In terms of IR structure, TorchScript allows types to vary in many
circumstances where MLIR requires pointer-identical types. In particular,
it is valid to pass any subtype in place of a type. For example, if an
`Optional[int]` is required somewhere in the IR, it is legal to pass a
value of just `int` (but not the other way around; see
`torch.prim.unchecked_cast`). In effect, every *use* can have a different
type.
We introduce a new op `torch.derefine` that models that impedance
mismatch. This op allows casting a value from one type to a type that it
is a subtype of to model this behavior.
Recommended review order:
- TorchOps.td for new torch.derefine (and updated docs for
`torch.prim.unchecked_cast`)
- new test code in if.py, loop.py, function-derefine.py
- new code in node_importer.cpp for handling derefinement insertion
- function_importer.cpp and utils changes in torch_to_mlir_utils.cpp
Properly handling derefinement on function boundaries required
relayering the code so that graph_importer.cpp/.h is now
function_importer.cpp/.h because only the `torch::jit::Function`
(actually the `c10::FunctionSchema` it holds) knows the derefined types that are
actually needed at the boundary (see `function-derefine.py` for a test).
Annoyingly, this churns all the functions which are now prefixed with
`__torch__.` but that is more correct anyway (that is their linkage name
in the `torch::jit::CompilationUnit`; the previous `mb.import_function`
was actually buggy in the case of functions calling each other as it
would reference their unqualified name).
With this change, we can import `resnet18` from `torchvision` :)
IR: https://gist.github.com/silvasean/6426a5272d8a6c7caae533fce05ab704
I could not find a corresponding ListIndex in prim, which seems to
translate to a __get_attr__ under the hood. I think the reason a tuple
Index op can exist is because Tuple's are supposed to be frozen, where
List operands can be mutable.
This arises when casting optionals, which happens a lot especially
around handling of default arguments (python `if arg is None` idiom).
In this case, the offending code for the model is in max_pool2d:
[code link](b3bf08e67f/torch/nn/functional.py (L657))
Used by resnet18.
It seems to originate from a helper `_verify_batch_size`:
[code link](b3bf08e67f/torch/nn/functional.py (L2099)).
I couldn't find a way to test `prim::RaiseException` without also having
`prim::Uninitialized`.
This primarily unlocks proper handling of free functions (that is,
functions that are not methods of any torch.nn.Module).
Recommended review order:
- `ivalue_importer.cpp` + `ivalue_import/functions*.py`
- `GlobalizeObjectGraph.cpp` + test case
- misc other stuff
The `torch::jit::CompilationUnit` is basically a backing store or
"context" holding all the possible functions in the program. The
previous code was not explicitly accessing this data structure, since it
just imported the `torch::jit::Function`'s that it saw attached to
methods.
Subtly, any time a TorchScript module called into a free function, the
free function gets incorporated into the torch::jit::CompilationUnit,
but doesn't show up anywhere when dumping the module, except in the
curious pattern:
```
%5 : Function = prim::Constant[name="adaptive_avg_pool2d"]()
%6 : Tensor = prim::CallFunction(%5, %input.1, %4)
```
That is, calls are indirect calls, and are accessed via `prim::Constant`
materializing a function object. Even stranger, the `name` attribute here
doesn't really even tell the full story -- it doesn't correspond to
anything. It turns out that the c10::FunctionType itself actually holds
a pointer to the `torch::jit::Function` in the compilation unit
directly (so there is actually no indirection in prim::CallMethod,
because any two values of the same FunctionType call the same
function!). E.g. when converting the IR to bytecode, the "name" is
ignored [code link](1d6bd15790/torch/csrc/jit/runtime/interpreter.cpp (L937)).
We do import `prim::CallFunction` as a `std.call_indirect` though
because it's more braindead to do it that way (it gets canonicalized to
a direct call easily).
With this, we can import BERT!
```
pt_util ~/tmp/bert.pt --import --exported-name=forward \
| npcomp-opt -torch-globalize-object-graph -inline -symbol-dce
```
https://gist.github.com/silvasean/fe7735ff5d065cc9216f7b0346d0e977
The test case here is a bit unconventional -- it isn't actually valid
Python. To figure out how to generate it I had to go search the PyTorch
codebase for "NumToTensor" and work backward. In this case I found
this
[code](649760e5f1/torch/csrc/jit/frontend/ir_emitter.cpp (L464))
which via a wild guess I was able to turn into a test case.
In this case it didn't take me too long, but when doing this kind of
"add a bunch of trivial stuff to bring up a real model", I'm starting to
think that we might skimp on test cases when it's fairly trivial and not
obvious how to test with a small test.
- `module_import -> ivalue_import`, as it mainly tests ivalue_importer.cpp
- `graph_import -> node_import`, as it mainly tests node_importer.cpp
- graph_importer.cpp does call into node_importer.cpp, but doesn't do
much.
This was getting pretty confusing. Also add README.md's in each
directory for more clarity.
Adds support for lowering a torch.nn.Conv2d module to the Torch Dialect through TorchScript import.
Generated IR can be viewed here:
https://gist.github.com/brycearden/6c0f790115c4577249372ef82768e6fd
Required implementing support for tuple in the ivalue importer and list in the node importer.
The first use case is to annotate certain program constructs as either
exported or private. In this commit we plumb it down to
GlobalizeObjectGraph which makes use of this information.
Recommended review order:
1. class_annotator.h/.cpp + `test/module_import/annotations/*`
- New abstractions to communicate with Python code and annotate.
2. IR changes in TorchOps.td
- Adding "private" attribute to various things.
3. ivalue_import.cpp changes
- Module + ClassAnnotator = annotated IR
4. GlobalizeObjectGraph.cpp + tests
- use new "private" attributes to create "private" IR.
- also, tweak some of the op deleting mechanics, which was triggering
some memory errors / assertions
With this, we can run the classifier through and inline it as follows:
```
frontends/pytorch/utils/pt_util.py --import --exported-name forward ~/tmp/classifier.pt \
| npcomp-opt -torch-globalize-object-graph -inline
```
IR: https://gist.github.com/silvasean/32dcad9f6270557f412094a77cecdd69
With this + manually setting private visibility on everything, a simple
classifier can be reduced to this IR, which is looking pretty lean and
mean:
https://gist.github.com/silvasean/19e7e2e21a61ff197aeac0dd864d188f
Also, include a utility script for importing `.pt` models.
```
pt_util.py --import classifier.pt | npcomp-opt -torch-globalize-object-graph
```
This required restructuring of how we model TorchScript on import. The
main difference is that now we split out a `torch.class_type` that holds
methods and declarations of the types of each slot. This is more
consistent with TorchScript (our previous representation was
"denormalized").
Recommended reading order:
1. check out the description of `torch.class_type` in `TorchOps.td` and
look at `test/Dialect/Torch/ops.mlir` and
`frontends/pytorch/test/module_import/` to familiarize with the new
representation.
- Just look at the new IR. The diff between the old names and new
names is confusing.
2. check out `test/Dialect/Torch/globalize-object-graph*.mlir`
and read along with the pass description in
`include/npcomp/Dialect/Torch/Transforms/Passes.td`
3. Read the code in `GlobalizeObjectGraph.cpp` and miscellaneous changes
in `ivalue_importer.cpp`, `TorchOps.cpp`, etc.
Modify ACAP_Dispatch to work with latest pytorch
-Remove boxed from convolution's m.impl
-Use redispatch and constrainted keyset to replace deprecated
callwithdispatchkey
This required some careful considerations when defining object identity
for tensors. See the comments for how we do it.
This also tracks some basic information for diagnostics.
This required some invasive surgery to graph_importer.h/cpp,
specifically moving most of it into node_importer.h/cpp and relayering
it. The abstraction that it had didn't work well in the recursive
setting that happens with prim::If.
The key observation is that torch::jit::Graph doesn't really correspond
directly to anything on the MLIR side. It's a weird combination of a
context, builder, and function and just holds a `torch::jit::Block`. It
is `torch::jit::Node` and `torch::jit::Block` which form the recursive
structure analogous to MLIR's operation/region/block. So
node_importer.h/cpp makes sense as a core building block.
As part of doing this, I did venture a bit into the AcapController code,
and realize now that there is functionality duplicated there with the
ivalue importer. Will refactor that soon.
It turns out that this was easiest to structure as a general IValue
importer, since torch module are just one of the possible IValue's.
We import the IValue object graph in a braindead fashion into basicpy
ops and a new `torch.nn_module` op that is used to model the
attributes/methods of a torch::jit::Module IValue. See `Torch/ops.mlir`
for an example, and also check out the .py import tests in
`frontends/pytorch/test/module_import`.
As part of this change, a few housekeeping tasks:
- extract some helpers from graph_importer.cpp
- more helpers around the C API
- misc touchups
This allows building NPCOMP as an external project of LLVM, similar to
how CIRCT can be built: https://github.com/llvm/circt/pull/227.
The CMake options to use this build style look like this:
```
-DLLVM_EXTERNAL_PROJECTS=npcomp \
-DLLVM_EXTERNAL_NPCOMP_SOURCE_DIR=/path/to/mlir-npcomp \
```
* 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
* Going through TODOs on the PyTorch side, this is a big cause of them (not being able to have constants for signed/unsigned).
* Added complex while in here since we're at the phase where it is better to just have things complete than partially done.
* Does not handle all features yet but should conservatively fail on unsupported things.
* Location tracking is still somewhat mismatched between what TorchScript and MLIR do. Likely need a better heuristic for tracking locations from defs for nodes that do not carry location.
* Sets the ground-work for a specialized/generic split but only implements the generic side.
* Had some evidence that this requires a recent bump of PT nightly (within the last month) to pick up pybind11 2.6, which includes some cross-module symbol fixes (vs the previously sync'd version). No source changes, but older versions fail to cast function types at runtime.
Note that unlike aten.matmul which has dynamic behavior
depending on the argument ranks (can do matrix-matrix, matrix-vector,
batch matmul, etc.), aten.mm is just a vanilla matrix
multiply, which can be lowered precisely to tcf.matmul.
The "test" is really just an example that I stared at while getting my
feet wet with this. We probably want something that actually tests this
as part of `ninja check-npcomp`.
After the recent change of cmake variables
from PYTHON_INCLUDE_DIRS to Python3_INCLUDE_DIRS
and PYTHON_LIBRARIES to Python3_LIBRARIES, there were
a few files that still had references to the old
variables. This patch fixes that.
* 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)
* Fixes#107
* I wouldn't say I love what had to be done here. Worth a conversation with the PT devs (probably as part of a rollup of a bunch of this stuff).
* A bit gross because I took the chance to upgrade all of the backend bits to the new MLIR Python bindings and we still co-mingle the old and new for now.
* Since the Python created PassManagers are configured for explicit nesting, I had to upgrade some of the pass pipelines to be explicit.
* The demo in mul_maximum_e2e.py now compiles, runs through PyTorch and through the JIT, prints and asserts the same results.
* I am not claiming that this is the prettiest API in this patch: consider that this is just directly using low-level APIs and there should be an intervening high level API.
* Conversions are very simple, suporting mul, maximum and add (alpha=1 only).
* Example added with pass pipeline needed to run.
* Much missing off of the golden path but sufficient for such simple cases.
* convolution, convolution_backward, _log_softmax, _log_softmax_backward_data, nll_loss_forward, nll_loss_backward, nll_loss2d_forward, nll_loss2d_backward, copy_
* Extends the recognition logic and metadata for handling inplace transformations, optional tensors, ints, lists and dropped args.
* The kernel_calls generated by test_conv_nllloss_grads.py now convert to ATen.
* The result *almost* comes out as a pure tensor program with the exception of the copy_ op, which I will do some followup work to deal with.
* More progress on #97
The current code was inserting all build_list ops
after the last constant op since it was assuming that all
elements being passed in were constants.
This patch replaces that patch with a new function that
inserts the build_list ops before the terminator.
Also modifies test_export_conv2d_fwd.py since its output
no longer matches.
TEST: Added test_export_cat.py which is the code in #102
* This is sufficient to capture the forward and backward pass and gradients of a convolutional model with an nllloss.
* As with the forward conv, the backward conv is a special case wrapped in an enigma on the PyTorch side. There aren't many like it, so special casing is just what we do.
* When I traced this, I found that the copy_ op is not yet boxing compatible so I had to map it manually. If there are many more like this, I'll probably do something a bit more clever to reduce duplication.
* This exposes new signature patterns that will need to be handled by the ATen lowering. Will take care of that next: It will be nice to have an e2e of a non-trivial case with full gradients.
* Fixes#97.
* None's out Device? args.
* Emits bool tensors if needed.
* Adds some stderr tracing to better see what is going on.
* Test case that exercises NLLLoss.
* This test case emits something for backward calculations but there are some issues still to be worked out, so that part is left out of the test case.
* Progress on #97
* Deletes prior code generator from previous attempt (moved some of it into this one).
* Renames old generated tablegen source to "Legacy".
* Generates ODS and import rules for most binary and unary arithmetic ops.
* Removes old generated ops and integration tests that were testing details of the prior setup.
* Two op interfaces, one for querying instance metadata and one for getting static data needed to construct an op from a generic form.
* For torch.generic_kernel ops, metadata is splatted in during capture from Torch (it comes from the op registry, which will work for either device capture or graph import).
* Moved the 'add' out of the generated set so I can experiment on it. It implements the TorchBuildableKernelOpInterface interface which provides its metadata.
* The ATenRecognizeKernelsPass pass generically lowers from a torch.generic_kernel to recognized ops that implement the TorchBuildableKernelOpInterface, handling the various types of transformations that we allow at this stage.
Sean Silva
Bryce Arden
Bairen Yi
Stanley Winata
mikeurbach
Stella Laurenzo
Phoenix Meadowlark
harsh-nod