llvm
/
torch-mlir
mirror of https://github.com/llvm/torch-mlir
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
2,938 Commits
47 Branches
1195 Tags
77 MiB
Commit Graph

756 Commits (7cd3368b206bbcfb9cf272bfa7e532e60f574fc8)

Author SHA1 Message Date
Yuanqiang Liu 689efc8917
[Torch] fix toBuiltinTensor() (#3415) 
* Let `toBuiltinTensor()` reflects the original dtype of
`!torch.vtensor`.
* Backend handles dtype conversion themselves.
 2024-06-08 09:36:32 +08:00
Rob Suderman 75af64fc12
[torch] Add support for f8 types for linalg conversion (#3436) 
Linalg conversion requires mapping for f8 types
 2024-06-07 13:59:38 -07:00
Sambhav Jain d0a818a03e
Representing Symbolic Shape Expressions in Torch Dialect (#3372) 
Torch Dialect with symbolic shape expressions:
```ll
module {                                                                                                                                                                                                     
  func.func @main(%arg0: !torch.vtensor<[?,?,3],f32>, %arg1: !torch.vtensor<[?,?,3],f32>) -> !torch.vtensor<[?,?,3],f32> {                                                                                   
    %0 = torch.symbolic_int "s0" {min_val = 5, max_val = 10} : !torch.int                                                                                                                                    
    %1 = torch.symbolic_int "s1" {min_val = 0, max_val = 100} : !torch.int                                                                                                                                   
    %2 = torch.symbolic_int "s3" {min_val = 0, max_val = 50} : !torch.int                                                                                                                                    
    
    torch.bind_symbolic_shape %arg0, [%0, %1], #affine_map<()[s0, s1] -> (s0, s1, 3)> : !torch.vtensor<[?,?,3],f32>                                                                                          
    torch.bind_symbolic_shape %arg1, [%0, %2], #affine_map<()[s0, s1] -> (s0, s1, 3)> : !torch.vtensor<[?,?,3],f32>                                                                                          
    
    %3 = torch.aten.tanh %arg0 : !torch.vtensor<[?,?,3],f32> -> !torch.vtensor<[?,?,3],f32>                                                                                                                  
    torch.bind_symbolic_shape %3, [%0, %1], #affine_map<()[s0, s1] -> (s0, s1, 3)> : !torch.vtensor<[?,?,3],f32>                                                                                             
    
    %4 = torch.aten.sigmoid %arg1 : !torch.vtensor<[?,?,3],f32> -> !torch.vtensor<[?,?,3],f32>                                                                                                               
    torch.bind_symbolic_shape %4, [%0, %2], #affine_map<()[s0, s1] -> (s0, s1, 3)> : !torch.vtensor<[?,?,3],f32>                                                                                             
    
    %5 = torch.prim.ListConstruct %3, %3, %4 : (!torch.vtensor<[?,?,3],f32>, !torch.vtensor<[?,?,3],f32>, !torch.vtensor<[?,?,3],f32>) -> !torch.list<vtensor>                                               
    %int1 = torch.constant.int 1                                                                                                                                                                             
    %6 = torch.aten.cat %5, %int1 : !torch.list<vtensor>, !torch.int -> !torch.vtensor<[?,?,3],f32>                                                                                                          
    torch.bind_symbolic_shape %6, [%0, %1, %2], #affine_map<()[s0, s1, s2] -> (s0, s1 * 2 + s2, 3)> : !torch.vtensor<[?,?,3],f32>                                                                            
    
    return %6 : !torch.vtensor<[?,?,3],f32>                                                                                                                                                                  
  }                                                                                                                                                                                                          
}              
```

For reference, this is the TorchDynamo exported program with symbolic
shape expressions that the above Torch dialect program is imported from:
```py
ExportedProgram:                                                                                                                                                                                             
    class GraphModule(torch.nn.Module):                                                                                                                                                                      
        def forward(self, x: "f32[s0, s1, 3]", y: "f32[s0, s3, 3]"):                                                                                                                                         
            # File: /home/sambhav.jain/workspaces/cruise/src/3p/torch-mlir/test/python/fx_importer/symbolic_shape_expr_test.py:31 in forward, code: a = torch.tanh(x)                                        
            tanh: "f32[s0, s1, 3]" = torch.ops.aten.tanh.default(x);  x = None                                                                                                                               
                                                                                                                                                                                                             
            # File: /home/sambhav.jain/workspaces/cruise/src/3p/torch-mlir/test/python/fx_importer/symbolic_shape_expr_test.py:32 in forward, code: b = torch.sigmoid(y)                                     
            sigmoid: "f32[s0, s3, 3]" = torch.ops.aten.sigmoid.default(y);  y = None                                                                                                                         
                                                                                                                                                                                                             
            # File: /home/sambhav.jain/workspaces/cruise/src/3p/torch-mlir/test/python/fx_importer/symbolic_shape_expr_test.py:33 in forward, code: return torch.cat((a, a, b), dim=1)                       
            cat: "f32[s0, 2*s1 + s3, 3]" = torch.ops.aten.cat.default([tanh, tanh, sigmoid], 1);  tanh = sigmoid = None                                                                                      
            return (cat,)                                                                                                                                                                                    
                                                                                                                                                                                                             
Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='x'), target=None, persistent=None), InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='y'), target=None, persistent=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='cat'), target=None)])                                               
Range constraints: {s0: ValueRanges(lower=5, upper=10, is_bool=False), s1: ValueRanges(lower=0, upper=100, is_bool=False), s3: ValueRanges(lower=0, upper=50, is_bool=False)} 
```

Huge credit to @stellaraccident for the inputs that helped evaluate the
various design options and arrive at the representation of choice.


- [x] Op definitions for symbolic_int and bind_symbolic_shape ops
- [x] fx_importer updates to import range constraints + create
symbolic_int ops
- [x] fx_importer changes for AffineMapAttr building + adding
bind_symbolic_shape ops
- [x] custom printer/parser for inlined AffineMap expressions in mlir
assembly
- [x] Dialect lit test
- [x] fx_importer python lit tests
- [ ] Cleanup pass to remove these ops (can add in a follow-on)
 2024-06-07 04:04:03 -07:00
Vivek Khandelwal 72837fbb3d
build: manually update PyTorch version (#3340) 
Set PyTorch and TorchVision version to nightly release 2024-05-14.

Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
 2024-06-06 22:23:40 +05:30
Vivek Khandelwal 661be2d5b0
[MLIR][Torch] Add TorchToLinalg lowering for AtenAvgPool3dOp (#3030) 
This commit also fixes the average pool op' test failing for
OnnxToLinalg lowering.

Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
 2024-06-04 22:12:34 +05:30
zjgarvey 8995c90879
[TorchToLinalg] add support for quantized group conv (#3341) 
This addresses 7 of the model failures I'm seeing in the test suite. See
[Shark-Turbine issue
#566](https://github.com/nod-ai/SHARK-Turbine/issues/566).

Need the op ```linalg.conv_2d_ngchw_gfchw_q``` to be added upstream
before merging this. See [llvm-project PR #92136
](https://github.com/llvm/llvm-project/pull/92136).

A small additional expansion to operand quantization is included in this
patch to address a model failure that occurs when unblocking the
quantized group convolutions in one of these onnx models.
 2024-06-03 21:57:44 +05:30
Vivek Khandelwal 6382dbbcc0
[ONNX] Add OnnxToTorch lowering for SpaceToDepth op (#3393) 
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
 2024-06-03 20:29:39 +05:30
Xinyu Yang 285b087a5d
[Torch] Emit rrelu and decompose it (#3250) 
as title
 2024-06-03 19:25:52 +08:00
Xinyu Yang 267052df2a
[Torch] decompose AtenLerpTensorOp (#3251) 
as title
 2024-06-03 15:25:09 +08:00
Xinyu Yang 23b53050de
[Torch]Support conv_transpose1d and conv_transpose3d (#3286) 
1. Support conv_transpose1d and conv_transpose3d
2. Fix bugs of convertTransposedConv func in
lib/Conversion/TorchToStablehlo/Linear.cpp
 2024-06-03 15:11:12 +08:00
Rob Suderman afca88a058
[NFC] Change to *cast instead of .*cast variants (#3405) 
Member casts have been deprecated. Changing over a bunch of the member
cast calls to the global templated variants to remove deprecation
warnings.
 2024-05-30 23:45:13 -07:00
Yuanqiang Liu 4e05e2cd1e
[Torch] support recompose of aten.split.with_sizes and aten.tensor_sp… (#3401) 
…lit.sections

* support recompose to aten.split.with_sizes and
aten.tensor_split.sections
* fix recompose of aten.chunk
 2024-05-31 09:56:47 +08:00
zjgarvey 074098d20c
Modifies onnx resize lowering to fix numerical issues (#3381) 
Updates:

- some unsupported modes are now going to report a match failure for
unsupported coordinate transformation modes.
- fixes a bug that was introduced in the last patch for resize (my
bad...)
- uses actual x and y coordinates for computing weights in bilinear
interpolation (rather than eps modified values)
- slightly simplifies the bilinear interpolation payload for readability
and performance
- passes coordinate transformation mode information from an onnx.Resize
op to the mode string for the aten._interpolate op. This allows us to
perform custom logic in the torch->linalg lowering to support
onnx.Resize options without losing the default behaviors of the
interpolate op.
 2024-05-30 20:34:37 -04:00
penguin_wwy 1f544c37d0
[NFC] Remove unused header files (#3386) 2024-05-30 14:30:36 +08:00
Yuanqiang Liu e0a5adb1db
[Torch] fix aten.linear's decomposition (#3391) 
* support aten.linear with more rank.
 2024-05-27 15:49:50 +08:00
Yuanqiang Liu 5bb1a65ec9
[Stablehlo] refactor reduction lowering and support aten.amin (#3383) 
* implement detailed lowering template pattern
`ConvertAtenReduceAllDimsOp` and `ConvertAtenReduceKeepDimOp`
* support `aten.amin`'s lowering.
 2024-05-23 20:40:20 +08:00
Angel Zhang 2e194e13d6
[Torch] Fix bugs for `Torch::AtenOneHotOp` (#3350) 
This PR fixes the bugs for `Torch::AtenOneHotOp` by:

1) Using `Torch::kUnknownSize` as the default value for `numClasses` in
   the pattern matching stage in `DecomposeAtenOneHotOp`
2) Adding `AtenIntScalarOp` to the patterns in `TorchToArith`
3) Handling both `int` and `float` types for `off` and `on` values in
`TorchOnnxToTorch` conversion

It also includes:

1) A new test in `TorchToArith/basic.mlir`, for `torch.aten.Int.Scalar`,
and
2) A new test in `decompose-complex-ops.mlir`, for `torch.aten.one_hot`

**Dependencies**

This PR is dependent on #3334.
 2024-05-22 17:19:08 +00:00
Xinyu Yang 4d7cdba4bf
[Torch] eliminate "getWithLeastStaticInformation" in DecomposeAtenTriuOp (#3330) 
I am trying to eliminate 'getWithLeastStaticInformation' in
DecomposeAtenTriuOp. Could you provide me with some suggestions?
@qingyunqu @zjgarvey 
See issue https://github.com/llvm/torch-mlir/issues/3312
 2024-05-22 23:16:57 +08:00
Sambhav Jain 6e485574e5
[Pipeline] Use dedicated simplification pipeline for TorchDynamo frontend (#3376) 
Discord Thread:
https://discord.com/channels/636084430946959380/1238330633328005243

## Context: 

[This](https://github.com/llvm/torch-mlir/blob/main/python/torch_mlir/fx.py#L61)
was updated to support e2e tests for the TorchDynamo frontend in
Torch-MLIR, where we run FX decompositions and import the FX IR to
generate Torch dialect, followed by
`torch-function-to-torch-backend-pipeline`, skipping only the shape/type
refinement for now. However, we should be able to skip many of the torch
simplification passes, as depicted in the [frontend
roadmap](https://github.com/llvm/torch-mlir/blob/main/docs/images/roadmap_frontend.png).

Based on IREE's TorchDynamo
[pipeline](https://github.com/iree-org/iree/blob/main/compiler/plugins/input/Torch/InputConversion/Passes.cpp#L29),
the only two passes we seem to require are: `ReduceOpVariantsPass` and
`DecomposeComplexOpsPass`. This is inline with our findings as well
based on initial exploration.

This PR creates a dedicated frontend simplification pipeline for
TorchDynamo / FX Importer which calls only `ReduceOpVariantsPass` and
`DecomposeComplexOpsPass`. We rely on the e2e fx_importer tests to
ensure we're not regressing by removing many of the passes that were
historically needed for TorchScript.

One notable change here is that we do not call the
`LowerToBackendContractPass` anymore, which used to call
`TorchSimplificationPipeline` iteratively until VerifyBackendContract
was clean. Some of this was required for the shape/type refinement to
converge, which seems a non-issue for Dynamo frontend. Do we anticipate
this (the iterative invocation of TorchSimplificationPipeline followed
by VerifyBackendContract) to be worth retaining in the Dynamo frontend
pipeline? If so, I can make those changes, PLMK.
 2024-05-22 05:23:18 -07:00
Yuanqiang Liu 8814d0ae64
[Torch] emit aten.dot and canonicalize it to aten.matmul (#3361) 
* canonicalize `aten.dot` to `aten.matmul`
 2024-05-18 22:45:14 +08:00
Xinyu Yang 7faba75696
[Torch] Decompose AtenMaskedScatterOp (#3353) 
Co-authored-by: Yuanqiang Liu <liuyuanqiang.yqliu@bytedance.com>
 2024-05-16 15:27:25 +08:00
Xinyu Yang a9edefb3cf
[Torch] Fix AtenSliceTensorOp::fold (#3345) 2024-05-16 11:42:43 +08:00
Peiming Liu ccb772cd0f
[sparse] propagate sparsity properly when decompose torch operations. (#3318) 2024-05-15 10:09:27 -07:00
Aaron St George ba32b9cee7
Don't fold `aten.clone` if result isn't same type as input (#3347) 
Similar to https://github.com/llvm/torch-mlir/pull/2824, we were seeing
some assertion failures after the addition checks around folders were
tightened up in LLVM: https://github.com/llvm/llvm-project/pull/75887 .
This PR essentially moves the logic that used to be applied at the LLVM
level into the folder, which seems to be the suggested fix.
 2024-05-16 00:07:45 +08:00
Xinyu Yang 6b95dd461d
[Torch] Fix PrimNumToTensorScalarOp::fold (#3339) 
In constant folding progress, a new constant op will be created
according to the origin op's result type.

See the code in TorchDialect.cpp.

```cpp
Operation *TorchDialect::materializeConstant(OpBuilder &builder,
                                             Attribute value, Type type,
                                             Location loc) {
  if (auto integerType = dyn_cast<Torch::IntType>(type))
    return builder.create<Torch::ConstantIntOp>(loc, cast<IntegerAttr>(value));

  if (auto floatType = dyn_cast<Torch::FloatType>(type))
    return builder.create<Torch::ConstantFloatOp>(loc, cast<FloatAttr>(value));

  if (auto numberType = dyn_cast<Torch::NumberType>(type)) {
    if (auto floatValue = dyn_cast<mlir::FloatAttr>(value)) {
      return builder.create<Torch::ConstantNumberOp>(loc, floatValue);
    } else if (auto intValue = dyn_cast<mlir::IntegerAttr>(value)) {
      return builder.create<Torch::ConstantNumberOp>(loc, intValue);
    }
  }

  if (isa<Torch::BoolType>(type)) {
    return builder.create<Torch::ConstantBoolOp>(loc, cast<IntegerAttr>(value));
  }

  if (isa<Torch::NoneType>(type))
    return builder.create<ConstantNoneOp>(loc);

  if (auto stringAttr = dyn_cast<StringAttr>(value))
    return builder.create<ConstantStrOp>(loc, stringAttr);

  if (auto elementsAttr = dyn_cast<ElementsAttr>(value)) {
    // Only !torch.vtensor can be constant folded. !torch.tensor has
    // non-trivial aliasing semantics which prevent deduplicating it.
    assert(isa<ValueTensorType>(type) && "should be a vtensor type!");
    return builder.create<ValueTensorLiteralOp>(loc, elementsAttr);
  }

  return nullptr;
}
```
So when the op has a tensor result type, it must be "ValueTensorType"
due to the **assert** statement. However, many fold methods in
TorchOps.cpp only have a judgment of "BaseTensorType".
 2024-05-15 20:54:19 +08:00
zjgarvey 911e723581
Expands Q Commuting Ops (#3332) 
After running the model tests in SHARK-TestSuite, I noticed a few model
failures due to half-fusion.

Notably, RDN_pytorch_vaiq_int8 had a depth=5 convolution chain with
multiple AtenViewOp's.
 2024-05-13 11:01:53 -07:00
zjgarvey 75d1d72059
Generalize Operand Quantization in FuseQuantizeOps (#3327) 
This change enables more customization with operand quantization, and
generalizes the patterns QuantizeOperands and QuantizeTransposeOperands
to QuantizeOperandsPastCommutingOps.

This allows for passing quantization through operations which are
functionally unaffected by quantization, such as view-like ops. The
purpose of this change is to address a myriad of quantization issues
seen in quantized onnx models that have some reshape-like operations
sandwiched in between a dequant and something like a matmul (whose other
operand is immediately quantizable).
 2024-05-12 20:49:59 -07:00
NeverRaR 1d4859699b
MaxPool1d lowering to linalg (#3295) 
Co-authored-by: root <root@i32b01216.sqa.eu95>
 2024-05-10 22:05:26 +05:30
penguin_wwy 64b59c7fc3
[FxImporter] Eliminate the dependency on the refinement pass (#3309) 2024-05-10 02:44:36 +08:00
aldesilv ec6d7aa5d2
OnnxToTorch lowering resize op (#3013) 
https://github.com/nod-ai/SHARK-Turbine/issues/358
adds a lowering from onnx to linalg for bilinear and nearest resize with
support for using scales or sizes to get resize shape. uses coordinate
transform half pixel for bilinear mode and asymmetrical for nearest
mode. See
https://github.com/onnx/onnx/blob/main/docs/Operators.md#Resize. Added
two passes -- one for bilinear and the other for nearest.
 2024-05-08 21:35:03 +00:00
Jiawei Wu 346a536c9f
[Torch Dialect] decompose all index_put-like op to aten.index_put.hacked_twin for stricter semantics (#3071) 
This PR decomposes all index_put-like op to aten.index_put.hacked_twin for stricter semantics, i.e., no None index in indices argument.
 2024-05-08 22:44:57 +08:00
Xinyu Yang abef114c0c
[torch] emit aten.Softshrink and aten.Hardshrink (#3248) 
as title
 2024-05-08 15:20:45 +08:00
Vivek Khandelwal e60160d793
Revert "Decompose AtenNonzeroOp" (#3289) 
Reverts llvm/torch-mlir#3281
 2024-05-06 09:52:04 -07:00
Xida Ren (Cedar) 1af00e6040
Decompose AtenNonzeroOp (#3281) 
This fixes some onnx lit tests not lowering to linalg in
https://github.com/nod-ai/SHARK-Turbine/issues/450
 2024-05-05 21:59:25 +08:00
Ze Zhang 11cd7cd9e7
Folder and Canonicalizer for PrimsConvertElementTypeOp and AtenMaxPool2dWithIndicesOp (#3272) 
While playing with TorchDynamo on ResNet18. I notice following issues:

- `prims.convert_element_type` can’t be canonicalized even if the input
and the output share the same type

- `aten.max_pool2d_with_indices` is always used instead of
`aten.max_pool2d`, even if the second returned output (indices) has no
user

This PR fixes above issues by adding a folder to the
PrimsConvertElementTypeOp and a canonicalizer to the
AtenMaxPool2dWithIndicesOp


Lit test:

`cmake --build build --target check-torch-mlir-all`

---------

Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
 2024-05-02 00:03:41 -07:00
Xida Ren (Cedar) 315dc6c3e3
[torch] `aten.eye` should use dynamic dims when no static dims are available (#3202) 
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
 2024-04-30 17:41:03 +00:00
Xinyu Yang f32ada993d
[Stablehlo] Improve the lowering of pool op in stablehlo (#3259) 
1. Handle case stride == None
2. add avgpool3d maxpool1d  maxpool3d lowering
 2024-05-01 00:06:13 +08:00
Vivek Khandelwal b1e2241479
[ONNX] Fix Onnx.Selu lowering and canonicalizer for IntImplicit op (#3221) 
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
 2024-04-29 04:00:01 +00:00
Yuanqiang Liu aed2cf3351
[Torch] emit aten.__contains__.str_list and add folder (#3249) 2024-04-29 10:51:17 +08:00
Xinyu Yang 5684dc0441
[Torch] emit aten.celu and decompose it (#3247) 
CELU(x)=max(0,x)+min(0,α∗(exp(x/α)−1))
 2024-04-28 17:23:40 +08:00
Yuanqiang Liu 46c0f3cad0
[Torch] emit aten.log_sigmoid and decompose it to log(sigmoid) (#3246) 2024-04-28 11:47:43 +08:00
Stella Laurenzo 5d4b803914 [NFC reformat] Run pre-commit on all files and format misc. 
This is part 1 of ~3, formatting all miscellaneous text files and CPP files matched by a first run of pre-commit. These tend to be low change-traffic and are likely not disruptive.

Subsequent patches will format Python files and remaining CPP files.
 2024-04-27 14:08:09 -07:00
penguin_wwy 6679728c56
Fix deprecated uses of cast/dyn_cast/dyn_cast_or_null/isa (#3243) 
Like #3130, gradually replace the deprecated code

https://github.com/llvm/mlir-www/blob/main/website/content/deprecation/_index.md#deprecated
 2024-04-27 14:00:56 -07:00
Yuanqiang Liu f173a06fa7
[Torch] emit aten.ne.str and add folder (#3242) 2024-04-28 00:58:50 +08:00
Yuanqiang Liu 634a796933
[Torch] fold aten.log (#3223) 2024-04-26 10:10:02 +08:00
Yuanqiang Liu b0ba3def93
[Torch] support AtenScalarImplicitOp canonicalize with float (#3231) 2024-04-26 02:36:13 +08:00
Yuanqiang Liu fab2696489
[Torch] support aten.trunc (#3219) 
decompose `trunc(x)` to `sign(x) * floor(abs(x))`
 2024-04-24 14:32:33 +08:00
Xinyu Yang 4da3d714cc
[Torch] Support AtenProdOp on linalg and stablehlo (#3215) 2024-04-24 11:14:04 +08:00
zjgarvey a8ba865fca
[torch] Adds Quantization Support for `aten.relu` (#3177) 
A choice was made to quantize the return type of Relu with a scale and
zero point copied from the input's quantization scheme. With this
choice, the torch-to-linalg conversion of quantized Relu essentially
computes max(input, zeroPoint) in the elementwise payload.
 2024-04-23 11:01:36 -07:00
Yuanqiang Liu db3842f2e8
[Stablehlo] support lowering sinh & cosh to stablehlo (#3213) 2024-04-23 19:54:58 +08:00