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/Conversion/TorchToLinalg/elementwise.mlir

87 lines
6.1 KiB
MLIR
Raw Normal View History

Move TorchConversion dialect and TorchTo* into torch-mlir
2021-09-24 11:59:12 +08:00
// RUN: torch-mlir-opt <%s -convert-torch-to-linalg -split-input-file -mlir-print-local-scope -verify-diagnostics | FileCheck %s
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK-LABEL: func @elementwise$unary(
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<[],f32>) -> !torch.vtensor<[],f32> {
Add TorchToIREE and factor out TorchConversion dialect. This converts a basic list op (torch.prim.ListConstruct) to the IREE dialect. ``` def forward(self, x: float): return [x, x] ``` turns into: ``` builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> { %0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float> return %0 : !torch.list<!torch.float> } ``` which turns into: ``` builtin.func @forward(%arg0: f64) -> !iree.list<f64> { %c1 = constant 1 : index %c0 = constant 0 : index %c2 = constant 2 : index %0 = iree.list.create %c2 : !iree.list<f64> iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64 iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64 return %0 : !iree.list<f64> } ``` As part of doing this, I realized that it was time to formalize the IR form that we reach right before running TorchTo{Linalg,Std,...}. We now call it the "Torch backend contract". We then lower the "Torch backend contract" to the "npcomp backend contract", which involves the new TorchConversion (`torch_c`) dialect, which holds ops that need to operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE list, etc.) and the `!torch` types. This made more sense, as I realized that if I didn't factor out `torch_c` then the Torch dialect would have a dependency on IREE dialect (we previously didn't notice this was an issue because we only depended on `builtin` types), which seemed wrong to me. Recommended review order: - TorchToIREE.cpp / `TorchToIREE/basic.mlir` - Look at the new structure of createTorchScriptToNpcompBackendPipeline. It now lives in TorchConversion/Transforms/Passes.cpp and cleanly calls into `Torch::createTorchScriptToTorchBackendPipeline` for the frontend lowering to the Torch backend contract. - Mechanical change extracting `torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new TorchConversion dialect, and a few passes specific to the lowering from the Torch backend contract to the npcomp backend contract. - Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that we convert lists as part of operand materialization, we need to use the original operands). Also added test for AtenMaxPool2dOp and fixed m_TorchConstantIntList. - TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that are created as part of operand materialization for conv/max pool/avg pool ops in TorchToLinalg.
2021-08-12 05:40:08 +08:00
// CHECK: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[],f32> -> tensor<f32>
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: %[[INIT_TENSOR:.*]] = linalg.init_tensor [] : tensor<f32>
// CHECK: %[[GENERIC:.*]] = linalg.generic {indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>], iterator_types = []} ins(%[[BUILTIN_TENSOR]] : tensor<f32>) outs(%[[INIT_TENSOR]] : tensor<f32>) {
// CHECK: ^bb0(%[[BBARG0:.*]]: f32, %{{.*}}: f32):
// CHECK: %[[TANH:.*]] = math.tanh %[[BBARG0]] : f32
// CHECK: linalg.yield %[[TANH]] : f32
// CHECK: } -> tensor<f32>
// CHECK: %[[CASTED:.*]] = tensor.cast %[[GENERIC:.*]] : tensor<f32> to tensor<f32>
Add TorchToIREE and factor out TorchConversion dialect. This converts a basic list op (torch.prim.ListConstruct) to the IREE dialect. ``` def forward(self, x: float): return [x, x] ``` turns into: ``` builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> { %0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float> return %0 : !torch.list<!torch.float> } ``` which turns into: ``` builtin.func @forward(%arg0: f64) -> !iree.list<f64> { %c1 = constant 1 : index %c0 = constant 0 : index %c2 = constant 2 : index %0 = iree.list.create %c2 : !iree.list<f64> iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64 iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64 return %0 : !iree.list<f64> } ``` As part of doing this, I realized that it was time to formalize the IR form that we reach right before running TorchTo{Linalg,Std,...}. We now call it the "Torch backend contract". We then lower the "Torch backend contract" to the "npcomp backend contract", which involves the new TorchConversion (`torch_c`) dialect, which holds ops that need to operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE list, etc.) and the `!torch` types. This made more sense, as I realized that if I didn't factor out `torch_c` then the Torch dialect would have a dependency on IREE dialect (we previously didn't notice this was an issue because we only depended on `builtin` types), which seemed wrong to me. Recommended review order: - TorchToIREE.cpp / `TorchToIREE/basic.mlir` - Look at the new structure of createTorchScriptToNpcompBackendPipeline. It now lives in TorchConversion/Transforms/Passes.cpp and cleanly calls into `Torch::createTorchScriptToTorchBackendPipeline` for the frontend lowering to the Torch backend contract. - Mechanical change extracting `torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new TorchConversion dialect, and a few passes specific to the lowering from the Torch backend contract to the npcomp backend contract. - Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that we convert lists as part of operand materialization, we need to use the original operands). Also added test for AtenMaxPool2dOp and fixed m_TorchConstantIntList. - TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that are created as part of operand materialization for conv/max pool/avg pool ops in TorchToLinalg.
2021-08-12 05:40:08 +08:00
// CHECK: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[CASTED]] : tensor<f32> -> !torch.vtensor<[],f32>
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[],f32>
// CHECK: }
func @elementwise$unary(%arg0: !torch.vtensor<[],f32>) -> !torch.vtensor<[],f32> {
%0 = torch.aten.tanh %arg0 : !torch.vtensor<[],f32> -> !torch.vtensor<[],f32>
return %0 : !torch.vtensor<[],f32>
}
// CHECK-LABEL: func @elementwise$binary(
// CHECK-SAME: %[[ARG0:.*]]: !torch.vtensor<[?,?],f32>,
// CHECK-SAME: %[[ARG1:.*]]: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?,?],f32> {
Add TorchToIREE and factor out TorchConversion dialect. This converts a basic list op (torch.prim.ListConstruct) to the IREE dialect. ``` def forward(self, x: float): return [x, x] ``` turns into: ``` builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> { %0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float> return %0 : !torch.list<!torch.float> } ``` which turns into: ``` builtin.func @forward(%arg0: f64) -> !iree.list<f64> { %c1 = constant 1 : index %c0 = constant 0 : index %c2 = constant 2 : index %0 = iree.list.create %c2 : !iree.list<f64> iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64 iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64 return %0 : !iree.list<f64> } ``` As part of doing this, I realized that it was time to formalize the IR form that we reach right before running TorchTo{Linalg,Std,...}. We now call it the "Torch backend contract". We then lower the "Torch backend contract" to the "npcomp backend contract", which involves the new TorchConversion (`torch_c`) dialect, which holds ops that need to operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE list, etc.) and the `!torch` types. This made more sense, as I realized that if I didn't factor out `torch_c` then the Torch dialect would have a dependency on IREE dialect (we previously didn't notice this was an issue because we only depended on `builtin` types), which seemed wrong to me. Recommended review order: - TorchToIREE.cpp / `TorchToIREE/basic.mlir` - Look at the new structure of createTorchScriptToNpcompBackendPipeline. It now lives in TorchConversion/Transforms/Passes.cpp and cleanly calls into `Torch::createTorchScriptToTorchBackendPipeline` for the frontend lowering to the Torch backend contract. - Mechanical change extracting `torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new TorchConversion dialect, and a few passes specific to the lowering from the Torch backend contract to the npcomp backend contract. - Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that we convert lists as part of operand materialization, we need to use the original operands). Also added test for AtenMaxPool2dOp and fixed m_TorchConstantIntList. - TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that are created as part of operand materialization for conv/max pool/avg pool ops in TorchToLinalg.
2021-08-12 05:40:08 +08:00
// CHECK: %[[BUILTIN_ARG0:.*]] = torch_c.to_builtin_tensor %[[ARG0]] : !torch.vtensor<[?,?],f32> -> tensor<?x?xf32>
// CHECK: %[[BUILTIN_ARG1:.*]] = torch_c.to_builtin_tensor %[[ARG1]] : !torch.vtensor<[?],f32> -> tensor<?xf32>
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[C0:.*]] = arith.constant 0 : index
Bump llvm-project to da289a174fc6617c7be37be2947480510fd4f02a - Build adjustments for `.cpp.inc` dialect files. - Renaming of `memref.dim` to `tensor.dim` for tensor case. Minor changes: - Renaming of `mlir::linalg::ReassociationIndices` to `mlir::ReassociationIndices`. - Adjust command line option parsing in npcomp-run-mlir.
2021-07-08 02:55:31 +08:00
// CHECK: %[[ARG0_DIM0:.*]] = tensor.dim %[[BUILTIN_ARG0]], %[[C0]] : tensor<?x?xf32>
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[C1:.*]] = arith.constant 1 : index
Bump llvm-project to da289a174fc6617c7be37be2947480510fd4f02a - Build adjustments for `.cpp.inc` dialect files. - Renaming of `memref.dim` to `tensor.dim` for tensor case. Minor changes: - Renaming of `mlir::linalg::ReassociationIndices` to `mlir::ReassociationIndices`. - Adjust command line option parsing in npcomp-run-mlir.
2021-07-08 02:55:31 +08:00
// CHECK: %[[ARG0_DIM1:.*]] = tensor.dim %[[BUILTIN_ARG0]], %[[C1]] : tensor<?x?xf32>
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[C0_2:.*]] = arith.constant 0 : index
Bump llvm-project to da289a174fc6617c7be37be2947480510fd4f02a - Build adjustments for `.cpp.inc` dialect files. - Renaming of `memref.dim` to `tensor.dim` for tensor case. Minor changes: - Renaming of `mlir::linalg::ReassociationIndices` to `mlir::ReassociationIndices`. - Adjust command line option parsing in npcomp-run-mlir.
2021-07-08 02:55:31 +08:00
// CHECK: %[[ARG1_DIM0:.*]] = tensor.dim %[[BUILTIN_ARG1]], %[[C0_2]] : tensor<?xf32>
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[LEGAL_SIZES:.*]] = arith.cmpi eq, %[[ARG0_DIM1]], %[[ARG1_DIM0]] : index
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: assert %[[LEGAL_SIZES]], "mismatched size for broadcast"
// CHECK: %[[INIT_TENSOR:.*]] = linalg.init_tensor [%[[ARG0_DIM0]], %[[ARG0_DIM1]]] : tensor<?x?xf32>
// CHECK: %[[GENERIC:.*]] = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%[[BUILTIN_ARG0]], %[[BUILTIN_ARG1]] : tensor<?x?xf32>, tensor<?xf32>) outs(%[[INIT_TENSOR]] : tensor<?x?xf32>) {
// CHECK: ^bb0(%[[LHS:.*]]: f32, %[[RHS:.*]]: f32, %{{.*}}: f32):
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[MUL:.*]] = arith.mulf %[[LHS]], %[[RHS]] : f32
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: linalg.yield %[[MUL]] : f32
// CHECK: } -> tensor<?x?xf32>
// CHECK: %[[CASTED:.*]] = tensor.cast %[[GENERIC:.*]] : tensor<?x?xf32> to tensor<?x?xf32>
Add TorchToIREE and factor out TorchConversion dialect. This converts a basic list op (torch.prim.ListConstruct) to the IREE dialect. ``` def forward(self, x: float): return [x, x] ``` turns into: ``` builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> { %0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float> return %0 : !torch.list<!torch.float> } ``` which turns into: ``` builtin.func @forward(%arg0: f64) -> !iree.list<f64> { %c1 = constant 1 : index %c0 = constant 0 : index %c2 = constant 2 : index %0 = iree.list.create %c2 : !iree.list<f64> iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64 iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64 return %0 : !iree.list<f64> } ``` As part of doing this, I realized that it was time to formalize the IR form that we reach right before running TorchTo{Linalg,Std,...}. We now call it the "Torch backend contract". We then lower the "Torch backend contract" to the "npcomp backend contract", which involves the new TorchConversion (`torch_c`) dialect, which holds ops that need to operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE list, etc.) and the `!torch` types. This made more sense, as I realized that if I didn't factor out `torch_c` then the Torch dialect would have a dependency on IREE dialect (we previously didn't notice this was an issue because we only depended on `builtin` types), which seemed wrong to me. Recommended review order: - TorchToIREE.cpp / `TorchToIREE/basic.mlir` - Look at the new structure of createTorchScriptToNpcompBackendPipeline. It now lives in TorchConversion/Transforms/Passes.cpp and cleanly calls into `Torch::createTorchScriptToTorchBackendPipeline` for the frontend lowering to the Torch backend contract. - Mechanical change extracting `torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new TorchConversion dialect, and a few passes specific to the lowering from the Torch backend contract to the npcomp backend contract. - Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that we convert lists as part of operand materialization, we need to use the original operands). Also added test for AtenMaxPool2dOp and fixed m_TorchConstantIntList. - TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that are created as part of operand materialization for conv/max pool/avg pool ops in TorchToLinalg.
2021-08-12 05:40:08 +08:00
// CHECK: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[CASTED]] : tensor<?x?xf32> -> !torch.vtensor<[?,?],f32>
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[?,?],f32>
func @elementwise$binary(%arg0: !torch.vtensor<[?,?],f32>, %arg1: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?,?],f32> {
%0 = torch.aten.mul.Tensor %arg0, %arg1 : !torch.vtensor<[?,?],f32>, !torch.vtensor<[?],f32> -> !torch.vtensor<[?,?],f32>
return %0 : !torch.vtensor<[?,?],f32>
}
// CHECK-LABEL: func @elementwise$ternary(
// CHECK: linalg.generic {indexing_maps = [
// CHECK-SAME: affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
// CHECK-SAME: affine_map<(d0, d1, d2) -> (d1, d2)>,
// CHECK-SAME: affine_map<(d0, d1, d2) -> (d2)>,
// CHECK-SAME: affine_map<(d0, d1, d2) -> (d0, d1, d2)>]
func @elementwise$ternary(%arg0: !torch.vtensor<[?,?,?],f32>, %arg1: !torch.vtensor<[?,?],f32>, %arg2: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?,?,?],f32> {
%0 = torch.aten.lerp.Tensor %arg0, %arg1, %arg2 : !torch.vtensor<[?,?,?],f32>, !torch.vtensor<[?,?],f32>, !torch.vtensor<[?],f32> -> !torch.vtensor<[?,?,?],f32>
return %0 : !torch.vtensor<[?,?,?],f32>
}
// CHECK-LABEL: func @elementwise$with_scalar_capture(
// CHECK-SAME: %[[VAL_0:.*]]: !torch.vtensor<[?],f32>,
// CHECK-SAME: %[[VAL_1:.*]]: !torch.vtensor<[],f32>) -> !torch.vtensor<[?],f32> {
// CHECK: %[[C1:.*]] = torch.constant.int 1
Add TorchToIREE and factor out TorchConversion dialect. This converts a basic list op (torch.prim.ListConstruct) to the IREE dialect. ``` def forward(self, x: float): return [x, x] ``` turns into: ``` builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> { %0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float> return %0 : !torch.list<!torch.float> } ``` which turns into: ``` builtin.func @forward(%arg0: f64) -> !iree.list<f64> { %c1 = constant 1 : index %c0 = constant 0 : index %c2 = constant 2 : index %0 = iree.list.create %c2 : !iree.list<f64> iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64 iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64 return %0 : !iree.list<f64> } ``` As part of doing this, I realized that it was time to formalize the IR form that we reach right before running TorchTo{Linalg,Std,...}. We now call it the "Torch backend contract". We then lower the "Torch backend contract" to the "npcomp backend contract", which involves the new TorchConversion (`torch_c`) dialect, which holds ops that need to operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE list, etc.) and the `!torch` types. This made more sense, as I realized that if I didn't factor out `torch_c` then the Torch dialect would have a dependency on IREE dialect (we previously didn't notice this was an issue because we only depended on `builtin` types), which seemed wrong to me. Recommended review order: - TorchToIREE.cpp / `TorchToIREE/basic.mlir` - Look at the new structure of createTorchScriptToNpcompBackendPipeline. It now lives in TorchConversion/Transforms/Passes.cpp and cleanly calls into `Torch::createTorchScriptToTorchBackendPipeline` for the frontend lowering to the Torch backend contract. - Mechanical change extracting `torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new TorchConversion dialect, and a few passes specific to the lowering from the Torch backend contract to the npcomp backend contract. - Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that we convert lists as part of operand materialization, we need to use the original operands). Also added test for AtenMaxPool2dOp and fixed m_TorchConstantIntList. - TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that are created as part of operand materialization for conv/max pool/avg pool ops in TorchToLinalg.
2021-08-12 05:40:08 +08:00
// CHECK: %[[BUILTIN_C1:.*]] = torch_c.to_i64 %[[C1]]
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> ()>, affine_map<(d0) -> (d0)>]
// CHECK: ^bb0(%[[LHS:.*]]: f32, %[[RHS:.*]]: f32, %{{.*}}: f32):
Update llvm-project to a54f4eae0e1d0ef5adccdcf9f6c2b518dc1101aa This brings in https://reviews.llvm.org/D110797. PRs that are in progress will need to use scripts provided by https://llvm.discourse.group/t/psa-removed-arithmetic-ops-from-standard/4455.
2021-10-16 02:34:29 +08:00
// CHECK: %[[ALPHA:.*]] = arith.sitofp %[[BUILTIN_C1]] : i64 to f32
// CHECK: %[[SCALED:.*]] = arith.mulf %[[RHS]], %[[ALPHA]] : f32
// CHECK: %[[RES:.*]] = arith.addf %[[LHS]], %[[SCALED]] : f32
Generalize support for elementwise ops. We plumb through e2e a fair number of interesting cases: - unary, binary, ternary elementwise ops - ops like `torch.aten.add.Tensor` that also take a scalar parameter - static size-1 broadcasting We allow the static size-1 broadcasting case, but emit a runtime error in the case of dynamic size-1 broadcasting. This seems like a sweet spot subset of things that can be lowered directly to linalg, while not being overly constraining to users. This is consistent with what IREE is doing for CHLO->Linalg lowering as well ([code](https://github.com/google/iree/blob/50bf7a87e465d2048c527bc27724edde40519b7e/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp#L1)). To test the static size-1 case, we added support for the `torch.aten.unsqueeze` op and lowering for it through `linalg.tensor_expand_shape`. This involved a generalization of `MaximizeValueSemantics` able to handle it (the solution there also works for `torch.aten.flatten.using_ints` which we need for ResNet anyway) Also, a few minor additional changes: - Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a large class of errors before we get to backend lowering (now that we are doing dialect conversion, the errors are way nicer if we just emit them up front rather than in the guts of a random pattern). - Minor change to RefBackend to allow `linalg.tensor_expand_shape`. Recommended review order: - e2e tests in elementwise.py - `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test - `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test - RefineTypes.cpp + tests - MaximizeValueSemantics changes + test - VerifyInvariantsBeforeBackendLowering pass + test
2021-06-26 08:25:09 +08:00
// CHECK: linalg.yield %[[RES]] : f32
// CHECK: } -> tensor<?xf32>
func @elementwise$with_scalar_capture(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vtensor<[],f32>) -> !torch.vtensor<[?],f32> {
%int1 = torch.constant.int 1
%0 = torch.aten.add.Tensor %arg0, %arg1, %int1 : !torch.vtensor<[?],f32>, !torch.vtensor<[],f32>, !torch.int -> !torch.vtensor<[?],f32>
return %0 : !torch.vtensor<[?],f32>
}
// CHECK-LABEL: func @elementwise$static_1(
// CHECK: linalg.generic {indexing_maps = [
// CHECK-SAME: affine_map<(d0) -> (d0)>,
// CHECK-SAME: affine_map<(d0) -> (0)>,
// CHECK-SAME: affine_map<(d0) -> (d0)>]
func @elementwise$static_1(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vtensor<[1],f32>) -> !torch.vtensor<[?],f32> {
%1 = torch.aten.mul.Tensor %arg0, %arg1 : !torch.vtensor<[?],f32>, !torch.vtensor<[1],f32> -> !torch.vtensor<[?],f32>
return %1 : !torch.vtensor<[?],f32>
}