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/reify-shape-calculations.mlir

245 lines
17 KiB
MLIR
Raw Normal View History

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
// RUN: torch-mlir-opt -torch-reify-shape-calculations -split-input-file %s | FileCheck %s
// CHECK: module {
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: func.func private @__torch_mlir_shape_fn.aten.tanh(
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 @basic(
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: %[[RESULT:.*]] = torch.shape.calculate {
// CHECK: %[[TANH:.*]] = torch.aten.tanh %[[ARG]] : !torch.vtensor -> !torch.vtensor
// CHECK: torch.shape.calculate.yield %[[TANH]] : !torch.vtensor
// CHECK: } shapes {
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: %[[SHAPE:.*]] = torch.aten.size %[[ARG]] : !torch.vtensor -> !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: %[[RESULT_SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.tanh(%[[SHAPE]]) : (!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
// CHECK: torch.shape.calculate.yield.shapes %[[RESULT_SHAPE]] : !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.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 @basic(%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.aten.tanh %arg0 : !torch.vtensor -> !torch.vtensor
return %0 : !torch.vtensor
}
// -----
// CHECK: module {
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
// CHECK: func.func private @__torch_mlir_shape_fn.aten.bernoulli.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
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 @valsem_ops(
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.vtensor,
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
// CHECK-SAME: %[[ARG1:.*]]: !torch.float,
// CHECK-SAME: %[[ARG2:.*]]: !torch.Generator) -> !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
// CHECK: %[[RESULT:.*]] = torch.shape.calculate {
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
// CHECK: %[[VALUE:.*]] = torch.valsem.aten.bernoulli.float %[[ARG0]], %[[ARG1]], %[[ARG2]] : !torch.vtensor, !torch.float, !torch.Generator -> !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
// CHECK: torch.shape.calculate.yield %[[VALUE]] : !torch.vtensor
// CHECK: } shapes {
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: %[[SHAPE:.*]] = torch.aten.size %[[ARG0]] : !torch.vtensor -> !torch.list<int>
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
// CHECK: %[[RESULT_SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.bernoulli.float(%[[SHAPE]], %{{.*}}) : (!torch.list<int>, !torch.float, !torch.any) -> !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
// CHECK: torch.shape.calculate.yield.shapes %[[RESULT_SHAPE]] : !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.vtensor
// CHECK: return %[[RESULT:.*]] : !torch.vtensor
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
func.func @valsem_ops(%arg0: !torch.vtensor, %arg1: !torch.float, %arg2: !torch.Generator) -> !torch.vtensor {
%0 = torch.valsem.aten.bernoulli.float %arg0, %arg1, %arg2 : !torch.vtensor, !torch.float, !torch.Generator -> !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
return %0 : !torch.vtensor
}
// -----
// CHECK: module {
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 private @__torch_mlir_shape_fn.aten.uniform(
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.any)
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 @adjust_shape_function_arg$torch.any(
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.vtensor,
// CHECK-SAME: %[[ARG1:.*]]: !torch.float) -> !torch.vtensor {
// CHECK: %[[NONE:.*]] = torch.constant.none
// CHECK: %[[RESULT:.*]] = torch.shape.calculate {
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
// CHECK: %[[UNIFORM:.*]] = torch.aten.uniform %[[ARG0]], %[[ARG1]], %[[ARG1]], %[[NONE]] : !torch.vtensor, !torch.float, !torch.float, !torch.none -> !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
// CHECK: torch.shape.calculate.yield %[[UNIFORM]] : !torch.vtensor
// CHECK: } shapes {
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: %[[ARG0_SHAPE:.*]] = torch.aten.size %[[ARG0]] : !torch.vtensor -> !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: %[[ANY:.*]] = torch.derefine %[[NONE]] : !torch.none to !torch.any
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: %[[SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.uniform(%[[ARG0_SHAPE]], %[[ARG1]], %[[ARG1]], %[[ANY]]) : (!torch.list<int>, !torch.float, !torch.float, !torch.any) -> !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
// CHECK: torch.shape.calculate.yield.shapes %[[SHAPE]] : !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.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 @adjust_shape_function_arg$torch.any(%arg0: !torch.vtensor, %arg1: !torch.float) -> !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
%none = torch.constant.none
Remove all but one of valsem ops + move fill.Scalar to elementwise (#1531) This commit removes almost all of the valsem ops, since the value semantics version of the ops now exist in PyTorch. The only op missing is `aten.bernoulli_.float`. In addition, this commit also simplifies the implementation of `aten.fill.Scalar` by moving it to the pattern that converts elementwise ops.
2022-10-28 23:06:11 +08:00
%0 = torch.aten.uniform %arg0, %arg1, %arg1, %none : !torch.vtensor, !torch.float, !torch.float, !torch.none -> !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
return %0 : !torch.vtensor
}
// -----
// torch.aten.add.Tensor also has a shape function that calls a "broadcast"
// helper, so this test also checks our logic for transitively pulling in
// callees of the shape functions.
// CHECK: module {
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: func.func private @__torch_mlir_shape_fn.aten.add.Tensor(
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 @adjust_shape_function_arg$scalar(
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.vtensor,
// CHECK-SAME: %[[ARG1:.*]]: !torch.vtensor) -> !torch.vtensor {
// CHECK: %[[INT1:.*]] = torch.constant.int 1
// CHECK: %[[RESULT:.*]] = torch.shape.calculate {
// CHECK: %[[ADD:.*]] = torch.aten.add.Tensor %[[ARG0]], %[[ARG1]], %[[INT1]] : !torch.vtensor, !torch.vtensor, !torch.int -> !torch.vtensor
// CHECK: torch.shape.calculate.yield %[[ADD]] : !torch.vtensor
// CHECK: } shapes {
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: %[[ARG0_SHAPE:.*]] = torch.aten.size %[[ARG0]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[ARG1_SHAPE:.*]] = torch.aten.size %[[ARG1]] : !torch.vtensor -> !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: %[[SCALAR_CONVERTED:.*]] = torch.aten.Float.Scalar %[[INT1]] : !torch.int -> !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: %[[RESULT_SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.add.Tensor(%[[ARG0_SHAPE]], %[[ARG1_SHAPE]], %[[SCALAR_CONVERTED]]) : (!torch.list<int>, !torch.list<int>, !torch.float) -> !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
// CHECK: torch.shape.calculate.yield.shapes %[[RESULT_SHAPE]] : !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.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 @adjust_shape_function_arg$scalar(%arg0: !torch.vtensor, %arg1: !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
%int1 = torch.constant.int 1
%0 = torch.aten.add.Tensor %arg0, %arg1, %int1 : !torch.vtensor, !torch.vtensor, !torch.int -> !torch.vtensor
return %0 : !torch.vtensor
}
// -----
// CHECK: module {
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: func.func private @__torch_mlir_shape_fn.aten.topk(
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 @multiple_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: %[[ARG:.*]]: !torch.tensor) -> (!torch.tensor, !torch.tensor) {
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: %[[INT3:.*]] = torch.constant.int 3
// CHECK: %[[INT1:.*]] = torch.constant.int 1
// CHECK: %[[RESULTS:.*]]:2 = torch.shape.calculate {
// CHECK: %[[TOP_VALUES:.*]], %[[TOPK_INDICES:.*]] = torch.aten.topk %[[ARG]], %[[INT3]], %[[INT1]], %[[TRUE]], %[[TRUE]] : !torch.tensor, !torch.int, !torch.int, !torch.bool, !torch.bool -> !torch.tensor, !torch.tensor
// CHECK: torch.shape.calculate.yield %[[TOP_VALUES]], %[[TOPK_INDICES]] : !torch.tensor, !torch.tensor
// CHECK: } shapes {
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: %[[ARG_SHAPE:.*]] = torch.aten.size %[[ARG]] : !torch.tensor -> !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: %[[TOPK_SHAPE_TUPLE:.*]] = func.call @__torch_mlir_shape_fn.aten.topk(%[[ARG_SHAPE]], %[[INT3]], %[[INT1]], %[[TRUE]], %[[TRUE]]) : (!torch.list<int>, !torch.int, !torch.int, !torch.bool, !torch.bool) -> !torch.tuple<list<int>, 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
// CHECK: %[[TOPK_SHAPE:.*]]:2 = torch.prim.TupleUnpack %[[TOPK_SHAPE_TUPLE]] : !torch.tuple<list<int>, list<int>> -> !torch.list<int>, !torch.list<int>
// CHECK: torch.shape.calculate.yield.shapes %[[TOPK_SHAPE]]#0, %[[TOPK_SHAPE]]#1 : !torch.list<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.tensor, !torch.tensor
// CHECK: return %[[RESULTS:.*]]#0, %[[RESULTS]]#1 : !torch.tensor, !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 @multiple_results(%arg0: !torch.tensor) -> (!torch.tensor, !torch.tensor) {
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
%true = torch.constant.bool true
%int3 = torch.constant.int 3
%int1 = torch.constant.int 1
%values, %indices = torch.aten.topk %arg0, %int3, %int1, %true, %true : !torch.tensor, !torch.int, !torch.int, !torch.bool, !torch.bool -> !torch.tensor, !torch.tensor
return %values, %indices : !torch.tensor, !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 @adjust_shape_function_arg$optional(
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.vtensor,
// CHECK-SAME: %[[ARG1:.*]]: !torch.vtensor) -> !torch.vtensor {
// CHECK: %[[RESULT:.*]] = torch.shape.calculate {
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: %[[CONV:.*]] = torch.aten.conv2d %[[ARG0]], %[[ARG1]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : !torch.vtensor, !torch.vtensor, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.int -> !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
// CHECK: torch.shape.calculate.yield %[[CONV]] : !torch.vtensor
// CHECK: } shapes {
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: %[[SHAPE0:.*]] = torch.aten.size %[[ARG0]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[SHAPE1:.*]] = torch.aten.size %[[ARG1]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[DEREFINED:.*]] = torch.derefine %{{.*}} : !torch.none to !torch.optional<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: %[[SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.conv2d(%[[SHAPE0]], %[[SHAPE1]], %[[DEREFINED]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) : (!torch.list<int>, !torch.list<int>, !torch.optional<list<int>>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.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
// CHECK: torch.shape.calculate.yield.shapes %[[SHAPE]] : !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.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 @adjust_shape_function_arg$optional(%arg0: !torch.vtensor, %arg1: !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
%int3 = torch.constant.int 3
%int4 = torch.constant.int 4
%int2 = torch.constant.int 2
%int1 = torch.constant.int 1
%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
%24 = torch.prim.ListConstruct %int2, %int2 : (!torch.int, !torch.int) -> !torch.list<int>
%25 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
%26 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
%29 = torch.aten.conv2d %arg0, %arg1, %none, %24, %25, %26, %int1 : !torch.vtensor, !torch.vtensor, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.int -> !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
return %29 : !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 @adjust_shape_function_arg$optional_tensor(
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: %[[FALSE:.*]] = torch.constant.bool false
// CHECK: %[[TRUE:.*]] = torch.constant.bool true
// CHECK: %[[C1EM5:.*]] = torch.constant.float 1.000000e-05
// CHECK: %[[C1EM1:.*]] = torch.constant.float 1.000000e-01
// CHECK: %[[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
// CHECK: %[[DEREFINED:.*]] = torch.derefine %[[ARG]] : !torch.vtensor to !torch.optional<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
// CHECK: %[[RESULT:.*]] = torch.shape.calculate {
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: %[[BN:.*]] = torch.aten.batch_norm %[[ARG]], %[[DEREFINED]], %[[NONE]], %[[NONE]], %[[NONE]], %[[FALSE]], %[[C1EM1]], %[[C1EM5]], %[[TRUE]] : !torch.vtensor, !torch.optional<vtensor>, !torch.none, !torch.none, !torch.none, !torch.bool, !torch.float, !torch.float, !torch.bool -> !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
// CHECK: torch.shape.calculate.yield %[[BN]] : !torch.vtensor
// CHECK: } shapes {
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: %[[ARG_SIZE:.*]] = torch.aten.size %[[ARG]] : !torch.vtensor -> !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: %[[NONE2:.*]] = 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
// CHECK: %[[IS:.*]] = torch.aten.__is__ %[[DEREFINED]], %[[NONE2]] : !torch.optional<vtensor>, !torch.none -> !torch.bool
// CHECK: %[[DEREFINED_OPTIONAL_SIZE:.*]] = torch.prim.If %[[IS]] -> (!torch.optional<list<int>>) {
// CHECK: %[[DEREFINE_NONE:.*]] = torch.derefine %[[NONE2]] : !torch.none to !torch.optional<list<int>>
// CHECK: torch.prim.If.yield %[[DEREFINE_NONE]] : !torch.optional<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: } else {
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: %[[DOWNCASTED:.*]] = torch.prim.unchecked_cast %[[DEREFINED]] : !torch.optional<vtensor> -> !torch.vtensor
// CHECK: %[[DOWNCASTED_SIZE:.*]] = torch.aten.size %[[DOWNCASTED]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[DEREFINE_DOWNCASTED_SIZE:.*]] = torch.derefine %[[DOWNCASTED_SIZE]] : !torch.list<int> to !torch.optional<list<int>>
// CHECK: torch.prim.If.yield %[[DEREFINE_DOWNCASTED_SIZE]] : !torch.optional<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: }
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: %[[DEREFINED_NONE1:.*]] = torch.derefine %[[NONE]] : !torch.none to !torch.optional<list<int>>
// CHECK: %[[DEREFINED_NONE2:.*]] = torch.derefine %[[NONE]] : !torch.none to !torch.optional<list<int>>
// CHECK: %[[DEREFINED_NONE3:.*]] = torch.derefine %[[NONE]] : !torch.none to !torch.optional<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: %[[BN_SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.batch_norm(%[[ARG_SIZE]], %[[DEREFINED_OPTIONAL_SIZE:.*]], %[[DEREFINED_NONE1]], %[[DEREFINED_NONE2]], %[[DEREFINED_NONE3]], %[[FALSE]], %[[C1EM1]], %[[C1EM5]], %[[TRUE]]) : (!torch.list<int>, !torch.optional<list<int>>, !torch.optional<list<int>>, !torch.optional<list<int>>, !torch.optional<list<int>>, !torch.bool, !torch.float, !torch.float, !torch.bool) -> !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
// CHECK: torch.shape.calculate.yield.shapes %[[BN_SHAPE]] : !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.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 @adjust_shape_function_arg$optional_tensor(%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
%false = torch.constant.bool false
%true = torch.constant.bool true
%float1.000000e-05 = torch.constant.float 1.000000e-05
%float1.000000e-01 = torch.constant.float 1.000000e-01
%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
%derefined_tensor = torch.derefine %arg0 : !torch.vtensor to !torch.optional<vtensor>
%0 = torch.aten.batch_norm %arg0, %derefined_tensor, %none, %none, %none, %false, %float1.000000e-01, %float1.000000e-05, %true : !torch.vtensor, !torch.optional<vtensor>, !torch.none, !torch.none, !torch.none, !torch.bool, !torch.float, !torch.float, !torch.bool -> !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
return %0 : !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 @adjust_shape_function_arg$list(
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.vtensor,
// CHECK-SAME: %[[ARG1:.*]]: !torch.vtensor) -> !torch.vtensor {
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 %[[ARG1]] : (!torch.vtensor) -> !torch.list<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
// CHECK: %[[VAL_3:.*]] = torch.shape.calculate {
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: %[[VAL_4:.*]] = torch.aten.index.Tensor %[[ARG0]], %[[LIST]] : !torch.vtensor, !torch.list<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
// CHECK: torch.shape.calculate.yield %[[VAL_4]] : !torch.vtensor
// CHECK: } shapes {
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: %[[ARG0_SHAPE:.*]] = torch.aten.size %[[ARG0]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[ADJUSTED_LIST:.*]] = torch.prim.ListConstruct : () -> !torch.list<optional<list<int>>>
// CHECK: %[[LIST_SIZE:.*]] = torch.aten.len.t %[[LIST]] : !torch.list<vtensor> -> !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: %[[CTRUE:.*]] = torch.constant.bool true
// CHECK: torch.prim.Loop %[[LIST_SIZE]], %[[CTRUE]], init() {
// CHECK: ^bb0(%[[ITER_NUM:.*]]: !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
// CHECK: %[[UNADJUSTED_ELEMENT:.*]] = torch.aten.__getitem__.t %[[LIST]], %[[ITER_NUM]] : !torch.list<vtensor>, !torch.int -> !torch.vtensor
// CHECK: %[[UNADJUSTED_ELEMENT_SHAPE:.*]] = torch.aten.size %[[UNADJUSTED_ELEMENT]] : !torch.vtensor -> !torch.list<int>
// CHECK: %[[ADJUSTED_ELEMENT:.*]] = torch.derefine %[[UNADJUSTED_ELEMENT_SHAPE]] : !torch.list<int> to !torch.optional<list<int>>
// CHECK: %{{.*}} = torch.aten.append.t %[[ADJUSTED_LIST]], %[[ADJUSTED_ELEMENT]] : !torch.list<optional<list<int>>>, !torch.optional<list<int>> -> !torch.list<optional<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.prim.Loop.condition %[[CTRUE]], iter()
// CHECK: } : (!torch.int, !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: %[[RESULT_SHAPE:.*]] = func.call @__torch_mlir_shape_fn.aten.index.Tensor(%[[ARG0_SHAPE]], %[[ADJUSTED_LIST]]) : (!torch.list<int>, !torch.list<optional<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
// CHECK: torch.shape.calculate.yield.shapes %[[RESULT_SHAPE]] : !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.vtensor
// CHECK: return %[[VAL_15:.*]] : !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 @adjust_shape_function_arg$list(%arg0: !torch.vtensor, %arg1: !torch.vtensor) -> !torch.vtensor {
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 %arg1 : (!torch.vtensor) -> !torch.list<vtensor>
%1 = torch.aten.index.Tensor %arg0, %0 : !torch.vtensor, !torch.list<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
return %1 : !torch.vtensor
}
Cast `number` to `float` when shape function takes Scalar arg (#1978) To keep things simple in shape functions, `Scalar` inputs are considered `float`s. This means that when inserting the shape functions into the IR, we must cast any `!torch.number`s into `float`s so that the operand type matches the expected type in the shape function. This commit adds the cast from `Scalar` to `float`.
2023-03-29 00:30:31 +08:00
// -----
// CHECK-LABEL: func.func @adjust_shape_function_arg$number(
// CHECK: %[[FLOAT:.*]] = torch.aten.Float.Scalar {{.*}} : !torch.number -> !torch.float
// CHECK: %[[VAL_9:.*]] = func.call @__torch_mlir_shape_fn.aten.arange(%[[FLOAT]], {{.*}}) : (!torch.float, {{.*}}
func.func @adjust_shape_function_arg$number(%arg0: !torch.number) -> !torch.vtensor {
%none = torch.constant.none
%1 = torch.aten.arange %arg0, %none, %none, %none, %none : !torch.number, !torch.none, !torch.none, !torch.none, !torch.none -> !torch.vtensor
return %1 : !torch.vtensor
}