|---|---|---|
| .. | ||
| csrc | ||
| lib | ||
| test | ||
| utils | ||
| CMakeLists.txt | ||
| LICENSE | ||
| README.md | ||
README.md
NPComp - PyTorch frontend integration
This directory contains optional components for interfacing PyTorch to NPComp. Integration is targeted at multiple levels:
- Via program capture with a ATen pseudo-device.
- Via IR-level integration with PyTorch (via tracing or scripting interfaces).
- Interfaces to facilitate checking against reference implementations and verification.
In all situations, the target dialects are maintained in the outer project, along with their lowerings to common intermediate dialects and backends. This directory should be purely about interfacing with the PyTorch/LibTorch components for extracting and executing programs.
The code in this directory is intended to integrate tightly with pytorch, and follows the code style for pytorch. See the overall documentation for frontends for further details about code layout and integration philosophy. In particular, this directory exists to provide a working frontend to an MLIR based pytorch compilation flow and is not intended to be contributed to the LLVM monorepo. If the project is successful, it makes more sense to either break it out as an independent project that depends on LLVM/MLIR/npcomp or contribute it upstream to PyTorch. However, as it will be quite some time before the components are in a state to support such a dependency, it is being carried in-tree in the interim.
Program capture with a ATen pseudo-device.
Integration with a pseudo-device is typified by code like the following:
import npcomp.frontends.pytorch as torch_mlir
dev = torch_mlir.mlir_device()
t0 = torch.randn((4,4), device=dev)
t1 = torch.randn((4,4)).to(dev)
t2 = t0 + t1
t2_mlir = torch_mlir.get_mlir( t2 )
t2_cpu = t2.to('cpu')
In this case t2_cpu contains the result of the computation, and t2_mlir
contains the mlir description of the computation. Tensors are allocated
directly on the virtual device using the device= argument, or computed on
the host and then moved to the virtual device using the to(dev)
call. Subsequent calls on those tensors construct a graph of computation, but
do not perform compute in most cases. This computation graph is returned in
MLIR format by the get_mlir call, or lazily evaluated to return a regular
pytorch tensor by the to(cpu) call.
This technique has several advantages and disadvantages. For training use cases, this technique generates a backward path automatically using the same method that pytorch natively uses. The resulting graph also tends to be simpler, since it will not reflect conditionals in the original python code. Lastly, it is natural if MLIR is being used as a frontend target for an actual device of some sort. In this case, the MLIR could go through a device-specific lowering path and the resulting code run on a device. The implementation of this technique is largely modeled after pytorch_xla.