Commit Graph

645 Commits (ce7abf4911f7defd5ad56a644c4be7242677eea3)

Author SHA1 Message Date
Jiahao Li ce7abf4911
Add dtype functions for ops that take dtype from 2nd operand (#1891) 2023-02-21 14:54:41 -08:00
Ramiro Leal-Cavazos 63945a2fd4
Change dtype functions interface to take ints tuple for each tensor (#1865)
The original design for the dtype functions outlined in
https://github.com/llvm/torch-mlir/issues/1462 was unable to properly
handle ops that take optional tensors as an input when the optional
tensor has a value of None. By the time the op gets imported into
torch-mlir, if an optional value is None, all information about the
original type is lost from the op type signature, preventing
torch-mlir from knowing if a value of None was from an optional tensor
or not, which was crucial in the original design since each tensor
argument must be turned into two separate arguments for the dtype
function.

This commit changes the interface to dtype functions such that each
tensor turns into a tuple of two ints, the first representing the rank
of the tensor and the second the dtype of the tensor. Since now there
is a one-to-one correspondence between the operands of an op and the
operands of its dtype function, there is no ambiguity about which
operand of the op corresponds with which operand of the dtype
function.

To test the implementation, this commit defines dtype functions for
the convolution ops, all of which take one optional tensor as an
argument.
2023-02-13 17:56:09 -08:00
Ramiro Leal-Cavazos 981ac88758
Add dtype functions for two tensor promotion ops (#1831)
This commit adds dtype functions for ops in RefineTypes under the
category of "Promote the two dtypes". The only ops not added here are
convolution ops, since they take an optional tensor argument, and the
dtype pipeline currently does not correctly handle that case. I will
add a follow up patch fixing this.

This commit also adds two helper functions that perform a very
thorough testing of dtype functions. The helper function
`_check_two_tensor_op` is able to independently test invalid input
dtypes and invalid output dtypes.

Lastly, this commit also XFAILs "MobilenetV3Module_basic".
2023-02-01 22:30:27 +00:00
Jiahao Li 83d4e89d25
Add dtype functions for floating point ops (#1813) 2023-01-20 18:39:41 +00:00
Ramiro Leal-Cavazos 8cae5ba507
Add dtype functions for comparison ops (#1806)
This commit adds dtype functions for comparison ops that always return
a tensor of dtype `i1`.
2023-01-16 22:32:23 +00:00
Ramiro Leal-Cavazos ac94ba22e2 Move dtype functions into their own section in lib gen file
In order to easily keep track of the dtype functions that have been
moved to `abstract_interp_lib_gen.py` and make it easier to add new
ones, this commit groups all the dtype functions together, rather than
having them interspersed between the shape functions.
2023-01-16 20:16:48 +00:00
Jiahao Li 4f94831fed
[LINALG][TOSA][MHLO] Add e2e support for aten bitwise ops (#1753) 2023-01-11 14:40:03 -08:00
Vivek Khandelwal fd236b2c89 [MLIR][TORCH] Add decomposition for prims.var and prims.sqrt op
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2023-01-11 17:39:10 +05:30
Vivek Khandelwal b966733e04 build: manually update PyTorch version
Set PyTorch and TorchVision version to nightly release 2023-01-08.

Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2023-01-11 17:39:10 +05:30
Gleb Kazantaev c8b867b876
Added support for aten::norm.ScalarOpt_dim (#1774)
* Added support for aten::norm.ScalarOpt_dim

* Disable NormalizeModule_basic for linalg
2023-01-10 13:08:25 -05:00
Jiahao Li 8dc5d985eb
Add e2e support for aten logical or/and/xor/not ops (#1761) 2023-01-03 18:11:25 -08:00
Ramiro Leal-Cavazos 273664ded6
[custom op] Replace `tanh` dtype function with `expm1` (#1769)
This commit replaces the `tanh` dtype function, which was being used
to test the implementation of dtype functions in
a710237437, with a dtype function for
`expm1`. The dtype function for `expm1` is identical to the `tanh`
one, so the same level of testing is maintained.

Currently, there are ops getting dtype information from the
`RefineTypes` pass and ops getting dtype information from the
`TorchDtypeRefinementPipeline`. Since each pass can only propagete
dtype information for the ops it knows how to handle, some models with
many ops handled in both passes require the two dtype propagation
passes to execute many times, reaching the iteration limit set in the
`LowerToBackendContractPass`. To temporarily avoid this issue while
the migration to `TorchDtypeRefinementPipeline` is finished, this
commit switches `tanh` to `expm1`, since the latter is used a lot less
in large models.
2023-01-03 14:18:26 -08:00
Srirammaswamy a88e3766e8
Add E2E support for LeakyRelu and LeakyReluBackward ops (#1733)
Co-authored-by: srirammaswamy <srirammaswamy@gmail.com>
2023-01-03 08:30:16 -08:00
Ashay Rane ac780529b4
Revert e2e support for aten logical or/and/xor/not ops (#1757)
This reverts commit eaab9be207, since it
is causing the post-merge CI tests to fail, causing subsequent PRs to be
blocked.  Specifically, the tests
`ElementwiseAtenLogicalAndOpPromoteBroadcastModule_basic` and
`ElementwiseAtenLogicalXorOpPromoteBroadcastModule_basic` fail because
the oracle does not match the computed result.  This patch reverts the
commit to make the post-merge builds green again.
2022-12-29 21:01:06 -06:00
Shivam Gupta 2f45959f0d
Prelu lowering to linalg (#1712)
Prelu lowering to linalg
2022-12-28 08:51:33 +05:30
Jiahao Li eaab9be207
Add e2e support for aten logical or/and/xor/not ops (#1752) 2022-12-26 10:23:38 +08:00
Ramiro Leal-Cavazos 3260a1ea6e
Allow passing traced `torch.nn.Module`s into `torch_mlir.compile` (#1743)
This commit adds support for passing to `torch_mlir.compile` the
result of running `torch.jit.trace` on a model by relaxing the
condition that checks if the model is already in JIT IR to allow any
`torch.jit.ScriptModule`.

Fixes https://github.com/llvm/torch-mlir/issues/1739
2022-12-22 08:39:55 -08:00
Jiahao Li 60a139271d
Add aten.std.correction op and its decomposition (#1731) 2022-12-21 21:02:40 -08:00
Jiahao Li 15b249777b
[Torch][MHLO] Decompose aten.copy op. Lower aten.rsqrt & sigmoid to mhlo. (#1734) 2022-12-22 10:13:59 +08:00
Chi_Liu b2cefc0b64
[TOSA] Add aten.masked_fill.Tensor/Scalar support (#1735) 2022-12-21 08:56:07 -08:00
Jae Hoon (Antonio) Kim 1d695239ff
Unrevert #1724 (#1737)
* Unrevert #1724

* Update pytorch requirements.txt
2022-12-20 11:17:21 -05:00
Abhishek Varma 66d7a412cb [RefineTypes] Fix knowledge dtype for `aten.embedding` op
-- The dtype of the result of `aten.embedding` should match that of
   the `weight` operand's (operand[0]) instead of hardcoding to f32.
-- This commit aims to provide a fix for the same.

Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
2022-12-20 19:56:12 +05:30
Ashay Rane dd1cf578a6
build: fix LTC code after upstream PyTorch change (#1727)
pytorch/pytorch@140a3139 reverted a change from yesterday, causing the
RollPyTorch action to break.  This patch reverts the corresponding
change in the torch-mlir LTC code.

This patch also re-enables tests that were previously marked as XFAIL.
2022-12-16 13:07:38 -06:00
ataheridezfouli-groq 17ee643aeb
[TORCH] Add Complex Number support (#1673)
Add Complex number dtype support to torch tensors. Add
aten.fft_fft op to test complex numbers.
2022-12-15 21:40:01 +00:00
Jae Hoon (Antonio) Kim a2a93891ea
Replace asIntArrayRefSlow with macro (#1724)
* Replace asIntArrayRefSlow with macro

* Update pytorch requirements.txt
2022-12-15 11:52:41 -05:00
Prashant Kumar 8ba77ae2a5 Yapf Format `refbacked.py`. 2022-12-15 21:19:52 +05:30
Prashant Kumar 564403e3a1 Add float16 support in the refbackend.
This will require https://reviews.llvm.org/D139121 patch to go through.
2022-12-15 21:19:52 +05:30
Sean Silva af9e8a5e63 [torchdynamo] Move to aot_autograd instead of raw make_fx
As [@ezyang suggested](https://github.com/pytorch/pytorch/issues/90276#issuecomment-1339791275),
use `torch._dynamo.optimizations.training.aot_autograd` instead of raw
`make_fx`. This is more future proof and gives us the backward pass and
functionalization. We don't currently get functionalization because of
https://github.com/pytorch/pytorch/issues/90759

This also incidentally fixes the source location handling, which makes
`lockstep_basic.py` give an accurate source location!
2022-12-15 01:55:50 -08:00
Ahmed S. Taei b1f6832849
Add aten.slice.Tensor & aten.cat folders (#1691) 2022-12-13 13:02:47 -08:00
Ramiro Leal-Cavazos a710237437
[custom op] Generalize shape library logic to work with dtypes (#1594)
* [custom op] Generalize shape library logic to work with dtypes

This commit generalizes the shape library logic, so that dtype rules
for ops can also be expressed using the same mechanism. In other
words, each op can now have a shape function and a dtype function
specified in Python that is imported during lowering to calculate the
shapes and dtypes throught a program. For more information about how
to specify a dtype function, see the updated
`docs/adding_a_shape_and_dtype_function.md`.

For those not familiar with how the shape library works, the file
`docs/calculations_lib.md` provides an overview.
2022-12-13 08:25:41 -08:00
Ashay Rane 430737b820
[cleanup] fix naming of private variable according to the style guide (#1704) 2022-12-12 09:04:46 -06:00
Vivek Khandelwal d4862ec611 [MLIR][TORCH] Add e2e support for aten.var_mean op
Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2022-12-12 15:46:54 +05:30
Vivek Khandelwal f783e19dcb Revert "[MLIR][TORCH] Fix mean and mean.dim op for large-sized inputs"
This reverts commit 55c7e66aa7.
2022-12-09 19:30:46 +05:30
Sean Silva 7731211d02 Remove eager_mode
This was an experimental attempt at rolling out own op-by-op executor
with `__torch_dispatch__`, but it proved difficult to make it robust.
Op-by-op execution is very easy to implement robustly now with the
PyTorch 2.0 stack, so we don't need eager_mode.

Downstream users were using eager_mode to implement lockstep numerical
accuracy debuggers. We implemented the same functionality with
TorchDynamo in https://github.com/llvm/torch-mlir/pull/1681 so now there
is not much reason to continue maintaining it.
2022-12-09 03:50:00 -08:00
Gleb Kazantaev 804f9f1f8f
Extended TorchMLIRLoweringContext with virtual CreateComputation method (#1699)
* Extended TorchMLIRLoweringContext with virtual CreateComputation method

* Fix device_data_cast return value
2022-12-08 15:57:07 -05:00
Sean Silva e8511840c3 [cleanup] Use a single function pipeline for TOSA->Linalg
This should run faster and is overall clearer.
2022-12-08 09:02:38 -08:00
Sean Silva 69171c246a [RefBackend] Add elementwise fusion and buffer deallocation
This gives some decent improvements to memory consumption and latency of
testing. I would have expected buffer-deallocation to actually make a
big difference to the final process RSS but it doesn't appear to. Also
running buffer-deallocation later in the pipeline results in
miscompiles. I didn't have the time or interest to dig in deeper, but
something is off.

(numbers below are taken from a single run, but I did do a few runs to make
sure that the variance wasn't that great)

- Linalg-on-Tensors shows memory consumption improvements and some slight speedups.
```
./tools/e2e_test.sh -s -v -c refbackend
fuse=0 dealloc=0
RSS: 3071.33 MB
real    3m58.204s
user    6m22.299s
sys     0m51.235s
fuse=1 dealloc=0
RSS: 2515.89 MB
real    3m34.797s
user    5m56.902s
sys     0m44.933s
fuse=1 dealloc=post-bufferize:
RSS: 2290.25 MB
real    3m42.242s
user    6m0.560s
sys     0m46.335s
```

- TOSA ResNet18 gets significantly faster and uses significantly less memory.
```
time ./tools/e2e_test.sh -s -v -c tosa -f ResNet18
fuse=0 dealloc=0
rss 1328.56 MB
real    0m50.303s
user    0m55.355s
sys     0m12.260s
fuse=1 dealloc=0
rss 859MB
real    0m30.454s
user    0m35.551s
sys     0m11.879s
fuse=1 dealloc=post-bufferize:
rss 851MB
real    0m30.313s
user    0m39.889s
sys     0m11.941s
```

Big thanks to Ramiro for the methodology here for measuring the RSS with
`psutil`:
https://gist.github.com/ramiro050/5b5c2501f7389c008d9029210772c3a8
2022-12-08 03:14:42 -08:00
Ramiro Leal-Cavazos dd35488da5
build: update llvm tag to 798fa4b4 (#1684)
- Support for non-prefixed accessors has been removed. See:
  https://reviews.llvm.org/D136727
- Rename `operands` to `methodOperands` in `prim.CallMethod` since the
  name `operands` overlaps with a builtin method name. See:
  https://reviews.llvm.org/D136727
- Add passes in refbackend to lower memref.subview. See:
  https://reviews.llvm.org/D136377
- Replace `CopyToValueTensorOps` first in `RewriteViewLikeSubgraph` in
  maximize-value-semantics.

  The current implementation of the `RewriteViewLikeSubgraph` pass in
  maximize-value-semantics creates temporarily invalid IR. In
  particular, given a forward slice starting from a
  `CopyToNonValueTensorOp` and ending in `CopyToValueTensorOp`s, the
  pass first replaces all uses of the `CopyToNonValueTensorOp` with
  its operand, which results in all the `CopyToValueTensorOp` users
  having their operand have type `!torch.vtensor`, which is invalid.

  The correct way to do things is to first replace all the
  `CopyToValueTensorOp`s with their operand, and then replace all uses
  of the `CopyToNonValueTensorOp` with its operand.

  This only started failing now because the generated accessor
  `getOperand` for the `CopyToValueTensorOp` now returns a
  `TypedValue<NonValueTensorType>`, which has an assert checking that
  the value returned is of the expected type.
2022-12-07 12:20:41 -08:00
Sean Silva b1f9e09f85 [torchdynamo] Add ResNet18 example with TorchDynamo
This is a minor variation on our other resnet18 examples swapping in
TorchDynamo.

We replicate the refbackend_torchdynamo_backend out of the e2e test
config to avoid making that appear like a public API.

Also, some minor cleanups to TorchDynamoTestConfig.
2022-12-07 09:25:27 -08:00
Sean Silva c956c39c86 [cleanup] Remove disabled e2e test
This test has been disabled a long time, and since RefBackend is so slow
we don't want to add this unnecessarily. I believe it is covered by
downstream testing such as the Shark Tank.
2022-12-07 06:36:48 -08:00
Vivek Khandelwal 3e4bb2bd8e [MLIR][TORCH] Add E2E support for randn and randn.generator op
Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2022-12-06 22:41:24 +05:30
Sean Silva 485c18bb2f [torchdynamo] Add "lockstep" numerical accuracy debugger.
Thanks to TorchDynamo's great layering and design, this is only about
100 lines of code for a basic lockstep debugger.

This should allow us to deprecate eager_mode, since AFAIK the only
interesting use case that it was really supporting is for downstream users to
write lockstep debuggers.

NOTE: The exact reporting and interface here is subject to change. Please
try it out and provide feedback (or patches :) ).
- make_fx should not drop source locations: https://github.com/pytorch/pytorch/issues/90276
- Report tensors better (huge tensors should be summarized)
- Maybe don't abort, but just warn?
- Allow customizing atol/rtol.
- How best to print the failing node? And include surrounding graph
context?
2022-12-06 07:57:45 -08:00
Vivek Khandelwal ef39b9ebb4 build: manually update PyTorch version
Set PyTorch and TorchVision version to nightly release 2022-12-05.

Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2022-12-05 22:44:32 +05:30
Vivek Khandelwal f416953600 [MLIR][TORCH] Add TorchConversionToMLProgram and MLProgramBufferize pass
This commit changes the `InsertRngGlobalsPass` to `TorchConversionToMLProgram`
pass. This commit also adds the `MLProgramBufferize` pass for the
bufferization of ml_program dialect ops to run on refbackend.

Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2022-12-02 13:20:46 +05:30
Sean Silva 88db99946b [torchdynamo] Use decompositions to support a few ops 2022-12-01 11:25:20 -08:00
Ramiro Leal-Cavazos b4b92c990e
Replace LCG algorithm with squares64 algorithm in AtenUniformOp (#1633)
This commit replaces the LCG algorithm that was being used by the
`TorchToLinalg` lowering of `AtenUniformOp` to generate random numbers
with the `squares64` algorithm, for the LCG algorithm was producing
tensors that were highly correlated with one another.

Squares64 algorithm: https://arxiv.org/abs/2004.06278

Closes https://github.com/llvm/torch-mlir/issues/1608
2022-12-01 08:30:10 -08:00
Ramiro Leal-Cavazos 0983a7f93a
Fix modulus calculation in LCG algorithm of refbackend (#1658)
The current implementation sets the `nextSeed` value to `temp & 127`,
which is wrong. The last step of the LCG algorithm for the multiplier
and increment chosen should be `temp % 2^{64} = temp & (1 <<
63)`. However, because we are dealing with i64 values, the modulus
operation happens automatically, so it is not needed.

See Donald Knuth's values for LCG here:
https://en.wikipedia.org/wiki/Linear_congruential_generator
2022-11-30 08:46:52 -08:00
Abhishek Varma c27c1791f1 [MLIR][TORCH] Add e2e support for `aten.amax` op
-- This commit adds e2e support for `atend.amax` op.

Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
2022-11-30 17:54:37 +05:30
Tanyo Kwok bbcdb38d99
Revert "Decompose torch.slice_scatter (#1622)" (#1659)
This reverts commit f3f2f10030.
2022-11-30 12:47:13 +08:00
Daniel Ellis e2de20575f
Automatically strip overloads for FX-based models. 2022-11-29 22:19:09 -05:00