* This has been anticipated for a long time in that it is quite hard to keep C++ binary compatibility across a system landscape as diverse as PyTorch, LLVM, and this project. This is why we based the PyTorch extension on the MLIR and NPCOMP C APIs only: that is the only sane linkage story for the entire matrix.
* Removes the few LLVM'isms in torch_mlir that had snuck in, using either STL or PyTorch support utilities. The new rule here is that LLVM C++ includes are forbidden at this level and (as stated in the design), torch_mlir should use the PyTorch runtime and support libraries (not introduce an incidental C++ dependency on LLVM).
* Also deletes mnist-playground as it was proving impossible to keep the grid of PyTorch vs system ABI divisions functioning. I am open to a less drastic course here (optional/disabled by default?)
* This gets us pretty close to just using PyTorch's extension builder API, which will be nice for distribution (i.e. it integrates well with the PyTorch ecosystem for deployment). I ended up just simplifying the in-tree CMake support for now.
* Fixes#138
* Incorporates source fixes.
* Uses upstream pybind11 detection logic.
* Patches CI.
* This may break the CI, which will need to be fixed manually in a followup.
* IREE doesn't have proper install support, so there is some temporary hoaky hacking in our CMakeLists.txt to shuttle some symlinks around.
* Reworked the original numpy e2e with IREE test to pipe through iree-translate.
* Removed all of the C++-level dependencies.
* Will generalize and apply to the PyTorch backend in a followup.
* In most situations, this eliminates the need to explicitly set a path to the Torch cmake files.
* Also upgrades to new Python3 find package. (should eliminate 2.x mismatches)
* Since PyTorch is located by asking Python where it is, this eliminates a lot of causes of mismatch. (one source of truth)
* Enables -gsplit-dwarf for both LLVM and NPCOMP, reducing the occurrence of the ~GB scale binaries.
* CMake shared linking seems incompatible with this, so shared objects are still "too big" but there are few of them.
* Reduces disk thrash on clean/install of everything.
* Need to have a dag of shared library deps in order to interop across python extensions (as presented in ODM).
* Introduced add_npcomp_library and friends to mirror the MLIR setup.
* Adds a libNPCOMP.so shared library.
* Redirects tools and extensions to link against libNPCOMP.so (instead of static libs).
* Moves all libraries to lib/, all binaries to bin/ and all python extensions to python/. The invariant is that the rpaths are setup to have a one level directory structure.
* Reworks the _torch_mlir extension to build like the others (still need to come up with a consolidated rule to do this instead of open coded).
* Includes an upstream version bump to pick up needed changes.
Sizes with dynamic linking (stripped, release, asserts enabled):
libNPCOMP.so: 43M (includes much of the underlying LLVM codegen deps)
libMLIR.so: 31M
_npcomp.so: 1.6M (python extension)
_torch_mlir.so: 670K (python extension)
npcomp-capi-ir-test: 6.3K
npcomp-opt: 351K
npcomp-run-mlir: 461K
mnist-playground: 530K
Still more can be done to normalize and optimize but this gets us structurally to the starting point.
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.
Mostly this is CMake cleanup. Several library dependencies are missing, which
is often revealed with shared library builds. Also, it's generally bad to
link directly against LLVM libraries because it fails when using
LLVM_LINK_LLVM_DYLIB. MLIR will pull in libLLVM.so, and there will be
duplicate linkage with the the explicit libraries. There may need to be more
refactoring here.
* Since the manylinux images do not hard-link against python libs (resolving them at runtime), the module must be built without resolving undefined references.
* For some reason, builds on this platform are stricter, exposing dependency ordering issues.
* The CMake bits to build the extension are still somewhat of a mess. I have better versions both upstream and in IREE and will be reconciling. For now, don't look too closely.
* These were separated originally for layering reasons that no longer apply.
* Most of the python extension code is under lib/ with just the module setup in python/.
* Uses local configs and unsupported annotation to disable optional tests.
* This separation was just an artifact of having initial trouble getting lit setup.
This ~totally reworks the existing "runtime" stuff to be more
principled and usable, such as from Python. It's still not fully
production-quality, mainly in the department of memory management (e.g.
it currently leaks memory; we need to figure out "who frees memrefs" +
the analysis and transformation needed to do that (maybe use upstream
buffer allocation pass?)).
The user API is in include/npcomp/runtime/UserAPI.h, though
include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper.
The stuff under {include,lib}/runtime is totally firewalled from the
compiler and tiny (<6kB, though no attention has gone into optimizing
that size). For example, we don't link in libSupport into the runtime,
instead having our own bare bones replacements for basics like ArrayRef
(the JITRuntime helps with bridging that gap, since it *can* depend on
all common LLVM utilities).
The overall features of npcomprt is that it exposes a module that
with multiple function entry points. Each function has arguments and
results that are tensor-valued, and npcomprt::Tensor is the runtime type
that is used to interact with that (and a npcomprt::Ref<T>
reference-counting wrapper is provided to wrap npcomprt::Tensor in the
common case).
From an implementation perspective, an npcomprt module at the
LLVM/object/binary level exposes a single module descriptor struct that
has pointers to other metadata (currently just a list of function
metadata descriptors). All interactions with the npcomp runtime are
keyed off of that module descriptor, including function lookups and
dispatching. This is done to dodge platform ABI issues and also allow
enough reflection to e.g. verify provided arguments.
Most of the compiler-side work here was in LowerToNpcomprtABI and
LowerToLLVM.
Also,
- Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting
annoying to type the underscores/caps.
- misc improvements to bash_helpers.sh
* Adds python bindings for invoking flow, HAL, and VM lowering pipelines.
* Adds pythong bindings for translating to VM module flatbuffer.
* Adds a new backend_test/iree directory and configure lit to find the IREE python rt bindings.
* Open code a simple_invoke.py that exercises the whole pipeline (need real APIs for a lot of this).
* Fails when invoking the function because I never implemented argument marshaling for scalars :(
* Plenty of stuff to do tomorrow.
* Conversions to std for numeric binary expressions, numeric to_boolean, and numeric comparisons.
* Added folders to constant ops to comply with requirements of the pass system.
* Extended the frontend with parameter/result annotation processing for primitives (can specify types for function arguments).
* Added (empty) directory/sources for IREEVM conversions. These are only enabled if IREE is enabled.
* This won't work for most people until some upstream changes percolate.
* Sequestered it behind a flag and a special configure script (cmake_configure_iree.sh) for now.
The code isn't super clean, but is a useful incremental step
establishing most of the boilerplate for future enhancements.
We can't print or return tensors yet so correctness TBD, but I've
stepped into the running code in the debugger so I know it definitely is
running.
This is the first step to building out an npcomp mini-runtime. The
mini-runtime doesn't have to be fancy or complex, but it should at least
be layered nicely (which this code and the current compiler interaction
with the "runtime" code is not). Now that we have boilerplate for e2e
execution in some form, we can build that out.
Stella Laurenzo
stephenneuendorffer
Sean Silva