llvm
/
torch-mlir
mirror of https://github.com/llvm/torch-mlir
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
torch-mlir/test/Dialect/Torch/canonicalize.mlir

1773 lines
95 KiB
MLIR
Raw Normal View History

[torch-mlir earthmoving (1/N)] C/C++ code movement. This creates the `external/torch-mlir` directory as an LLVM_EXTERNAL_PROJECTS-compatible project (analogous to `iree-dialects`) and completes movement/rename of all pure MLIR C/C++ compiler code into there. The next step will be to move all the Python code / code that links/includes PyTorch C++ code (which currently lives in `frontends/pytorch`) into a subdirectory here. I call this "earthmoving" because it is mostly mechanical changes and renames. As a quick summary (we can change this down the road easily) - C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch` - CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet` - preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_` - CMake `NPCOMPFoo -> TorchMLIRFoo` The goal of this is to create a standalone project creating a center of mass for entry into the MLIR ecosystem from PyTorch, suitable in scope for eventual inclusion/ownership in PyTorch. The idea is that `external/torch-mlir` will some day be pulled out into its own repository, and then npcomp will simply pull it in as a submodule. Layering-wise, what lives in `torch-mlir` lowers code from PyTorch (currently TorchScript, but TorchFX or pytorch/xla-style tracing are possible extensions) down to what we have been calling the "Torch backend contract" which is cleaned up IR (inlining, simplifcation, conversion to value tensors, ...) entirely in the `torch` dialect. This is the branching off point for further lowering, of which npcomp takes one opinion (outside `torch-mlir` of course!), namely the `TorchConversion` dialect/transforms which lower to IR suitable for IREE and other linalg-on-tensors based lower-level compilers. Summary of changes: - move `{include,lib,test}/Dialect/Torch` into `torch-mlir` - move relevant parts of CAPI into `torch-mlir`. - leave a few things related to the `torch-mlir` Python build commented out, which should be resolved in a subsequent change.
2021-09-10 03:24:10 +08:00
// RUN: torch-mlir-opt %s -canonicalize | FileCheck %s
Add recognition/folder/lowering for aten::__is__, aten::ne.int, and aten::dim Interestingly, TorchScript has its own op (`torch::jit::Operator`) registry separate from the dispatcher (it is a superset of the dispatcher). This is where the "prim" ops and some "aten" ops (that should probably be renamed to "prim") live. In particular, `aten::__is__` is in that latter category of "aten but really prim". This registry is also the source of truth for what the TorchScript interpreter calls into when it executes. The bulk of the "not part of the dispatcher" ops live in https://github.com/pytorch/pytorch/blob/09feb5f579d40b550559360870907963e20c6954/torch/csrc/jit/runtime/register_prim_ops.cpp#L82 And the registry itself lives in: https://github.com/pytorch/pytorch/blob/09feb5f579d40b550559360870907963e20c6954/torch/csrc/jit/runtime/operator.cpp#L196 This fold further reduces the IR of ResNet by folding away some more not-taken branches. These not-taken branches in ResNet require first-class handling of the list type which we don't yet have on any backend.
2021-04-28 06:15:50 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__range_length$fold() -> (!torch.int, !torch.int, !torch.int, !torch.int) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[INT1:.*]] = torch.constant.int 1
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[INT3:.*]] = torch.constant.int 3
Bump LLVM (#958)
2022-06-23 11:23:46 +08:00
// CHECK: %[[INTM1:.*]] = torch.constant.int -1
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[NEG_STEP:.*]] = torch.aten.__range_length %[[INT1]], %[[INT3]], %[[INTM1]] : !torch.int, !torch.int, !torch.int -> !torch.int
// CHECK: return %[[INT2]], %[[INT2]], %[[INT1]], %[[NEG_STEP]] : !torch.int, !torch.int, !torch.int, !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__range_length$fold() -> (!torch.int, !torch.int, !torch.int, !torch.int) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int3 = torch.constant.int 3
%int4 = torch.constant.int 4
%int2 = torch.constant.int 2
%int1 = torch.constant.int 1
%int0 = torch.constant.int 0
%int-1 = torch.constant.int -1
%0 = torch.aten.__range_length %int0, %int4, %int2 : !torch.int, !torch.int, !torch.int -> !torch.int
%1 = torch.aten.__range_length %int1, %int4, %int2 : !torch.int, !torch.int, !torch.int -> !torch.int
%2 = torch.aten.__range_length %int1, %int3, %int2 : !torch.int, !torch.int, !torch.int -> !torch.int
%3 = torch.aten.__range_length %int1, %int3, %int-1 : !torch.int, !torch.int, !torch.int -> !torch.int
return %0, %1, %2, %3 : !torch.int, !torch.int, !torch.int, !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__is__
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__is__(%arg0: !torch.list<int>, %arg1: !torch.none) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.__is__ %arg0, %arg1 : !torch.list<int>, !torch.none -> !torch.bool
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
return %0 : !torch.bool
Add recognition/folder/lowering for aten::__is__, aten::ne.int, and aten::dim Interestingly, TorchScript has its own op (`torch::jit::Operator`) registry separate from the dispatcher (it is a superset of the dispatcher). This is where the "prim" ops and some "aten" ops (that should probably be renamed to "prim") live. In particular, `aten::__is__` is in that latter category of "aten but really prim". This registry is also the source of truth for what the TorchScript interpreter calls into when it executes. The bulk of the "not part of the dispatcher" ops live in https://github.com/pytorch/pytorch/blob/09feb5f579d40b550559360870907963e20c6954/torch/csrc/jit/runtime/register_prim_ops.cpp#L82 And the registry itself lives in: https://github.com/pytorch/pytorch/blob/09feb5f579d40b550559360870907963e20c6954/torch/csrc/jit/runtime/operator.cpp#L196 This fold further reduces the IR of ResNet by folding away some more not-taken branches. These not-taken branches in ResNet require first-class handling of the list type which we don't yet have on any backend.
2021-04-28 06:15:50 +08:00
}
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__is__$derefine_is_none
Fold __is__ and unchecked_cast of derefine The added e2e maxpool testcase from #545 was not getting a static shape due to an unfolded prim.If when RefineTypes was called. This was because of unfolded torch.iaten.__is__ and torch.prim.unchecked_cast operators with torch.derefine operands.
2022-01-28 21:35:40 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__is__$derefine_is_none(%arg0: !torch.list<int>, %arg1: !torch.none) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.derefine %arg0 : !torch.list<int> to !torch.optional<list<int>>
%1 = torch.aten.__is__ %0, %arg1 : !torch.optional<list<int>>, !torch.none -> !torch.bool
Fold __is__ and unchecked_cast of derefine The added e2e maxpool testcase from #545 was not getting a static shape due to an unfolded prim.If when RefineTypes was called. This was because of unfolded torch.iaten.__is__ and torch.prim.unchecked_cast operators with torch.derefine operands.
2022-01-28 21:35:40 +08:00
return %1 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__is__$none_is_none
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__is__$none_is_none(%arg0: !torch.none, %arg1: !torch.none) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%0 = torch.aten.__is__ %arg0, %arg1 : !torch.none, !torch.none -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__is__$is_none$derefine(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %{{.*}}: !torch.vtensor) -> !torch.bool {
// CHECK: %[[RESULT:.*]] = torch.constant.bool false
// CHECK: return %[[RESULT]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__is__$is_none$derefine(%arg0: !torch.vtensor) -> !torch.bool {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%none = torch.constant.none
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.derefine %arg0 : !torch.vtensor to !torch.optional<vtensor>
%1 = torch.aten.__is__ %0, %none : !torch.optional<vtensor>, !torch.none -> !torch.bool
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__isnot__
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__isnot__(%arg0: !torch.list<int>, %arg1: !torch.none) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.__isnot__ %arg0, %arg1 : !torch.list<int>, !torch.none -> !torch.bool
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__isnot__$none_isnot_none
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__isnot__$none_isnot_none(%arg0: !torch.none, %arg1: !torch.none) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%0 = torch.aten.__isnot__ %arg0, %arg1 : !torch.none, !torch.none -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.bool() -> !torch.bool {
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.bool() -> !torch.bool {
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
%a = torch.constant.bool true
%b = torch.constant.bool false
%0 = torch.aten.ne.bool %a, %b: !torch.bool, !torch.bool -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.bool$same_operand(
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
// CHECK-SAME: %[[ARG0:.*]]: !torch.bool) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.bool$same_operand(%arg0: !torch.bool) -> !torch.bool {
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
%0 = torch.aten.ne.bool %arg0, %arg0: !torch.bool, !torch.bool -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.bool$different_operand(
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
// CHECK-SAME: %[[ARG0:.*]]: !torch.bool) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: %[[RET:.*]] = torch.aten.ne.bool %[[ARG0]], %[[FALSE]] : !torch.bool, !torch.bool -> !torch.bool
// CHECK: return %[[RET]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.bool$different_operand(%a: !torch.bool) -> !torch.bool {
Add aten.ne.bool to make CI pass
2021-10-21 23:50:01 +08:00
%b = torch.constant.bool false
%0 = torch.aten.ne.bool %a, %b: !torch.bool, !torch.bool -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.size$canonicalize_to_list(
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[2,3],f32>) -> !torch.list<int> {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: %[[C2:.*]] = torch.constant.int 2
// CHECK: %[[C3:.*]] = torch.constant.int 3
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK: %[[LIST:.*]] = torch.prim.ListConstruct %[[C2]], %[[C3]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[LIST]] : !torch.list<int>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.size$canonicalize_to_list(%arg0: !torch.vtensor<[2,3],f32>) -> !torch.list<int> {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.size %arg0 : !torch.vtensor<[2,3],f32> -> !torch.list<int>
return %0 : !torch.list<int>
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
}
// One size unknown, so cannot canonicalize.
// TODO: For unknown sizes, insert the equivalent of a "dim" op.
// Then this will only require static rank.
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.size$unknown_size(
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[?,3],f32>) -> !torch.list<int> {
// CHECK: %[[SIZE:.*]] = torch.aten.size %[[ARG]] : !torch.vtensor<[?,3],f32> -> !torch.list<int>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.size$unknown_size(%arg0: !torch.vtensor<[?,3],f32>) -> !torch.list<int> {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.size %arg0 : !torch.vtensor<[?,3],f32> -> !torch.list<int>
return %0 : !torch.list<int>
Add aten::len.t, aten::size, and aten::gt.int primitive ops Also add some canonicalizations that finally reduce ResNet down to a single block.
2021-04-30 06:38:11 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.int$same_operand(
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> !torch.bool {
// CHECK-NEXT: %[[FALSE:.*]] = torch.constant.bool false
// CHECK-NEXT: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.int$same_operand(%arg0: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%0 = torch.aten.ne.int %arg0, %arg0 : !torch.int, !torch.int -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int4_0 = torch.constant.int 4
%2 = torch.aten.ne.int %int4, %int4_0 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ne.int$different_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ne.int$different_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%2 = torch.aten.ne.int %int4, %int5 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$different_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$different_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%2 = torch.aten.eq.int %int4, %int5 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$same_operand(
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> !torch.bool {
// CHECK-NEXT: %[[F:.*]] = torch.constant.bool true
// CHECK-NEXT: return %[[F]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$same_operand(%arg0: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%0 = torch.aten.eq.int %arg0, %arg0 : !torch.int, !torch.int -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int4_0 = torch.constant.int 4
%2 = torch.aten.eq.int %int4, %int4_0 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$of_size.int(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$of_size.int(%arg0: !torch.tensor) -> !torch.bool {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%int0 = torch.constant.int 0
%0 = torch.aten.size.int %arg0, %int0 : !torch.tensor, !torch.int -> !torch.int
%1 = torch.aten.eq.int %0, %int-1 : !torch.int, !torch.int -> !torch.bool
return %1 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$of_size.int_lhs_constant(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$of_size.int_lhs_constant(%arg0: !torch.tensor) -> !torch.bool {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%int0 = torch.constant.int 0
%0 = torch.aten.size.int %arg0, %int0 : !torch.tensor, !torch.int -> !torch.int
%1 = torch.aten.eq.int %int-1, %0 : !torch.int, !torch.int -> !torch.bool
return %1 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int$no_change_minus1(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.bool {
// CHECK: %[[CM1:.*]] = torch.constant.int -1
// CHECK: %[[RESULT:.*]] = torch.aten.eq.int %[[CM1]], %[[ARG]] : !torch.int, !torch.int -> !torch.bool
// CHECK: return %[[RESULT]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int$no_change_minus1(%arg0: !torch.int) -> !torch.bool {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%1 = torch.aten.eq.int %int-1, %arg0 : !torch.int, !torch.int -> !torch.bool
return %1 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.int$evaluate_to_true() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.int$evaluate_to_true() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%2 = torch.aten.lt.int %int4, %int5 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.int$same_operand(
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.int$same_operand(%arg0: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%2 = torch.aten.lt.int %arg0, %arg0: !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int4_0 = torch.constant.int 4
%2 = torch.aten.lt.int %int4, %int4_0 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.le.int$evaluate_to_true() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.le.int$evaluate_to_true() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%2 = torch.aten.le.int %int4, %int5 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.le.int$same_operand(
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> !torch.bool {
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.le.int$same_operand(%arg0: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%2 = torch.aten.le.int %arg0, %arg0: !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.le.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.le.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int4_0 = torch.constant.int 4
%2 = torch.aten.le.int %int4, %int4_0 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.gt.int$evaluate_to_true() -> !torch.bool {
Add folders for torch.aten.gt.int / torch.aten.ne.int This fixes a "regression" on ResNet where we weren't folding away all the control flow. For now, our policy is to "optimize hard enough" to make that control flow go away, because we don't yet have a way to lower to the backend the stuff guarded by the control flow (RaiseException, string operations, etc.). It remains to be seen how much optimization we decide to do at this level in the fullness of time -- the torch op set is not particularly well-designed (at least not idiomatically for MLIR) for general optimization. Ideally, with really good backend support for various features, all the heavy optimization will happen at that layer on `std` ops and `scf` control flow. But I have a suspicion we might end up needing more optimization earlier in the pipeline.
2021-06-17 02:05:08 +08:00
// CHECK-NEXT: %[[T:.*]] = torch.constant.bool true
// CHECK-NEXT: return %[[T]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.gt.int$evaluate_to_true() -> !torch.bool {
Add folders for torch.aten.gt.int / torch.aten.ne.int This fixes a "regression" on ResNet where we weren't folding away all the control flow. For now, our policy is to "optimize hard enough" to make that control flow go away, because we don't yet have a way to lower to the backend the stuff guarded by the control flow (RaiseException, string operations, etc.). It remains to be seen how much optimization we decide to do at this level in the fullness of time -- the torch op set is not particularly well-designed (at least not idiomatically for MLIR) for general optimization. Ideally, with really good backend support for various features, all the heavy optimization will happen at that layer on `std` ops and `scf` control flow. But I have a suspicion we might end up needing more optimization earlier in the pipeline.
2021-06-17 02:05:08 +08:00
%int2 = torch.constant.int 2
%int4 = torch.constant.int 4
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%0 = torch.aten.gt.int %int4, %int2 : !torch.int, !torch.int -> !torch.bool
Add folders for torch.aten.gt.int / torch.aten.ne.int This fixes a "regression" on ResNet where we weren't folding away all the control flow. For now, our policy is to "optimize hard enough" to make that control flow go away, because we don't yet have a way to lower to the backend the stuff guarded by the control flow (RaiseException, string operations, etc.). It remains to be seen how much optimization we decide to do at this level in the fullness of time -- the torch op set is not particularly well-designed (at least not idiomatically for MLIR) for general optimization. Ideally, with really good backend support for various features, all the heavy optimization will happen at that layer on `std` ops and `scf` control flow. But I have a suspicion we might end up needing more optimization earlier in the pipeline.
2021-06-17 02:05:08 +08:00
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.gt.int$evaluate_to_false() -> !torch.bool {
More folding for aten.gt.int, aten.ne.int and Aten__Getitem__TOp. - Fold more for aten.gt.int, aten.ne.int and Aten__Getitem__TOp - Some format cleaning up
2021-06-19 10:33:14 +08:00
// CHECK-NEXT: %[[T:.*]] = torch.constant.bool false
// CHECK-NEXT: return %[[T]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.gt.int$evaluate_to_false() -> !torch.bool {
More folding for aten.gt.int, aten.ne.int and Aten__Getitem__TOp. - Fold more for aten.gt.int, aten.ne.int and Aten__Getitem__TOp - Some format cleaning up
2021-06-19 10:33:14 +08:00
%int2 = torch.constant.int 2
%int4 = torch.constant.int 4
%0 = torch.aten.gt.int %int2, %int4 : !torch.int, !torch.int -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.int$evaluate_to_false() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.int$evaluate_to_false() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%2 = torch.aten.ge.int %int4, %int5 : !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.int$same_operand(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.int$same_operand(%arg0: !torch.int) -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%2 = torch.aten.ge.int %arg0, %arg0: !torch.int, !torch.int -> !torch.bool
return %2 : !torch.bool
Add folders for torch.aten.gt.int / torch.aten.ne.int This fixes a "regression" on ResNet where we weren't folding away all the control flow. For now, our policy is to "optimize hard enough" to make that control flow go away, because we don't yet have a way to lower to the backend the stuff guarded by the control flow (RaiseException, string operations, etc.). It remains to be seen how much optimization we decide to do at this level in the fullness of time -- the torch op set is not particularly well-designed (at least not idiomatically for MLIR) for general optimization. Ideally, with really good backend support for various features, all the heavy optimization will happen at that layer on `std` ops and `scf` control flow. But I have a suspicion we might end up needing more optimization earlier in the pipeline.
2021-06-17 02:05:08 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.int$same_value() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.int$same_value() -> !torch.bool {
More folding for aten.gt.int, aten.ne.int and Aten__Getitem__TOp. - Fold more for aten.gt.int, aten.ne.int and Aten__Getitem__TOp - Some format cleaning up
2021-06-19 10:33:14 +08:00
%int4 = torch.constant.int 4
%int4_0 = torch.constant.int 4
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%2 = torch.aten.ge.int %int4, %int4_0 : !torch.int, !torch.int -> !torch.bool
More folding for aten.gt.int, aten.ne.int and Aten__Getitem__TOp. - Fold more for aten.gt.int, aten.ne.int and Aten__Getitem__TOp - Some format cleaning up
2021-06-19 10:33:14 +08:00
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.float$evaluate_to_true() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.float$evaluate_to_true() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
%float4 = torch.constant.float 4.0
%float5 = torch.constant.float 5.0
%2 = torch.aten.lt.float %float4, %float5 : !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.float$same_operand(
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
// CHECK-SAME: %{{.*}}: !torch.float) -> !torch.bool {
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.float$same_operand(%arg0: !torch.float) -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
%2 = torch.aten.lt.float %arg0, %arg0: !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.lt.float$same_value() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.lt.float$same_value() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
%float4 = torch.constant.float 4.0
%float4_0 = torch.constant.float 4.0
%2 = torch.aten.lt.float %float4, %float4_0 : !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.gt.float$evaluate_to_true() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
// CHECK-NEXT: %[[T:.*]] = torch.constant.bool true
// CHECK-NEXT: return %[[T]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.gt.float$evaluate_to_true() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
%float2 = torch.constant.float 2.0
%float4 = torch.constant.float 4.0
%0 = torch.aten.gt.float %float4, %float2 : !torch.float, !torch.float -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.gt.float$evaluate_to_false() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
// CHECK-NEXT: %[[T:.*]] = torch.constant.bool false
// CHECK-NEXT: return %[[T]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.gt.float$evaluate_to_false() -> !torch.bool {
Add folder for aten.gt/lt.float
2022-02-11 05:25:25 +08:00
%float2 = torch.constant.float 2.0
%float4 = torch.constant.float 4.0
%0 = torch.aten.gt.float %float2, %float4 : !torch.float, !torch.float -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @comparison_with_torch.aten.size.int(
Fix for 0-size dim inferred incorrectly. The issue was in the canonicalizer for torch.aten.ge.int -- in cases where the operands were swapped, it would miscompile. This issue is fixed and folding support generalized to `torch.aten.size.int < 0` as well. Fixes #716
2022-03-31 05:10:51 +08:00
// CHECK-SAME: %[[ARG0:.*]]: !torch.vtensor<[?,2],unk>) -> (!torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool) {
// CHECK: %[[SIZE:.*]] = torch.aten.size.int %[[ARG0]], %int0 : !torch.vtensor<[?,2],unk>, !torch.int -> !torch.int
// CHECK: %[[GE_0_LHS:.*]] = torch.aten.ge.int %int0, %[[SIZE]] : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[LT_0_LHS:.*]] = torch.aten.lt.int %int0, %[[SIZE]] : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[EQ_0_LHS:.*]] = torch.aten.eq.int %int0, %[[SIZE]] : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[NE_0_LHS:.*]] = torch.aten.ne.int %int0, %[[SIZE]] : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[GT_0_RHS:.*]] = torch.aten.gt.int %[[SIZE]], %int0 : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[LE_0_RHS:.*]] = torch.aten.le.int %[[SIZE]], %int0 : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[EQ_0_RHS:.*]] = torch.aten.eq.int %[[SIZE]], %int0 : !torch.int, !torch.int -> !torch.bool
// CHECK: %[[NE_0_RHS:.*]] = torch.aten.ne.int %[[SIZE]], %int0 : !torch.int, !torch.int -> !torch.bool
// CHECK: return %true, %true, %false, %false, %[[GE_0_LHS]], %[[LT_0_LHS]], %[[EQ_0_LHS]], %[[NE_0_LHS]], %[[GT_0_RHS]], %[[LE_0_RHS]], %[[EQ_0_RHS]], %[[NE_0_RHS]] : !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @comparison_with_torch.aten.size.int(%arg0: !torch.vtensor<[?,2],unk>) -> (!torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int0 = torch.constant.int 0
Fix for 0-size dim inferred incorrectly. The issue was in the canonicalizer for torch.aten.ge.int -- in cases where the operands were swapped, it would miscompile. This issue is fixed and folding support generalized to `torch.aten.size.int < 0` as well. Fixes #716
2022-03-31 05:10:51 +08:00
%0 = torch.aten.size.int %arg0, %int0 : !torch.vtensor<[?,2],unk>, !torch.int -> !torch.int
// Cases we can fold.
%1 = torch.aten.le.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
%2 = torch.aten.ge.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
%3 = torch.aten.lt.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
%4 = torch.aten.gt.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
// Cases we cannot fold.
%5 = torch.aten.ge.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
%6 = torch.aten.lt.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
%7 = torch.aten.eq.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
%8 = torch.aten.ne.int %int0, %0 : !torch.int, !torch.int -> !torch.bool
%9 = torch.aten.gt.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
%10 = torch.aten.le.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
%11 = torch.aten.eq.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
%12 = torch.aten.ne.int %0, %int0 : !torch.int, !torch.int -> !torch.bool
return %1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12 : !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool, !torch.bool
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
}
Fix for 0-size dim inferred incorrectly. The issue was in the canonicalizer for torch.aten.ge.int -- in cases where the operands were swapped, it would miscompile. This issue is fixed and folding support generalized to `torch.aten.size.int < 0` as well. Fixes #716
2022-03-31 05:10:51 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.float$different_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.float$different_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
%float4 = torch.constant.float 4.0
%float5 = torch.constant.float 5.0
%2 = torch.aten.eq.float %float4, %float5 : !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.float$same_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.float$same_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
%float4 = torch.constant.float 4.0
%float4_0 = torch.constant.float 4.0
%2 = torch.aten.eq.float %float4, %float4_0 : !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.str$different_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.str$different_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
%str4 = torch.constant.str "4"
%str5 = torch.constant.str "5"
%2 = torch.aten.eq.str %str4, %str5 : !torch.str, !torch.str -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.str$same_operand(
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
// CHECK-SAME: %{{.*}}: !torch.str) -> !torch.bool {
// CHECK-NEXT: %[[F:.*]] = torch.constant.bool true
// CHECK-NEXT: return %[[F]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.str$same_operand(%arg0: !torch.str) -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
%0 = torch.aten.eq.str %arg0, %arg0 : !torch.str, !torch.str -> !torch.bool
return %0 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.str$same_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.str$same_value() -> !torch.bool {
Add support for constant_pad_nd Note that to enable folding of the code coming from an example like the ConstantPad2dStaticModule e2e test, support for other operations had to be added/improved: - aten::neg.int - aten::eq.float - aten::eq.str - prim::Uninitialized
2022-01-11 15:42:53 +08:00
%str4 = torch.constant.str "4"
%str4_0 = torch.constant.str "4"
%2 = torch.aten.eq.str %str4, %str4_0 : !torch.str, !torch.str -> !torch.bool
return %2 : !torch.bool
}
Add E2E support for aten.len.str (#969)
2022-07-08 01:41:55 +08:00
// CHECK-LABEL: func.func @torch.aten.len.str() -> !torch.int {
// CHECK: %[[INT7:.*]] = torch.constant.int 7
// CHECK: return %[[INT7]] : !torch.int
func.func @torch.aten.len.str() -> !torch.int {
%str = torch.constant.str "example"
%2 = torch.aten.len.str %str : !torch.str -> !torch.int
return %2 : !torch.int
}
// CHECK-LABEL: func.func @torch.aten.len.str$empty() -> !torch.int {
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: return %[[INT0]] : !torch.int
func.func @torch.aten.len.str$empty() -> !torch.int {
%str = torch.constant.str ""
%2 = torch.aten.len.str %str : !torch.str -> !torch.int
return %2 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__not__
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__not__() -> !torch.bool {
Fix compilation errors from MT model With the following changes the compilation can continue until RefineTypes pass: - Add operators without ODS into `torch_ods_gen.py` - Add some new optional and list types in `TorchTypes.td` - Add some folders for aten int type comparator ops - Modify GlobalizeObjectGraph.cpp. For global slots that's not used, dont check if an aliased value is stored in more than one of global slots. This can work around a failure where the same tensor is stored in multiple "version" slots which are not used.
2021-08-11 09:28:50 +08:00
%false = torch.constant.bool false
%ret = torch.aten.__not__ %false : !torch.bool -> !torch.bool
return %ret: !torch.bool
More folding for aten.gt.int, aten.ne.int and Aten__Getitem__TOp. - Fold more for aten.gt.int, aten.ne.int and Aten__Getitem__TOp - Some format cleaning up
2021-06-19 10:33:14 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.max.int$identity(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.int {
// CHECK: return %[[ARG]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.max.int$identity(%arg0: !torch.int) -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.prim.max.int %arg0, %arg0 : !torch.int, !torch.int -> !torch.int
return %0 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.max.int$constant() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: return %[[INT3]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.max.int$constant() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%int3 = torch.constant.int 3
%0 = torch.prim.max.int %int-1, %int3 : !torch.int, !torch.int -> !torch.int
return %0 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.min.self_int$basic() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[M1:.*]] = torch.constant.int -1
// CHECK: return %[[M1]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.min.self_int$basic() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int-1, %int0, %int1 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%1 = torch.prim.min.self_int %0 : !torch.list<int> -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.min.self_int$nofold$dynamic(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: torch.prim.min.self_int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.min.self_int$nofold$dynamic(%arg0: !torch.int) -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int-1 = torch.constant.int -1
%int0 = torch.constant.int 0
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int-1, %int0, %arg0: (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%1 = torch.prim.min.self_int %0 : !torch.list<int> -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.len.t$of_size(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<*,f32>) -> !torch.int {
// CHECK: %[[DIM:.*]] = torch.aten.dim %[[ARG]] : !torch.vtensor<*,f32> -> !torch.int
// CHECK: return %[[DIM]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.len.t$of_size(%arg0: !torch.vtensor<*,f32>) -> !torch.int {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.size %arg0 : !torch.vtensor<*,f32> -> !torch.list<int>
%1 = torch.aten.len.t %0 : !torch.list<int> -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %1 : !torch.int
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.dim$with_shape(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[?,?,?],f32>) -> !torch.int {
// CHECK: %[[DIM:.*]] = torch.constant.int 3
// CHECK: return %[[DIM]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.dim$with_shape(%arg0: !torch.vtensor<[?,?,?],f32>) -> !torch.int {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%0 = torch.aten.dim %arg0 : !torch.vtensor<[?,?,?],f32> -> !torch.int
return %0 : !torch.int
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.len.t$of_build_list(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.int {
// CHECK: %[[LEN:.*]] = torch.constant.int 4
// CHECK: return %[[LEN]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.len.t$of_build_list(%arg0: !torch.int) -> !torch.int {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %arg0, %arg0, %arg0, %arg0 : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
%1 = torch.aten.len.t %0 : !torch.list<int> -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %1 : !torch.int
Add aten::len.t, aten::size, and aten::gt.int primitive ops Also add some canonicalizations that finally reduce ResNet down to a single block.
2021-04-30 06:38:11 +08:00
}
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.len.t$no_fold_list_mutated()
func.func @torch.aten.len.t$no_fold_list_mutated() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int4 = torch.constant.int 4
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct : () -> !torch.list<int>
%1 = torch.aten.append.t %0, %int4 : !torch.list<int>, !torch.int -> !torch.list<int>
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: torch.aten.len.t
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%2 = torch.aten.len.t %0 : !torch.list<int> -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %2 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: %[[C5:.*]] = torch.constant.int 5
// CHECK: return %[[C5]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.t() -> !torch.int {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%int4 = torch.constant.int 4
%int5 = torch.constant.int 5
%int1 = torch.constant.int 1
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int4, %int5 : (!torch.int, !torch.int) -> !torch.list<int>
%1 = torch.aten.__getitem__.t %0, %int1 : !torch.list<int>, !torch.int -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %1 : !torch.int
Add TorchList type and prim::ListConstruct #218
2021-06-05 06:57:21 +08:00
}
// Not canonicalized because of passed in index
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t$no_change_test0(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: %[[C5:.*]] = torch.constant.int 5
Bump LLVM (#958)
2022-06-23 11:23:46 +08:00
// CHECK: %[[C4:.*]] = torch.constant.int 4
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK: %[[LIST:.*]] = torch.prim.ListConstruct %[[C4]], %[[C5]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[ITEM:.*]] = torch.aten.__getitem__.t %[[LIST]], %arg0 : !torch.list<int>, !torch.int -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: return %[[ITEM]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.t$no_change_test0(%arg0: !torch.int) -> !torch.int {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%int5 = torch.constant.int 5
%int4 = torch.constant.int 4
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int4, %int5 : (!torch.int, !torch.int) -> !torch.list<int>
%1 = torch.aten.__getitem__.t %0, %arg0 : !torch.list<int>, !torch.int -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %1 : !torch.int
Add TorchList type and prim::ListConstruct #218
2021-06-05 06:57:21 +08:00
}
// Not canonicalized because of passed in list
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t$no_change_test1(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: %[[C5:.*]] = torch.constant.int 5
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK: %[[ITEM:.*]] = torch.aten.__getitem__.t %arg0, %[[C5]] : !torch.list<int>, !torch.int -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK: return %[[ITEM]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.t$no_change_test1(%arg0: !torch.list<int>) -> !torch.int {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%int5 = torch.constant.int 5
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.__getitem__.t %arg0, %int5 : !torch.list<int>, !torch.int -> !torch.int
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %0 : !torch.int
Add TorchList type and prim::ListConstruct #218
2021-06-05 06:57:21 +08:00
}
Introduce native `!torch.none` type. - Add `torch.constant.none` op to construct it (naming is chosen to be analogous to Torch's representation of a prim::Constant with NoneType, rather than using the "singleton" terminology of Basicpy).
2021-06-15 02:36:10 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t$getitem_of_size(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[TENSOR:.*]]: !torch.tensor,
// CHECK-SAME: %[[INDEX:.*]]: !torch.int) -> !torch.int {
// CHECK: %[[RESULT:.*]] = torch.aten.size.int %[[TENSOR]], %[[INDEX]] : !torch.tensor, !torch.int -> !torch.int
// CHECK: return %[[RESULT]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.t$getitem_of_size(%arg0: !torch.tensor, %arg1: !torch.int) -> !torch.int {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.aten.size %arg0 : !torch.tensor -> !torch.list<int>
%1 = torch.aten.__getitem__.t %0, %arg1 : !torch.list<int>, !torch.int -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t$negative_index() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[INT8:.*]] = torch.constant.int 8
// CHECK: return %[[INT8]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.t$negative_index() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int7 = torch.constant.int 7
%int8 = torch.constant.int 8
%int-1 = torch.constant.int -1
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int7, %int8 : (!torch.int, !torch.int) -> !torch.list<int>
%1 = torch.aten.__getitem__.t %0, %int-1 : !torch.list<int>, !torch.int -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.t$invalid_index() -> !torch.int {
func.func @torch.aten.__getitem__.t$invalid_index() -> !torch.int {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%int7 = torch.constant.int 7
%int8 = torch.constant.int 8
%int-1 = torch.constant.int -100
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int7, %int8 : (!torch.int, !torch.int) -> !torch.list<int>
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: torch.aten.__getitem__.t
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%1 = torch.aten.__getitem__.t %0, %int-1 : !torch.list<int>, !torch.int -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
[MLIR][TORCH] Add support for multiple indexing tensors for aten.index.Tensor (#1097) - Includes a canonicalizer for `aten.add.t`needed for successfully lowering the shape function - Only offers support for statically sized index tensors when there is more than one - Dynamic shape support remains for single indexing tensors
2022-07-29 07:00:02 +08:00
// Not canonicalized because of mutated lhs list
// CHECK-LABEL: func.func @torch.aten.add.t$no_canonicalize_lhs_mutated()
func.func @torch.aten.add.t$no_canonicalize_lhs_mutated() -> !torch.list<int> {
%int4 = torch.constant.int 4
%0 = torch.prim.ListConstruct : () -> !torch.list<int>
%1 = torch.prim.ListConstruct : () -> !torch.list<int>
%2 = torch.aten.append.t %0, %int4 : !torch.list<int>, !torch.int -> !torch.list<int>
// CHECK: torch.aten.add.t
%3 = torch.aten.add.t %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.list<int>
return %3 : !torch.list<int>
}
// Not canonicalized because of mutated rhs list
// CHECK-LABEL: func.func @torch.aten.add.t$no_canonicalize_rhs_mutated()
func.func @torch.aten.add.t$no_canonicalize_rhs_mutated() -> !torch.list<int> {
%int4 = torch.constant.int 4
%0 = torch.prim.ListConstruct : () -> !torch.list<int>
%1 = torch.prim.ListConstruct : () -> !torch.list<int>
%2 = torch.aten.append.t %1, %int4 : !torch.list<int>, !torch.int -> !torch.list<int>
// CHECK: torch.aten.add.t
%3 = torch.aten.add.t %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.list<int>
return %3 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.add.t$concat(
// CHECK-SAME: %[[ARG0:.*]]: !torch.int,
// CHECK-SAME: %[[ARG1:.*]]: !torch.int) -> !torch.list<int> {
// CHECK: %[[LIST:.*]] = torch.prim.ListConstruct %[[ARG0]], %[[ARG1]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[LIST]] : !torch.list<int>
func.func @torch.aten.add.t$concat(%arg0: !torch.int, %arg1: !torch.int) -> !torch.list<int> {
%0 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%1 = torch.prim.ListConstruct %arg1 : (!torch.int) -> !torch.list<int>
%2 = torch.aten.add.t %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.list<int>
return %2 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.add.t$concat_empty(
// CHECK-SAME: %[[ARG0:.*]]: !torch.int) -> !torch.list<int> {
// CHECK: %[[LIST:.*]] = torch.prim.ListConstruct %[[ARG0]] : (!torch.int) -> !torch.list<int>
// CHECK: return %[[LIST]] : !torch.list<int>
func.func @torch.aten.add.t$concat_empty(%arg0: !torch.int) -> !torch.list<int> {
%0 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%1 = torch.prim.ListConstruct : () -> !torch.list<int>
%2 = torch.aten.add.t %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.list<int>
return %2 : !torch.list<int>
}
[MLIR] Add canonicalizer for aten.slice.t op (#1413) * [MLIR] Add canonicalizer for aten.slice.t op * Add mlir tests and strength the canonicalizer * rename variable Co-authored-by: Vremold <xremold@gamil.com>
2022-09-27 05:35:50 +08:00
// CHECK-LABEL: func.func @torch.aten.slice.t$basic() -> !torch.list<int> {
// CHECK: %int0 = torch.constant.int 0
// CHECK: %int1 = torch.constant.int 1
// CHECK: %[[RET:.*]] = torch.prim.ListConstruct %int0, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RET]] : !torch.list<int>
func.func @torch.aten.slice.t$basic() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.slice.t %0, %int0, %int-1, %int1 : !torch.list<int>, !torch.int, !torch.int, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.slice.t$none_start() -> !torch.list<int> {
// CHECK: %int0 = torch.constant.int 0
// CHECK: %int1 = torch.constant.int 1
// CHECK: %[[RET:.*]] = torch.prim.ListConstruct %int0, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RET]] : !torch.list<int>
func.func @torch.aten.slice.t$none_start() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%none = torch.constant.none
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.slice.t %0, %none, %int-1, %int1 : !torch.list<int>, !torch.none, !torch.int, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.slice.t$none_end() -> !torch.list<int> {
// CHECK: %int0 = torch.constant.int 0
// CHECK: %int1 = torch.constant.int 1
// CHECK: %int2 = torch.constant.int 2
// CHECK: %[[RET:.*]] = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RET]] : !torch.list<int>
func.func @torch.aten.slice.t$none_end() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%none = torch.constant.none
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.slice.t %0, %int0, %none, %int1 : !torch.list<int>, !torch.int, !torch.none, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.slice.t$start_exceed_range() -> !torch.list<int> {
// CHECK: %int0 = torch.constant.int 0
// CHECK: %int1 = torch.constant.int 1
// CHECK: %[[RET:.*]] = torch.prim.ListConstruct %int0, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RET]] : !torch.list<int>
func.func @torch.aten.slice.t$start_exceed_range() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%int-1000 = torch.constant.int -1000
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.slice.t %0, %int-1000, %int-1, %int1 : !torch.list<int>, !torch.int, !torch.int, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.slice.t$end_exceed_range() -> !torch.list<int> {
// CHECK: %int0 = torch.constant.int 0
// CHECK: %int1 = torch.constant.int 1
// CHECK: %int2 = torch.constant.int 2
// CHECK: %[[RET:.*]] = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RET]] : !torch.list<int>
func.func @torch.aten.slice.t$end_exceed_range() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%int1000 = torch.constant.int 1000
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.slice.t %0, %int0, %int1000, %int1 : !torch.list<int>, !torch.int, !torch.int, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
// Not canonicalized because of mutated l list
// CHECK-LABEL: func.func @torch.aten.slice.t$no_canonicalize_l_mutated()
func.func @torch.aten.slice.t$no_canonicalize_l_mutated() -> !torch.list<int> {
%int0 = torch.constant.int 0
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int-1 = torch.constant.int -1
%0 = torch.prim.ListConstruct %int0, %int1, %int2 : (!torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: torch.aten.slice.t
%2 = torch.aten.slice.t %0, %int0, %int-1, %int1 : !torch.list<int>, !torch.int, !torch.int, !torch.int -> !torch.list<int>
%3 = torch.aten.append.t %0, %int-1 : !torch.list<int>, !torch.int -> !torch.list<int>
return %2 : !torch.list<int>
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int_list$fold$literals_of_different_sizes
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[RET:.*]] = torch.constant.bool false
// CHECK: return %[[RET]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int_list$fold$literals_of_different_sizes(%arg0: !torch.int) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct : () -> !torch.list<int>
%1 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%2 = torch.aten.eq.int_list %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.bool
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int_list$fold$same_literal
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[RET:.*]] = torch.constant.bool true
// CHECK: return %[[RET]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.eq.int_list$fold$same_literal(%arg0: !torch.int) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%1 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%2 = torch.aten.eq.int_list %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.bool
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.eq.int_list$no_fold$different_literals(
func.func @torch.aten.eq.int_list$no_fold$different_literals(%arg0: !torch.int, %arg1: !torch.int) -> !torch.bool {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %arg0 : (!torch.int) -> !torch.list<int>
%1 = torch.prim.ListConstruct %arg1 : (!torch.int) -> !torch.list<int>
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: torch.aten.eq.int_list
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%2 = torch.aten.eq.int_list %0, %1 : !torch.list<int>, !torch.list<int> -> !torch.bool
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.Float.Scalar$constant_fold_int_to_float() -> !torch.float {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK: %[[VAL_0:.*]] = torch.constant.float 3.000000e+00
// CHECK: return %[[VAL_0]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.Float.Scalar$constant_fold_int_to_float() -> !torch.float {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.constant.int 3
%1 = torch.aten.Float.Scalar %0 : !torch.int -> !torch.float
return %1 : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.Float.Scalar$identity(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[VAL_0:.*]]: !torch.float) -> !torch.float {
// CHECK: return %[[VAL_0]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.Float.Scalar$identity(%arg0: !torch.float) -> !torch.float {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.aten.Float.Scalar %arg0 : !torch.float -> !torch.float
return %0 : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.constant.none$constantlike() -> (!torch.none, !torch.none) {
Introduce native `!torch.none` type. - Add `torch.constant.none` op to construct it (naming is chosen to be analogous to Torch's representation of a prim::Constant with NoneType, rather than using the "singleton" terminology of Basicpy).
2021-06-15 02:36:10 +08:00
// CHECK: %[[C:.*]] = torch.constant.none
// CHECK: return %[[C]], %[[C]] : !torch.none, !torch.none
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.constant.none$constantlike() -> (!torch.none, !torch.none) {
Introduce native `!torch.none` type. - Add `torch.constant.none` op to construct it (naming is chosen to be analogous to Torch's representation of a prim::Constant with NoneType, rather than using the "singleton" terminology of Basicpy).
2021-06-15 02:36:10 +08:00
%0 = torch.constant.none
%1 = torch.constant.none
return %0, %1 : !torch.none, !torch.none
}
Add !torch.str type. - Remove dependence on `!basicpy.BytesType`. - Add `torch.constant.str "s"` analogous to `torch.constant.none`.
2021-06-15 23:29:06 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.constant.str$constantlike() -> (!torch.str, !torch.str, !torch.str) {
Add !torch.str type. - Remove dependence on `!basicpy.BytesType`. - Add `torch.constant.str "s"` analogous to `torch.constant.none`.
2021-06-15 23:29:06 +08:00
// CHECK: %[[S:.*]] = torch.constant.str "s"
Bump LLVM (#958)
2022-06-23 11:23:46 +08:00
// CHECK: %[[T:.*]] = torch.constant.str "t"
Add !torch.str type. - Remove dependence on `!basicpy.BytesType`. - Add `torch.constant.str "s"` analogous to `torch.constant.none`.
2021-06-15 23:29:06 +08:00
// CHECK: return %[[S]], %[[S]], %[[T]] : !torch.str, !torch.str, !torch.str
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.constant.str$constantlike() -> (!torch.str, !torch.str, !torch.str) {
Add !torch.str type. - Remove dependence on `!basicpy.BytesType`. - Add `torch.constant.str "s"` analogous to `torch.constant.none`.
2021-06-15 23:29:06 +08:00
%0 = torch.constant.str "s"
%1 = torch.constant.str "s"
%2 = torch.constant.str "t"
return %0, %1, %2 : !torch.str, !torch.str, !torch.str
}
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.constant.bool$constantlike() -> (!torch.bool, !torch.bool, !torch.bool) {
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
// CHECK: %[[T:.*]] = torch.constant.bool true
Bump LLVM (#958)
2022-06-23 11:23:46 +08:00
// CHECK: %[[F:.*]] = torch.constant.bool false
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
// CHECK: return %[[T]], %[[T]], %[[F]] : !torch.bool, !torch.bool, !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.constant.bool$constantlike() -> (!torch.bool, !torch.bool, !torch.bool) {
Add `!torch.bool` type. This finishes removing the dependence on the basicpy dialect! Changes: - Add `!torch.bool` type and replace use of `!basicpy.BoolType` in Torch-related code. - Rename BuiltinTensorize to BackendTypeConversion since now it handles bool conversions (and, when we add !torch.int and !torch.float, it will handle those as well), and generalize the related utilities (I also moved them to Torch/Transforms since they aren't really part of Torch/IR). - Add `torch.to_i1` and `torch.from_i1` ops for materializations - [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td - Remove dependency of `torch` dialect on `basicpy` dialect and also `std` dialect. For `std`, we use some call related ops, but the `torch` dialect itself never produces them (we have passes that do though). This is fairly mechanical. Recommended review order: - New stuff in Torch/IR - New BuiltinTypeConversion files. - Mechnical fixups elsewhere.
2021-06-16 07:47:53 +08:00
%0 = torch.constant.bool true
%1 = torch.constant.bool true
%2 = torch.constant.bool false
return %0, %1, %2 : !torch.bool, !torch.bool, !torch.bool
}
Add `torch.prim.If` This removes the use of `scf.if`, which required laundering back and forth between `i1` and `!torch.bool` in the frontend. We will eventually lower this op to `scf.if`, but this results in a cleaner IR and layering at the frontend.
2021-06-17 01:23:26 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.If$erase_dead_branch(
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.int {
// CHECK-NEXT: %[[RET:.*]] = torch.aten.add.int %[[ARG]], %[[ARG]] : !torch.int, !torch.int -> !torch.int
// CHECK-NEXT: return %[[RET]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.If$erase_dead_branch(%arg0: !torch.int) -> !torch.int {
Add `torch.prim.If` This removes the use of `scf.if`, which required laundering back and forth between `i1` and `!torch.bool` in the frontend. We will eventually lower this op to `scf.if`, but this results in a cleaner IR and layering at the frontend.
2021-06-17 01:23:26 +08:00
%true = torch.constant.bool true
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%0 = torch.prim.If %true -> (!torch.int) {
%1 = torch.aten.add.int %arg0, %arg0 : !torch.int, !torch.int -> !torch.int
torch.prim.If.yield %1 : !torch.int
Add `torch.prim.If` This removes the use of `scf.if`, which required laundering back and forth between `i1` and `!torch.bool` in the frontend. We will eventually lower this op to `scf.if`, but this results in a cleaner IR and layering at the frontend.
2021-06-17 01:23:26 +08:00
} else {
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
%1 = torch.aten.mul.int %arg0, %arg0 : !torch.int, !torch.int -> !torch.int
torch.prim.If.yield %1 : !torch.int
Add `torch.prim.If` This removes the use of `scf.if`, which required laundering back and forth between `i1` and `!torch.bool` in the frontend. We will eventually lower this op to `scf.if`, but this results in a cleaner IR and layering at the frontend.
2021-06-17 01:23:26 +08:00
}
Add `!torch.int` type. This replaces the ad-hoc use of `i64` throughout the Torch layer, and helps to keep it crystal clear the distinction between `!torch.int` (which is modeling the Python `int` type) and the various types that serve as dtypes of tensors, which are a totally different type universe. Changes: - `!torch.int` type and C bindings. - Change `torch.constant.int` parser to not need the `: i64` at the end. - `m_TorchConstantInt` matcher to aid with matching constants. - BackendTypeConversion changes for `!torch.int` -> `i64` type conversion. - Refactor finalizing patterns in FinalizingBackendTypeConversionPass (they were getting very repetitive). - Mechanical rewriting of `!torch.int` to `i64` in all the tests, and `AnyTorchIntType` to `Torch_IntType` in the `.td` files.
2021-06-17 06:53:15 +08:00
return %0 : !torch.int
Add `torch.prim.If` This removes the use of `scf.if`, which required laundering back and forth between `i1` and `!torch.bool` in the frontend. We will eventually lower this op to `scf.if`, but this results in a cleaner IR and layering at the frontend.
2021-06-17 01:23:26 +08:00
}
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.If$no_fold$side_effect(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG0:.*]]: !torch.bool) {
// CHECK: %[[STR:.*]] = torch.constant.str "str"
// CHECK: torch.prim.If %[[ARG0]] -> () {
// CHECK: torch.prim.RaiseException %[[STR]], %[[STR]] : !torch.str, !torch.str
// CHECK: torch.prim.If.yield
// CHECK: } else {
// CHECK: torch.prim.If.yield
// CHECK: }
// CHECK: return
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.If$no_fold$side_effect(%arg0: !torch.bool) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%str = torch.constant.str "str"
torch.prim.If %arg0 -> () {
torch.prim.RaiseException %str, %str : !torch.str, !torch.str
torch.prim.If.yield
} else {
torch.prim.If.yield
}
return
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.If$fold_same_result(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[PRED:.*]]: !torch.bool,
// CHECK-SAME: %[[ARG1:.*]]: !torch.int) -> (!torch.int, !torch.int) {
// CHECK-NEXT: return %[[ARG1]], %[[ARG1]] : !torch.int, !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.If$fold_same_result(%arg0: !torch.bool, %arg1: !torch.int) -> (!torch.int, !torch.int) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0, %1 = torch.prim.If %arg0 -> (!torch.int, !torch.int) {
torch.prim.If.yield %arg1, %arg1 : !torch.int, !torch.int
} else {
torch.prim.If.yield %arg1, %arg1 : !torch.int, !torch.int
}
return %0, %1: !torch.int, !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.If$fold_same_result$subset_of_results(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[PRED:.*]]: !torch.bool,
// CHECK-SAME: %[[ARG1:.*]]: !torch.int,
// CHECK-SAME: %[[ARG2:.*]]: !torch.int) -> (!torch.int, !torch.int) {
// CHECK: %[[IF_RESULT:.*]] = torch.prim.If %[[PRED]] -> (!torch.int) {
// CHECK: torch.prim.If.yield %[[ARG1]] : !torch.int
// CHECK: } else {
// CHECK: torch.prim.If.yield %[[ARG2]] : !torch.int
// CHECK: }
// CHECK: return %[[ARG1]], %[[IF_RESULT:.*]] : !torch.int, !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.If$fold_same_result$subset_of_results(%arg0: !torch.bool, %arg1: !torch.int, %arg2: !torch.int) -> (!torch.int, !torch.int) {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0, %1 = torch.prim.If %arg0 -> (!torch.int, !torch.int) {
torch.prim.If.yield %arg1, %arg1: !torch.int, !torch.int
} else {
torch.prim.If.yield %arg1, %arg2: !torch.int, !torch.int
}
return %0, %1: !torch.int, !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.TupleUnpack(
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK-SAME: %[[ARG0:.*]]: !torch.tensor,
// CHECK-SAME: %[[ARG1:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: return %[[ARG0]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.TupleUnpack(%arg0: !torch.tensor, %arg1: !torch.tensor) -> !torch.tensor{
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%123 = torch.prim.TupleConstruct %arg0, %arg1: !torch.tensor, !torch.tensor -> !torch.tuple<tensor, tensor>
%124:2 = torch.prim.TupleUnpack %123 : !torch.tuple<tensor, tensor> -> !torch.tensor, !torch.tensor
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
return %124#0 : !torch.tensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__contains__.str(
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK-SAME: %[[K0:.*]]: !torch.str, %[[V0:.*]]: !torch.tensor,
// CHECK-SAME: %[[K1:.*]]: !torch.str,
// CHECK-SAME: %[[V1:.*]]: !torch.tensor) -> !torch.bool {
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: %[[DICT:.*]] = torch.prim.DictConstruct
// CHECK-SAME: keys(%[[K0]], %[[K1]] : !torch.str, !torch.str)
// CHECK-SAME: values(%[[V0]], %[[V1]] : !torch.tensor, !torch.tensor)
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: -> !torch.dict<str, tensor>
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__contains__.str(%k0 : !torch.str, %v0: !torch.tensor, %k1: !torch.str, %v1: !torch.tensor) -> !torch.bool{
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%dict = torch.prim.DictConstruct keys(%k0, %k1: !torch.str, !torch.str) values(%v0, %v1: !torch.tensor, !torch.tensor) -> !torch.dict<str, tensor>
%pred = torch.aten.__contains__.str %dict, %k0 : !torch.dict<str, tensor>, !torch.str -> !torch.bool
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
return %pred : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__contains__.str$with_dict_modified(
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK-SAME: %[[K0:.*]]: !torch.str, %[[V0:.*]]: !torch.tensor,
// CHECK-SAME: %[[K1:.*]]: !torch.str, %[[V1:.*]]: !torch.tensor) -> !torch.bool {
// CHECK: %[[DICT:.*]] = torch.prim.DictConstruct
// CHECK-SAME: keys(%[[K0]], %[[K1]] : !torch.str, !torch.str)
// CHECK-SAME: values(%[[V0]], %[[V1]] : !torch.tensor, !torch.tensor)
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: -> !torch.dict<str, tensor>
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: torch.aten._set_item.str %[[DICT]], %[[K0]], %[[V1]] :
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: !torch.dict<str, tensor>, !torch.str, !torch.tensor
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[RET:.*]] = torch.aten.__contains__.str %[[DICT]], %[[K0]] :
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: !torch.dict<str, tensor>, !torch.str -> !torch.bool
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: return %[[RET]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__contains__.str$with_dict_modified(%k0 : !torch.str, %v0: !torch.tensor, %k1: !torch.str, %v1: !torch.tensor) -> !torch.bool{
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%dict = torch.prim.DictConstruct keys(%k0, %k1: !torch.str, !torch.str) values(%v0, %v1: !torch.tensor, !torch.tensor) -> !torch.dict<str, tensor>
torch.aten._set_item.str %dict, %k0, %v1 : !torch.dict<str, tensor>, !torch.str, !torch.tensor
%pred = torch.aten.__contains__.str %dict, %k0 : !torch.dict<str, tensor>, !torch.str -> !torch.bool
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
return %pred : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.__getitem__.Dict_str(
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK-SAME: %[[K0:.*]]: !torch.str, %[[V0:.*]]: !torch.tensor,
// CHECK-SAME: %[[K1:.*]]: !torch.str, %[[V1:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: %[[DICT:.*]] = torch.prim.DictConstruct
// CHECK-SAME: keys(%[[K0]], %[[K1]] : !torch.str, !torch.str)
// CHECK-SAME: values(%[[V0]], %[[V1]] : !torch.tensor, !torch.tensor)
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: -> !torch.dict<str, tensor>
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: return %[[V0]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.__getitem__.Dict_str(%k0 : !torch.str, %v0: !torch.tensor, %k1: !torch.str, %v1: !torch.tensor) -> !torch.tensor {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%dict = torch.prim.DictConstruct keys(%k0, %k1: !torch.str, !torch.str) values(%v0, %v1: !torch.tensor, !torch.tensor) -> !torch.dict<str, tensor>
%v = torch.aten.__getitem__.Dict_str %dict, %k0 : !torch.dict<str, tensor>, !torch.str -> !torch.tensor
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
return %v : !torch.tensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.add.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST9:.*]] = torch.constant.int 9
// CHECK: return %[[CST9]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.add.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst4 = torch.constant.int 4
%cst5 = torch.constant.int 5
%ret = torch.aten.add.int %cst4, %cst5: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.sub.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST1:.*]] = torch.constant.int 1
// CHECK: return %[[CST1]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.sub.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst6 = torch.constant.int 6
%cst5 = torch.constant.int 5
%ret = torch.aten.sub.int %cst6, %cst5: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.mul.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST30:.*]] = torch.constant.int 30
// CHECK: return %[[CST30]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.mul.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst6 = torch.constant.int 6
%cst5 = torch.constant.int 5
%ret = torch.aten.mul.int %cst6, %cst5: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.mul.int$with_zero() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST0:.*]] = torch.constant.int 0
// CHECK: return %[[CST0]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.mul.int$with_zero() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst6 = torch.constant.int 6
%cst0 = torch.constant.int 0
%ret = torch.aten.mul.int %cst6, %cst0: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.floordiv.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST3:.*]] = torch.constant.int 3
// CHECK: return %[[CST3]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.floordiv.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst18 = torch.constant.int 18
%cst5 = torch.constant.int 5
%ret = torch.aten.floordiv.int %cst18, %cst5: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.remainder.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
// CHECK: %[[CST3:.*]] = torch.constant.int 3
// CHECK: return %[[CST3]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.remainder.int() -> !torch.int {
Add more folders to fold away branches Added folders to a few binary computing ops, `TupleUnpack`, `__contains__.str` and `__getitem__.Dict_str`.
2021-08-18 01:59:47 +08:00
%cst18 = torch.constant.int 18
%cst5 = torch.constant.int 5
%ret = torch.aten.remainder.int %cst18, %cst5: !torch.int, !torch.int -> !torch.int
return %ret : !torch.int
}
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
torch: fold prim.dtype(bf16) to integer constant 15 (#1012) A prior patch (63538de2) that added support for bfloat16 type did not add the canonicalization pattern to fold `torch.prim.dtype` operations on bfloat16 tensors into the integer constant 15. This patch fixes the problem.
2022-07-07 09:21:43 +08:00
// CHECK-LABEL: func.func @torch.prim.dtype$bfloat16(
// CHECK-SAME: %[[T:.*]]: !torch.tensor<*,bf16>) -> !torch.int {
// CHECK: %[[CST:.*]] = torch.constant.int 15
// CHECK: return %[[CST]] : !torch.int
func.func @torch.prim.dtype$bfloat16(%t : !torch.tensor<*,bf16>) -> !torch.int {
%ret = torch.prim.dtype %t: !torch.tensor<*,bf16> -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.dtype$float(
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<*,f32>) -> !torch.int {
// CHECK: %[[CST:.*]] = torch.constant.int 6
// CHECK: return %[[CST]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.dtype$float(%t : !torch.tensor<*,f32>) -> !torch.int {
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
%ret = torch.prim.dtype %t: !torch.tensor<*,f32> -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.dtype$bool(
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<*,i1>) -> !torch.int {
// CHECK: %[[CST:.*]] = torch.constant.int 11
// CHECK: return %[[CST]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.dtype$bool(%t : !torch.tensor<*,i1>) -> !torch.int {
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
%ret = torch.prim.dtype %t: !torch.tensor<*,i1> -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.dtype$int64(
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<*,si64>) -> !torch.int {
// CHECK: %[[CST:.*]] = torch.constant.int 4
// CHECK: return %[[CST]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.dtype$int64(%t : !torch.tensor<*,si64>) -> !torch.int {
MT model compilation minor changes This contains the following changes: - Fix optional knowledge propagation. The initial knowledge should always be NotNone for the operations we implemented. - Add Folder for `prim.dtype`
2021-09-02 03:53:52 +08:00
%ret = torch.prim.dtype %t: !torch.tensor<*,si64> -> !torch.int
return %ret : !torch.int
}
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.size.int$neg_dim(
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<[2,3],f32>) -> !torch.int {
// CHECK: %[[RET:.*]] = torch.constant.int 2
// CHECK: return %[[RET]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.size.int$neg_dim(%t: !torch.tensor<[2,3],f32>) -> !torch.int {
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
%int-2 = torch.constant.int -2
%ret = torch.aten.size.int %t, %int-2 : !torch.tensor<[2,3],f32>, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.size.int$pos_dim(
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<[2,3],f32>) -> !torch.int {
// CHECK: %[[RET:.*]] = torch.constant.int 3
// CHECK: return %[[RET]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.size.int$pos_dim(%t: !torch.tensor<[2,3],f32>) -> !torch.int {
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
%int1 = torch.constant.int 1
%ret = torch.aten.size.int %t, %int1 : !torch.tensor<[2,3],f32>, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.size.int$invalid_dim(
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<[2,3],f32>) -> !torch.int {
// CHECK: %[[CST3:.*]] = torch.constant.int 3
// CHECK: %[[RET:.*]] = torch.aten.size.int %[[T]], %[[CST3]] : !torch.tensor<[2,3],f32>, !torch.int -> !torch.int
// CHECK: return %[[RET]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.size.int$invalid_dim(%t: !torch.tensor<[2,3],f32>) -> !torch.int {
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
%int3 = torch.constant.int 3
%ret = torch.aten.size.int %t, %int3 : !torch.tensor<[2,3],f32>, !torch.int -> !torch.int
return %ret : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.unchecked_cast$derefine_identity(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.int {
// CHECK: return %[[ARG]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.unchecked_cast$derefine_identity(%arg0: !torch.int) -> !torch.int {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.derefine %arg0 : !torch.int to !torch.optional<int>
%1 = torch.prim.unchecked_cast %0 : !torch.optional<int> -> !torch.int
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.derefine$of_unchecked_cast(
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.optional<int>) -> !torch.optional<int> {
// CHECK: return %[[ARG]] : !torch.optional<int>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.derefine$of_unchecked_cast(%arg0: !torch.optional<int>) -> !torch.optional<int> {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.unchecked_cast %arg0 : !torch.optional<int> -> !torch.int
%1 = torch.derefine %0 : !torch.int to !torch.optional<int>
return %1 : !torch.optional<int>
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.derefine$use_allows_type_refinement(
Add minimal support for Union types. A recent PyTorch commit made ConstantPad2d call a helper function with a `Union[int, float]` type annotated. This commit adds minimal support for representing and dealing with that. https://github.com/pytorch/pytorch/pull/73287 Changes: - Adding support for `!torch.union<T1, T2, T3>`/`Torch::UnionType`, along with the importer and CAPI code. - Add support in isValidSubtype for union types. - Adding a canonicalizer for `torch.derefine` to help simplify some code that derefines to a UnionType (this also fixes #664). There is still more work to do for really supporting UnionType well, such as canonicalizing UnionType's so that they can be compared with pointer equality.
2022-03-30 06:57:31 +08:00
// CHECK-SAME: %{{.*}}: !torch.int) -> (!torch.vtensor, !torch.optional<int>) {
// CHECK: %[[NONE:.*]] = torch.constant.none
// CHECK: %[[DEREFINED:.*]] = torch.derefine %[[NONE]] : !torch.none to !torch.optional<int>
// For the use that allows type refinement, we replace it with the refined value.
// CHECK: %[[ARANGE:.*]] = torch.aten.arange.start %{{.*}}, %{{.*}}, %[[NONE]], %{{.*}}, %{{.*}}, %{{.*}} : !torch.int, !torch.int, !torch.none, !torch.none, !torch.none, !torch.none -> !torch.vtensor
// For the use that does not allow type refinement, don't replace.
// CHECK: return %[[ARANGE]], %[[DEREFINED]] : !torch.vtensor, !torch.optional<int>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.derefine$use_allows_type_refinement(%arg0: !torch.int) -> (!torch.vtensor, !torch.optional<int>) {
Add minimal support for Union types. A recent PyTorch commit made ConstantPad2d call a helper function with a `Union[int, float]` type annotated. This commit adds minimal support for representing and dealing with that. https://github.com/pytorch/pytorch/pull/73287 Changes: - Adding support for `!torch.union<T1, T2, T3>`/`Torch::UnionType`, along with the importer and CAPI code. - Add support in isValidSubtype for union types. - Adding a canonicalizer for `torch.derefine` to help simplify some code that derefines to a UnionType (this also fixes #664). There is still more work to do for really supporting UnionType well, such as canonicalizing UnionType's so that they can be compared with pointer equality.
2022-03-30 06:57:31 +08:00
%none = torch.constant.none
%optional = torch.derefine %none : !torch.none to !torch.optional<int>
%ret = torch.aten.arange.start %arg0, %arg0, %optional, %none, %none, %none: !torch.int, !torch.int, !torch.optional<int>, !torch.none, !torch.none, !torch.none -> !torch.vtensor
return %ret, %optional : !torch.vtensor, !torch.optional<int>
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$downcast_first(
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: return %[[T]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.tensor_static_info_cast$downcast_first(%t: !torch.tensor) -> !torch.tensor {
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
%downcast = torch.tensor_static_info_cast %t : !torch.tensor to !torch.tensor<[?,?],f64>
%upcast = torch.tensor_static_info_cast %downcast : !torch.tensor<[?,?],f64> to !torch.tensor
return %upcast: !torch.tensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$upcast_first(
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
// CHECK-SAME: %[[T:.*]]: !torch.tensor<[?,?],f64>) -> !torch.tensor<[?,?],f64> {
// CHECK: return %[[T]] : !torch.tensor<[?,?],f64>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.tensor_static_info_cast$upcast_first(%t: !torch.tensor<[?,?],f64>) -> !torch.tensor<[?,?],f64> {
E2e support for aten.softmax.int and aten.embedding - Added a DecomposeComplexOps pass to decompose complex torchOps. - Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to `visitReductionAlongDimIntOp`. - Moved some helper functions into torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files. - Added support for f64 tensor as argument and return types.
2021-10-16 06:23:59 +08:00
%upcast = torch.tensor_static_info_cast %t : !torch.tensor<[?,?],f64> to !torch.tensor
%downcast = torch.tensor_static_info_cast %upcast : !torch.tensor to !torch.tensor<[?,?],f64>
return %downcast: !torch.tensor<[?,?],f64>
}
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$refine(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[],f32>) -> !torch.vtensor {
// CHECK-NEXT: %[[RESULT:.*]] = torch.aten.relu %[[ARG]] : !torch.vtensor<[],f32> -> !torch.vtensor
// CHECK-NEXT: return %[[RESULT]] : !torch.vtensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.tensor_static_info_cast$refine(%arg0: !torch.vtensor<[], f32>) -> !torch.vtensor {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.tensor_static_info_cast %arg0 : !torch.vtensor<[],f32> to !torch.vtensor
%1 = torch.aten.relu %0 : !torch.vtensor -> !torch.vtensor
return %1 : !torch.vtensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$no_refine(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor) -> !torch.vtensor {
// CHECK: %[[CAST:.*]] = torch.tensor_static_info_cast %[[ARG]] : !torch.vtensor to !torch.vtensor<[],f32>
// CHECK: %[[RESULT:.*]] = torch.aten.relu %[[CAST]] : !torch.vtensor<[],f32> -> !torch.vtensor
// CHECK: return %[[RESULT]] : !torch.vtensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.tensor_static_info_cast$no_refine(%arg0: !torch.vtensor) -> !torch.vtensor {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.tensor_static_info_cast %arg0 : !torch.vtensor to !torch.vtensor<[],f32>
%1 = torch.aten.relu %0 : !torch.vtensor<[],f32> -> !torch.vtensor
return %1 : !torch.vtensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$refine_allowed_ops(
Make conditions for type refinement of static cast less strict (#680) This commit adds support for type refinement when `torch.tensor_static_info_cast`s are involved, even when there are users of the casted tensor that don't allow type refinements. Originally the canonicalization pattern for `torch.tensor_static_info_cast` would check if all the users of the casted tensor allowed type refinements before making any changes. This means that if at least one of the users did not allow type refinements, the pattern would fail. This becomes an issue when doing shape calculations because the calculations need the shape information of each input tensor to be available before the calculation can be simplified.
2022-03-19 00:10:12 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[],f32>) -> !torch.tuple<vtensor, vtensor> {
// CHECK: %[[CAST:.*]] = torch.tensor_static_info_cast %[[ARG]] : !torch.vtensor<[],f32> to !torch.vtensor
// CHECK: %[[RELU:.*]] = torch.aten.relu %[[ARG]] : !torch.vtensor<[],f32> -> !torch.vtensor
// CHECK: %[[RESULT:.*]] = torch.prim.TupleConstruct %[[CAST]], %[[RELU]] : !torch.vtensor, !torch.vtensor -> !torch.tuple<vtensor, vtensor>
// CHECK: return %[[RESULT]] : !torch.tuple<vtensor, vtensor>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.tensor_static_info_cast$refine_allowed_ops(%arg0: !torch.vtensor<[], f32>) -> !torch.tuple<vtensor, vtensor> {
Make conditions for type refinement of static cast less strict (#680) This commit adds support for type refinement when `torch.tensor_static_info_cast`s are involved, even when there are users of the casted tensor that don't allow type refinements. Originally the canonicalization pattern for `torch.tensor_static_info_cast` would check if all the users of the casted tensor allowed type refinements before making any changes. This means that if at least one of the users did not allow type refinements, the pattern would fail. This becomes an issue when doing shape calculations because the calculations need the shape information of each input tensor to be available before the calculation can be simplified.
2022-03-19 00:10:12 +08:00
%0 = torch.tensor_static_info_cast %arg0 : !torch.vtensor<[],f32> to !torch.vtensor
%1 = torch.aten.relu %0 : !torch.vtensor -> !torch.vtensor
// prim.TupleConstruct does not allow type refinements
%2 = torch.prim.TupleConstruct %0, %1 : !torch.vtensor, !torch.vtensor -> !torch.tuple<vtensor, vtensor>
return %2 : !torch.tuple<vtensor, vtensor>
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.TupleIndex(
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
// CHECK-SAME: %[[T0:.*]]: !torch.tensor, %[[T1:.*]]: !torch.tensor, %[[T2:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: return %[[T1]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.TupleIndex(%t0: !torch.tensor, %t1: !torch.tensor, %t2: !torch.tensor) -> !torch.tensor {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.TupleConstruct %t0, %t1, %t2 : !torch.tensor, !torch.tensor, !torch.tensor -> !torch.tuple<tensor, tensor, tensor>
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
%int1 = torch.constant.int 1
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%1 = torch.prim.TupleIndex %0, %int1 : !torch.tuple<tensor, tensor, tensor>, !torch.int -> !torch.tensor
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
return %1 : !torch.tensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.TupleIndex$out_of_bound(
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
// CHECK-SAME: %[[T0:.*]]: !torch.tensor, %[[T1:.*]]: !torch.tensor, %[[T2:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: %[[INDEX3:.*]] = torch.constant.int 3
// CHECK: %[[TUPLE:.*]] = torch.prim.TupleConstruct %[[T0]], %[[T1]], %[[T2]] :
// CHECK-SAME: !torch.tensor, !torch.tensor, !torch.tensor ->
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: !torch.tuple<tensor, tensor, tensor>
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
// CHECK: %[[RET:.*]] = torch.prim.TupleIndex %[[TUPLE]], %[[INDEX3]] :
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-SAME: !torch.tuple<tensor, tensor, tensor>, !torch.int -> !torch.tensor
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
// CHECK: return %[[RET]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.TupleIndex$out_of_bound(%t0: !torch.tensor, %t1: !torch.tensor, %t2: !torch.tensor) -> !torch.tensor {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.TupleConstruct %t0, %t1, %t2 : !torch.tensor, !torch.tensor, !torch.tensor -> !torch.tuple<tensor, tensor, tensor>
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
%int3 = torch.constant.int 3
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%1 = torch.prim.TupleIndex %0, %int3 : !torch.tuple<tensor, tensor, tensor>, !torch.int -> !torch.tensor
Add support for multiple return values This change is to unblock the work of some backprop ops returning more than one tensors. We will need to think of a more scalable approach in the future if more flexible return types combinations are needed.
2021-11-08 23:56:40 +08:00
return %1 : !torch.tensor
}
Add folder for torch.aten.Int.Tensor This is to fold the common pattern from Bert inference like: ``` %111 = torch.prim.NumToTensor.Scalar %110 : !torch.int -> !torch.vtensor<[],si64> %112 = torch.aten.Int.Tensor %111 : !torch.vtensor<[],si64> -> !torch.int ```
2021-11-30 02:39:37 +08:00
torch.prim.TupleIndex: Adjust tensor types when folding. In cases where a refinement/derefinement was needed, we didn't fold. Fixes https://github.com/llvm/torch-mlir/issues/863
2022-05-19 21:12:58 +08:00
// CHECK-LABEL: func.func @torch.prim.TupleIndex$adjust_type$tensor(
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[7],f32>) -> !torch.tensor {
// CHECK: %[[RETURN:.*]] = torch.tensor_static_info_cast %[[ARG]] : !torch.tensor<[7],f32> to !torch.tensor
// CHECK: return %[[RETURN]] : !torch.tensor
func.func @torch.prim.TupleIndex$adjust_type$tensor(%arg0: !torch.tensor<[7],f32>) -> !torch.tensor {
Fix TupleIndex canonicalizer. It would change the result type.
2022-05-03 17:12:09 +08:00
%int0 = torch.constant.int 0
torch.prim.TupleIndex: Adjust tensor types when folding. In cases where a refinement/derefinement was needed, we didn't fold. Fixes https://github.com/llvm/torch-mlir/issues/863
2022-05-19 21:12:58 +08:00
%0 = torch.prim.TupleConstruct %arg0 : !torch.tensor<[7],f32> -> !torch.tuple<tensor<[7],f32>>
%1 = torch.prim.TupleIndex %0, %int0 : !torch.tuple<tensor<[7],f32>>, !torch.int -> !torch.tensor
Fix TupleIndex canonicalizer. It would change the result type.
2022-05-03 17:12:09 +08:00
return %1 : !torch.tensor
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.prim.unchecked_cast$derefine
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
// CHECK-next: return %arg0 : !torch.list<int>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.prim.unchecked_cast$derefine(%arg0: !torch.list<int>) -> !torch.list<int> {
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.derefine %arg0 : !torch.list<int> to !torch.optional<list<int>>
%1 = torch.prim.unchecked_cast %0 : !torch.optional<list<int>> -> !torch.list<int>
return %1 : !torch.list<int>
Fold __is__ and unchecked_cast of derefine The added e2e maxpool testcase from #545 was not getting a static shape due to an unfolded prim.If when RefineTypes was called. This was because of unfolded torch.iaten.__is__ and torch.prim.unchecked_cast operators with torch.derefine operands.
2022-01-28 21:35:40 +08:00
}
Add folder for torch.aten.Int.Tensor This is to fold the common pattern from Bert inference like: ``` %111 = torch.prim.NumToTensor.Scalar %110 : !torch.int -> !torch.vtensor<[],si64> %112 = torch.aten.Int.Tensor %111 : !torch.vtensor<[],si64> -> !torch.int ```
2021-11-30 02:39:37 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.Int.Tensor(
Add folder for torch.aten.Int.Tensor This is to fold the common pattern from Bert inference like: ``` %111 = torch.prim.NumToTensor.Scalar %110 : !torch.int -> !torch.vtensor<[],si64> %112 = torch.aten.Int.Tensor %111 : !torch.vtensor<[],si64> -> !torch.int ```
2021-11-30 02:39:37 +08:00
// CHECK-SAME: %[[NUM:.*]]: !torch.int) -> !torch.int {
// CHECK: %[[T:.*]] = torch.prim.NumToTensor.Scalar %[[NUM]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[NUM]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.Int.Tensor(%arg0: !torch.int) -> !torch.int {
Add folder for torch.aten.Int.Tensor This is to fold the common pattern from Bert inference like: ``` %111 = torch.prim.NumToTensor.Scalar %110 : !torch.int -> !torch.vtensor<[],si64> %112 = torch.aten.Int.Tensor %111 : !torch.vtensor<[],si64> -> !torch.int ```
2021-11-30 02:39:37 +08:00
%tensor = torch.prim.NumToTensor.Scalar %arg0: !torch.int -> !torch.vtensor<[],si64>
%scalar = torch.aten.Int.Tensor %tensor : !torch.vtensor<[],si64> -> !torch.int
return %scalar : !torch.int
}
[MLIR][TORCH] Add E2E support for `aten.squeeze` op This commit adds lowering of `aten.Squeeze` op into `linalg.TensorCollapseShape` op. The size 1 dynamic dimensions are not handled as a part of this commit. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-25 04:19:13 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.Float.Tensor(
[LINALG] Add E2E support for `aten.[Bool.Tensor|Float.Tensor]` op - This commit adds lowering of `aten.Bool.Tensor` and `aten.Float.Tensor` op as a part of `convert-torch-to-linalg` pass. - It also adds support for returning bool types. - It also fixes lowering of the `aten.Int.Tensor` op for non-zero rank input tensors. - If a scalar number is converted to a 0-d tensor and passed on to the `aten.Float.Tensor` op, it folds to the scalar number. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2022-02-09 19:55:14 +08:00
// CHECK-SAME: %[[NUM:.*]]: !torch.float) -> !torch.float {
// CHECK: %[[T:.*]] = torch.prim.NumToTensor.Scalar %[[NUM]] : !torch.float -> !torch.vtensor<[],f64>
// CHECK: return %[[NUM]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.Float.Tensor(%arg0: !torch.float) -> !torch.float {
[LINALG] Add E2E support for `aten.[Bool.Tensor|Float.Tensor]` op - This commit adds lowering of `aten.Bool.Tensor` and `aten.Float.Tensor` op as a part of `convert-torch-to-linalg` pass. - It also adds support for returning bool types. - It also fixes lowering of the `aten.Int.Tensor` op for non-zero rank input tensors. - If a scalar number is converted to a 0-d tensor and passed on to the `aten.Float.Tensor` op, it folds to the scalar number. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2022-02-09 19:55:14 +08:00
%tensor = torch.prim.NumToTensor.Scalar %arg0: !torch.float -> !torch.vtensor<[],f64>
%scalar = torch.aten.Float.Tensor %tensor : !torch.vtensor<[],f64> -> !torch.float
return %scalar : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.squeeze$zero_rank(
[MLIR][TORCH] Add E2E support for `aten.squeeze` op This commit adds lowering of `aten.Squeeze` op into `linalg.TensorCollapseShape` op. The size 1 dynamic dimensions are not handled as a part of this commit. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-25 04:19:13 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[],f32>) -> !torch.tensor<[],f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<[],f32>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.squeeze$zero_rank(%arg0: !torch.tensor<[],f32>) -> !torch.tensor<[],f32> {
[MLIR][TORCH] Add E2E support for `aten.squeeze` op This commit adds lowering of `aten.Squeeze` op into `linalg.TensorCollapseShape` op. The size 1 dynamic dimensions are not handled as a part of this commit. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-25 04:19:13 +08:00
%0 = torch.aten.squeeze %arg0 : !torch.tensor<[],f32> -> !torch.tensor<[],f32>
return %0 : !torch.tensor<[],f32>
}
[TORCH][MLIR] Add E2E support for `aten.squeeze.dim` op This commit adds lowering of `aten.squeeze.dim` op into `linalg.TensorCollapseShape` op. Here, the dim(th) dimension of the input tensor is not supposed to be dynamic. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-30 22:50:55 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.squeeze.dim$zero_rank(
[TORCH][MLIR] Add E2E support for `aten.squeeze.dim` op This commit adds lowering of `aten.squeeze.dim` op into `linalg.TensorCollapseShape` op. Here, the dim(th) dimension of the input tensor is not supposed to be dynamic. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-30 22:50:55 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[],f32>) -> !torch.tensor<[],f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<[],f32>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.squeeze.dim$zero_rank(%arg0: !torch.tensor<[],f32>) -> !torch.tensor<[],f32> {
[TORCH][MLIR] Add E2E support for `aten.squeeze.dim` op This commit adds lowering of `aten.squeeze.dim` op into `linalg.TensorCollapseShape` op. Here, the dim(th) dimension of the input tensor is not supposed to be dynamic. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-11-30 22:50:55 +08:00
%int0 = torch.constant.int 0
%0 = torch.aten.squeeze.dim %arg0, %int0 : !torch.tensor<[],f32>, !torch.int -> !torch.tensor<[],f32>
return %0 : !torch.tensor<[],f32>
}
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
Fold away type_as (#1358)
2022-09-13 06:59:12 +08:00
// CHECK-LABEL: func.func @torch.aten.type_as$same(
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[?,?],f32>) -> !torch.tensor<[?,?],f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<[?,?],f32>
func.func @torch.aten.type_as$same(%arg0: !torch.tensor<[?,?],f32>) -> !torch.tensor<[?,?],f32> {
%0 = torch.aten.type_as %arg0, %arg0 : !torch.tensor<[?,?],f32>, !torch.tensor<[?,?],f32> -> !torch.tensor<[?,?],f32>
return %0 : !torch.tensor<[?,?],f32>
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.to.dtype$same_dtype(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<*,f32>) -> !torch.tensor<*,f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<*,f32>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.to.dtype$same_dtype(%arg0: !torch.tensor<*,f32>) -> !torch.tensor<*,f32> {
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%none = torch.constant.none
%false = torch.constant.bool false
%int6 = torch.constant.int 6
%0 = torch.aten.to.dtype %arg0, %int6, %false, %false, %none : !torch.tensor<*,f32>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.tensor<*,f32>
return %0 : !torch.tensor<*,f32>
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.to.dtype$no_fold$unk_dtype(
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor) -> !torch.tensor {
// CHECK: %[[RESULT:.*]] = torch.aten.to.dtype %[[ARG]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : !torch.tensor, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.tensor
// CHECK: return %[[RESULT]] : !torch.tensor
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.to.dtype$no_fold$unk_dtype(%arg0: !torch.tensor) -> !torch.tensor {
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
%none = torch.constant.none
%false = torch.constant.bool false
%int6 = torch.constant.int 6
Introduce new shape library design. See the documentation in `docs/shape_lib.md` and `docs/adding_a_shape_function.md` for an overview of the system. This completely overhauls how we represent shape functions. In particular, RefineTypes does not infer shapes anymore (only dtypes). Shape functions are now written in (TorchScript'able) Python. Recommended review order: 1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`. 1. Code and tests for ReifyShapeCalculations, DropShapeCalculations. 1. Code and tests for SimplifyShapeCalculations. 1. shape_lib_gen.py 1. Code and tests for new RefineTypes pass. 1. Random folders/canonicalizers in TorchOps.cpp and associated test in `canonicalize.mlir`. 1. New ReadOnly trait inferred from the registry. 1. Any miscellaneous remaining stuff. Example `-print-ir-after-all` for ElementwiseUnaryModule: [IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212). Example `-print-ir-after-all` for ElementwiseBinaryModule: [IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
2022-03-10 08:44:22 +08:00
%0 = torch.aten.to.dtype %arg0, %int6, %false, %false, %none : !torch.tensor, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.tensor
return %0 : !torch.tensor
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.view$1D(
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[?],f32>) -> !torch.tensor<[?],f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<[?],f32>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.view$1D(%arg0: !torch.tensor<[?],f32>) -> !torch.tensor<[?],f32> {
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
%int-1 = torch.constant.int -1
Elide `!torch.` prefix in nested dialect types. This leads to much more succinct types in many cases: ``` !torch.list<!torch.int> !torch.list<int> !torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>> !torch.tuple<list<int>, list<int>> !torch.optional<!torch.list<!torch.int>> !torch.optional<list<int>> !torch.list<list<list<tensor>>> !torch.list<!torch.list<!torch.list<!torch.tensor>>> ``` I would like to take this further and allow omitting the `!torch.` prefix in all cases, but that's harder -- for example, we currently use `FuncOp` for functions, and so I don't think we can customize the printing there. It seems like it will be a longer road to getting that level of customization.
2022-03-16 07:22:56 +08:00
%0 = torch.prim.ListConstruct %int-1 : (!torch.int) -> !torch.list<int>
%1 = torch.aten.view %arg0, %0 : !torch.tensor<[?],f32>, !torch.list<int> -> !torch.tensor<[?],f32>
[TORCH][MLIR] Fold trivial cases of `aten.to.dtype` and `aten.view` op - It folds `aten.to.dtype` when the input tensor type and result type are exactly same. - It folds `aten.view` when the rank of both the input tensor type and result type is unity. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2021-12-23 20:04:29 +08:00
return %1 : !torch.tensor<[?],f32>
}
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.div.float$fold_zero_dividend(
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
// CHECK: %[[CST0:.*]] = torch.constant.float 0.000000e+00
// CHECK: return %[[CST0]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.div.float$fold_zero_dividend() -> !torch.float {
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
%float0 = torch.constant.float 0.0
%float5 = torch.constant.float 5.0
%0 = torch.aten.div.float %float0, %float5 : !torch.float, !torch.float -> !torch.float
return %0 : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.div.float$fold_one_divisor(
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
// CHECK: %[[CST4:.*]] = torch.constant.float 4.000000e+00
// CHECK: return %[[CST4]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.div.float$fold_one_divisor() -> !torch.float {
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
%float4 = torch.constant.float 4.0
%float1 = torch.constant.float 1.0
%0 = torch.aten.div.float %float4, %float1 : !torch.float, !torch.float -> !torch.float
return %0 : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.div.float$fold_cst_operands(
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
// CHECK: %[[CST2:.*]] = torch.constant.float 2.000000e+00
// CHECK: return %[[CST2]] : !torch.float
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.div.float$fold_cst_operands() -> !torch.float {
[MLIR][TORCH] Add E2E support for aten.div.float op This commit adds lowering of `aten.div.float` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 20:06:41 +08:00
%float4 = torch.constant.float 4.0
%float2 = torch.constant.float 2.0
%0 = torch.aten.div.float %float4, %float2 : !torch.float, !torch.float -> !torch.float
return %0 : !torch.float
}
[MLIR][TORCH] Add E2E support for aten.to.dtype_layout op This commit decomposes `aten.to.dtype_layout` op into `aten.to.dtype` op. This commit also fixes the formatting for the file type_conversion.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-27 19:07:40 +08:00
[MLIR][TORCH] Add E2E support for `aten.[div.int|bitwise_or.Tensor]` ops This commit adds lowering of `aten.div.int` and `aten.bitwise_or.Tensor` ops. Both these ops are required in order to support bloom_560m model. Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
2022-10-06 21:11:52 +08:00
// CHECK-LABEL: func.func @torch.aten.div.int$fold_cst_operands(
// CHECK: %[[CST:.*]] = torch.constant.float 5.000000e-01
// CHECK: return %[[CST]] : !torch.float
func.func @torch.aten.div.int$fold_cst_operands() -> !torch.float {
%int2 = torch.constant.int 2
%int4 = torch.constant.int 4
%0 = torch.aten.div.int %int2, %int4 : !torch.int, !torch.int -> !torch.float
return %0 : !torch.float
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.to.dtype_layout$same_dtype(
[MLIR][TORCH] Add E2E support for aten.to.dtype_layout op This commit decomposes `aten.to.dtype_layout` op into `aten.to.dtype` op. This commit also fixes the formatting for the file type_conversion.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-27 19:07:40 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.tensor<[?,?],f32>) -> !torch.tensor<[?,?],f32> {
// CHECK-NEXT: return %[[ARG]] : !torch.tensor<[?,?],f32>
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.to.dtype_layout$same_dtype(%arg0: !torch.tensor<[?,?],f32>) -> !torch.tensor<[?,?],f32> {
[MLIR][TORCH] Add E2E support for aten.to.dtype_layout op This commit decomposes `aten.to.dtype_layout` op into `aten.to.dtype` op. This commit also fixes the formatting for the file type_conversion.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-27 19:07:40 +08:00
%none = torch.constant.none
%false = torch.constant.bool false
%int6 = torch.constant.int 6
%0 = torch.aten.to.dtype_layout %arg0, %int6, %none, %none, %none, %false, %false, %none : !torch.tensor<[?,?],f32>, !torch.int, !torch.none, !torch.none, !torch.none, !torch.bool, !torch.bool, !torch.none -> !torch.tensor<[?,?],f32>
return %0 : !torch.tensor<[?,?],f32>
}
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.float$same_operand(
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
// CHECK-SAME: %{{.*}}: !torch.float) -> !torch.bool {
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.float$same_operand(%arg0: !torch.float) -> !torch.bool {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
%2 = torch.aten.ge.float %arg0, %arg0: !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.float$same_value() -> !torch.bool {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: return %[[TRUE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.float$same_value() -> !torch.bool {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
%float4 = torch.constant.float 4.0
%float4_0 = torch.constant.float 4.0
%2 = torch.aten.ge.float %float4, %float4_0: !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ge.float$different_value() -> !torch.bool {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
// CHECK: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: return %[[FALSE]] : !torch.bool
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ge.float$different_value() -> !torch.bool {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
%float4 = torch.constant.float 4.0
%float4_0 = torch.constant.float 5.0
%2 = torch.aten.ge.float %float4, %float4_0: !torch.float, !torch.float -> !torch.bool
return %2 : !torch.bool
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ceil.float$fold_cst() -> !torch.int {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
// CHECK: %[[CST2:.*]] = torch.constant.int 2
// CHECK: return %[[CST2]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ceil.float$fold_cst() -> !torch.int {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
%float = torch.constant.float 1.5
%1 = torch.aten.ceil.float %float : !torch.float -> !torch.int
return %1 : !torch.int
}
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
// CHECK-LABEL: func.func @torch.aten.ceil.float$no_fold(
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
// CHECK-SAME: %[[ARG:.*]]: !torch.float) -> !torch.int {
// CHECK: %[[RESULT:.*]] = torch.aten.ceil.float %[[ARG]] : !torch.float -> !torch.int
// CHECK: return %[[RESULT]] : !torch.int
mlir: bump llvm tag to 5380e3 (#856) In addition to updating the llvm-project submodule, this patch also: 1. updates shape functions and tests so that `func` and `call` operations refer to the `func` dialect 2. avoid duplicate registration of dialects
2022-05-17 03:54:35 +08:00
func.func @torch.aten.ceil.float$no_fold(%arg0 : !torch.float) -> !torch.int {
[MLIR][TORCH] E2E support for [ge|ceil].float, [ge|ne|gt].float_int op This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`, `aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op. This commit also fixes formatting for the file scalar.py and scalar_comparison.py. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-04-25 21:12:45 +08:00
%1 = torch.aten.ceil.float %arg0 : !torch.float -> !torch.int
return %1 : !torch.int
}
[MLIR][TORCH] Add E2E support for aten.sqrt.int op This commit adds lowering of `aten.sqrt.int` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-05-19 22:54:16 +08:00
// CHECK-LABEL: func.func @torch.aten.sqrt.int$fold_cst() -> !torch.float {
// CHECK: %[[CST:.*]] = torch.constant.float 2.2360679774997898
// CHECK: return %[[CST]] : !torch.float
func.func @torch.aten.sqrt.int$fold_cst() -> !torch.float {
%int = torch.constant.int 5
%0 = torch.aten.sqrt.int %int : !torch.int -> !torch.float
return %0 : !torch.float
}
// CHECK-LABEL: func.func @torch.aten.sqrt.int$no_fold(
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.float {
// CHECK: %[[RESULT:.*]] = torch.aten.sqrt.int %[[ARG]] : !torch.int -> !torch.float
// CHECK: return %[[RESULT]] : !torch.float
func.func @torch.aten.sqrt.int$no_fold(%arg0 : !torch.int) -> !torch.float {
%0 = torch.aten.sqrt.int %arg0 : !torch.int -> !torch.float
return %0 : !torch.float
}
[MLIR][TORCH] Add E2E support for aten.Bool.[float|int] op This commit adds lowering of `aten.Bool.float` and `aten.Bool.int` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-05-20 16:26:52 +08:00
// CHECK-LABEL: func.func @torch.aten.Bool.float$fold_cst() -> !torch.bool {
// CHECK: %[[CST2:.*]] = torch.constant.bool true
// CHECK: return %[[CST2]] : !torch.bool
func.func @torch.aten.Bool.float$fold_cst() -> !torch.bool {
%float = torch.constant.float 1.5
%1 = torch.aten.Bool.float %float : !torch.float -> !torch.bool
return %1 : !torch.bool
}
// CHECK-LABEL: func.func @torch.aten.Bool.int$fold_cst() -> !torch.bool {
// CHECK: %[[CST2:.*]] = torch.constant.bool true
// CHECK: return %[[CST2]] : !torch.bool
func.func @torch.aten.Bool.int$fold_cst() -> !torch.bool {
%int = torch.constant.int 2
%1 = torch.aten.Bool.int %int : !torch.int -> !torch.bool
return %1 : !torch.bool
}
Add canonicalization for aten.add.tensor op
2022-06-18 02:49:36 +08:00
// CHECK-LABEL: func.func @torch.aten.add.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT0]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR3:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR3]] : !torch.vtensor<[],si64>
func.func @torch.aten.add.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int0 : !torch.int -> !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.add.Tensor %0, %1, %int3 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: @torch.aten.add.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR2]] : !torch.vtensor<[],si64>
func.func @torch.aten.add.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense<0> : tensor<si64>) : !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.add.Tensor %0, %1, %int3 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
torch: fold prim.dtype(bf16) to integer constant 15 (#1012) A prior patch (63538de2) that added support for bfloat16 type did not add the canonicalization pattern to fold `torch.prim.dtype` operations on bfloat16 tensors into the integer constant 15. This patch fixes the problem.
2022-07-07 09:21:43 +08:00
}
canonicalizer: propagate type information across copy and cast ops (#1030) Prior to this patch, the canonicalizers for `AtenSizeOp` and `AtenSizeIntOp` succeeded only if the tensor operand's type information included the size of the requested dimension(s). We can extend the set of optimizable cases by propagating types across operations whose result type matches the input tensor type. Specifically, this patch enables the canonicalizers for `AtenSizeOp` and `AtenSizeIntOp` to see past `tensor_static_info_cast`, `copy.to_vtensor`, and `copy.to_tensor` ops until it reaches the first op whose result type contains size information for the requested dimensions, with a maximum bound of 6 parent lookups to avoid indefinite compilation times. All other encountered ops cause the canonicalizer to give up.
2022-07-13 03:38:37 +08:00
// CHECK-LABEL: @torch.aten.size$copy(
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[2,3],f32>) -> !torch.list<int> {
// CHECK: %[[TWO:.*]] = torch.constant.int 2
// CHECK: %[[THREE:.*]] = torch.constant.int 3
// CHECK: %[[LIST:.*]] = torch.prim.ListConstruct %[[TWO]], %[[THREE]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[LIST]] : !torch.list<int>
// CHECK: }
func.func @torch.aten.size$copy(%arg0: !torch.vtensor<[2,3],f32>) -> !torch.list<int> {
%cast = torch.tensor_static_info_cast %arg0 : !torch.vtensor<[2,3],f32> to !torch.vtensor
%non_value_tensor = torch.copy.to_tensor %cast : !torch.tensor
%value_tensor = torch.copy.to_vtensor %non_value_tensor : !torch.vtensor
%size = torch.aten.size %value_tensor : !torch.vtensor -> !torch.list<int>
return %size : !torch.list<int>
}
// CHECK-LABEL: @torch.aten.size.int$copy(
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[2,3],f32>) -> !torch.int {
// CHECK: %[[TWO:.*]] = torch.constant.int 2
// CHECK: return %[[TWO]] : !torch.int
// CHECK: }
func.func @torch.aten.size.int$copy(%arg0: !torch.vtensor<[2,3],f32>) -> !torch.int {
%cast = torch.tensor_static_info_cast %arg0 : !torch.vtensor<[2,3],f32> to !torch.vtensor
%non_value_tensor = torch.copy.to_tensor %cast : !torch.tensor
%value_tensor = torch.copy.to_vtensor %non_value_tensor : !torch.vtensor
%zero = torch.constant.int 0
%size = torch.aten.size.int %value_tensor, %zero : !torch.vtensor, !torch.int -> !torch.int
return %size : !torch.int
}
[MLIR][TORCH] Add canonicalization pattern for prim.ListUnpack op This commit adds the canonicalization pattern for the `prim.ListUnpack` op. Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
2022-07-15 20:10:23 +08:00
// CHECK-LABEL: func.func @prim.ListUnpack$fold_list(
// CHECK-SAME: %[[ARG0:.*]]: !torch.vtensor<[2,3],f32>,
// CHECK-SAME: %[[ARG1:.*]]: !torch.vtensor<[2,3],f32>) -> (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>) {
// CHECK: return %[[ARG0]], %[[ARG1]] : !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>
func.func @prim.ListUnpack$fold_list(%arg0: !torch.vtensor<[2,3],f32>, %arg1: !torch.vtensor<[2,3],f32>) -> (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>) {
%0 = torch.prim.ListConstruct %arg0, %arg1 : (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>) -> !torch.list<vtensor>
%1:2 = torch.prim.ListUnpack %0 : !torch.list<vtensor> -> !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>
return %1#0, %1#1 : !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>
}
Generalize canonicalization pattern for more aten.sub/div/mul/add op (#1209) Generalize canonicalization pattern for more sub/div/mul/add op, but for AtenDivTensorModeOp in 'trunc' rounding mode, we try to fold it.
2022-08-16 13:24:08 +08:00
// CHECK-LABEL: func.func @torch.aten.div.Tensor_mode$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT3]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR2]] : !torch.vtensor<[],si64>
func.func @torch.aten.div.Tensor_mode$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int6 = torch.constant.int 6
%str = torch.constant.str "floor"
%0 = torch.vtensor.literal(dense<2> : tensor<si64>) : !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int6 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.div.Tensor_mode %1, %0, %str : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.str -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.div.Tensor_mode$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR3:.*]] = torch.prim.NumToTensor.Scalar %[[INT3]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR3]] : !torch.vtensor<[],si64>
func.func @torch.aten.div.Tensor_mode$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int6 = torch.constant.int 6
%int2 = torch.constant.int 2
%str = torch.constant.str "floor"
%0 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int6 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.div.Tensor_mode %1, %0, %str : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.str -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.add.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT0]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR1]] : !torch.vtensor<[],si64>
func.func @torch.aten.add.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int0 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.add.Scalar %0, %int2, %int3 : !torch.vtensor<[],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.add.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR0]] : !torch.vtensor<[],si64>
func.func @torch.aten.add.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense<0> : tensor<si64>) : !torch.vtensor<[],si64>
%2 = torch.aten.add.Scalar %0, %int2, %int3 : !torch.vtensor<[],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.sub.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT_6:.*]] = torch.constant.int -6
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT0]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR3:.*]] = torch.prim.NumToTensor.Scalar %[[INT_6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR3]] : !torch.vtensor<[],si64>
func.func @torch.aten.sub.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int0 : !torch.int -> !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.sub.Tensor %0, %1, %int3 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: @torch.aten.sub.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT_6:.*]] = torch.constant.int -6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT_6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR2]] : !torch.vtensor<[],si64>
func.func @torch.aten.sub.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense<0> : tensor<si64>) : !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.sub.Tensor %0, %1, %int3 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.sub.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT_6:.*]] = torch.constant.int -6
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT0]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT_6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR1]] : !torch.vtensor<[],si64>
func.func @torch.aten.sub.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int0 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.sub.Scalar %0, %int2, %int3 : !torch.vtensor<[],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.sub.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT_6:.*]] = torch.constant.int -6
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT_6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR0]] : !torch.vtensor<[],si64>
func.func @torch.aten.sub.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense<0> : tensor<si64>) : !torch.vtensor<[],si64>
%2 = torch.aten.sub.Scalar %0, %int2, %int3 : !torch.vtensor<[],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.mul.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6]] = torch.constant.int 6
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR0]] : !torch.vtensor<[],si64>
func.func @torch.aten.mul.Scalar$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense<2> : tensor<si64>) : !torch.vtensor<[],si64>
%2 = torch.aten.mul.Scalar %0, %int3 : !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.mul.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR1]] : !torch.vtensor<[],si64>
func.func @torch.aten.mul.Scalar$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.mul.Scalar %0, %int3 : !torch.vtensor<[],si64>, !torch.int -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.mul.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6]] = torch.constant.int 6
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR0]] : !torch.vtensor<[],si64>
func.func @torch.aten.mul.Tensor$canonicalize_literal_0d() -> !torch.vtensor<[],si64> {
%0 = torch.vtensor.literal(dense<2> : tensor<si64>) : !torch.vtensor<[],si64>
%1 = torch.vtensor.literal(dense<3> : tensor<si64>) : !torch.vtensor<[],si64>
%2 = torch.aten.mul.Tensor %0, %1 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64> -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.mul.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[PR0:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT3]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR2]] : !torch.vtensor<[],si64>
func.func @torch.aten.mul.Tensor$canonicalize_numtotensor_0d() -> !torch.vtensor<[],si64> {
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%0 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int3 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.mul.Tensor %0, %1 : !torch.vtensor<[],si64>, !torch.vtensor<[],si64> -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.div.Tensor_mode$canonicalize_numtotensor_0d_trunc() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT2]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR3:.*]] = torch.prim.NumToTensor.Scalar %[[INT3]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR3]] : !torch.vtensor<[],si64>
func.func @torch.aten.div.Tensor_mode$canonicalize_numtotensor_0d_trunc() -> !torch.vtensor<[],si64> {
%int6 = torch.constant.int 6
%int2 = torch.constant.int 2
%str = torch.constant.str "trunc"
%0 = torch.prim.NumToTensor.Scalar %int2 : !torch.int -> !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int6 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.div.Tensor_mode %1, %0, %str : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.str -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
}
// CHECK-LABEL: func.func @torch.aten.div.Tensor_mode$canonicalize_literal_0d_trunc() -> !torch.vtensor<[],si64> {
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT6:.*]] = torch.constant.int 6
// CHECK: %[[PR1:.*]] = torch.prim.NumToTensor.Scalar %[[INT6]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: %[[PR2:.*]] = torch.prim.NumToTensor.Scalar %[[INT3]] : !torch.int -> !torch.vtensor<[],si64>
// CHECK: return %[[PR2]] : !torch.vtensor<[],si64>
func.func @torch.aten.div.Tensor_mode$canonicalize_literal_0d_trunc() -> !torch.vtensor<[],si64> {
%int6 = torch.constant.int 6
%str = torch.constant.str "trunc"
%0 = torch.vtensor.literal(dense<2> : tensor<si64>) : !torch.vtensor<[],si64>
%1 = torch.prim.NumToTensor.Scalar %int6 : !torch.int -> !torch.vtensor<[],si64>
%2 = torch.aten.div.Tensor_mode %1, %0, %str : !torch.vtensor<[],si64>, !torch.vtensor<[],si64>, !torch.str -> !torch.vtensor<[],si64>
return %2 : !torch.vtensor<[],si64>
Fold away type_as (#1358)
2022-09-13 06:59:12 +08:00
}
[MLIR][TORCH] Add aten.sort.int op Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
2022-11-18 19:47:07 +08:00
// CHECK-LABEL: func.func @torch.aten.sort.int$reverse_false() -> !torch.list<int> {
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[INT1:.*]] = torch.constant.int 1
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[RESULT:.*]] = torch.prim.ListConstruct %[[INT0]], %[[INT1]], %[[INT2]], %[[INT3]] : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RESULT]] : !torch.list<int>
func.func @torch.aten.sort.int$reverse_false() -> !torch.list<int> {
%false = torch.constant.bool false
%int1 = torch.constant.int 1
%int0 = torch.constant.int 0
%int3 = torch.constant.int 3
%int2 = torch.constant.int 2
%0 = torch.prim.ListConstruct %int1, %int0, %int3, %int2 : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
torch.aten.sort.int %0, %false : !torch.list<int>, !torch.bool
return %0 : !torch.list<int>
}
// CHECK-LABEL: func.func @torch.aten.sort.int$reverse_true() -> !torch.list<int> {
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT2:.*]] = torch.constant.int 2
// CHECK: %[[INT1:.*]] = torch.constant.int 1
// CHECK: %[[INT0:.*]] = torch.constant.int 0
// CHECK: %[[RESULT:.*]] = torch.prim.ListConstruct %[[INT3]], %[[INT2]], %[[INT1]], %[[INT0]] : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
// CHECK: return %[[RESULT]] : !torch.list<int>
func.func @torch.aten.sort.int$reverse_true() -> !torch.list<int> {
%true = torch.constant.bool true
%int1 = torch.constant.int 1
%int0 = torch.constant.int 0
%int3 = torch.constant.int 3
%int2 = torch.constant.int 2
%0 = torch.prim.ListConstruct %int1, %int0, %int3, %int2 : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
torch.aten.sort.int %0, %true : !torch.list<int>, !torch.bool
return %0 : !torch.list<int>
}