torch-mlir/test/Conversion/TorchToLinalg/basic.mlir

223 lines
13 KiB
MLIR
Raw Normal View History

// RUN: torch-mlir-opt <%s -convert-torch-to-linalg -split-input-file -verify-diagnostics | FileCheck %s
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
// CHECK-LABEL: func.func @torch.aten.mm$basic(
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// CHECK-SAME: %[[LHS_VTENSOR:.*]]: !torch.vtensor<[?,?],f32>,
// CHECK-SAME: %[[RHS_VTENSOR:.*]]: !torch.vtensor<[?,?],f32>) -> !torch.vtensor<[?,2],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: %[[LHS:.*]] = torch_c.to_builtin_tensor %[[LHS_VTENSOR]] : !torch.vtensor<[?,?],f32> -> tensor<?x?xf32>
// CHECK: %[[RHS:.*]] = torch_c.to_builtin_tensor %[[RHS_VTENSOR]] : !torch.vtensor<[?,?],f32> -> tensor<?x?xf32>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[LHS_DIM_0:.*]] = tensor.dim %[[LHS]], %[[C0]] : tensor<?x?xf32>
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[LHS_DIM_1:.*]] = tensor.dim %[[LHS]], %[[C1]] : tensor<?x?xf32>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[RHS_DIM_0:.*]] = tensor.dim %[[RHS]], %[[C0]] : tensor<?x?xf32>
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[RHS_DIM_1:.*]] = tensor.dim %[[RHS]], %[[C1]] : tensor<?x?xf32>
// CHECK: %[[EQ:.*]] = arith.cmpi eq, %[[LHS_DIM_1]], %[[RHS_DIM_0]] : index
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
// CHECK: assert %[[EQ]], "mismatching contracting dimension for torch.aten.mm"
// CHECK: %[[INIT_TENSOR:.*]] = linalg.init_tensor [%[[LHS_DIM_0]], %[[RHS_DIM_1]]] : tensor<?x?xf32>
// CHECK: %[[CF0:.*]] = arith.constant 0.000000e+00 : f32
// CHECK: %[[ZEROFILL:.*]] = linalg.fill ins(%[[CF0]] : f32) outs(%[[INIT_TENSOR]] : tensor<?x?xf32>) -> tensor<?x?xf32>
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
// CHECK: %[[MATMUL:.*]] = linalg.matmul ins(%[[LHS]], %[[RHS]] : tensor<?x?xf32>, tensor<?x?xf32>) outs(%[[ZEROFILL]] : tensor<?x?xf32>) -> tensor<?x?xf32>
// CHECK: %[[CASTED:.*]] = tensor.cast %[[MATMUL]] : tensor<?x?xf32> to tensor<?x2xf32>
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_VTENSOR:.*]] = torch_c.from_builtin_tensor %[[CASTED]] : tensor<?x2xf32> -> !torch.vtensor<[?,2],f32>
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// CHECK: return %[[RESULT_VTENSOR]] : !torch.vtensor<[?,2],f32>
func.func @torch.aten.mm$basic(%arg0: !torch.vtensor<[?,?],f32>, %arg1: !torch.vtensor<[?,?],f32>) -> !torch.vtensor<[?,2],f32> {
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
%0 = torch.aten.mm %arg0, %arg1 : !torch.vtensor<[?,?],f32>, !torch.vtensor<[?,?],f32> -> !torch.vtensor<[?,2],f32>
return %0 : !torch.vtensor<[?,2],f32>
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
}
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// -----
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
// If the operands are missing dtype, we cannot lower it.
func.func @torch.aten.mm$no_convert$missing_dtype(%arg0: !torch.vtensor, %arg1: !torch.vtensor) -> !torch.vtensor {
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// expected-error@+1 {{failed to legalize}}
%0 = torch.aten.mm %arg0, %arg1 : !torch.vtensor, !torch.vtensor -> !torch.vtensor
return %0 : !torch.vtensor
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
}
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// -----
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
// Correctly handle the case that operands are statically the wrong rank
// (rank 1 vs rank 2 expected for matmul.)
func.func @torch.aten.mm$no_convert$wrong_rank(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vtensor<[?],f32>) -> !torch.vtensor<[?,?],f32> {
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// expected-error@+1 {{failed to legalize}}
%0 = torch.aten.mm %arg0, %arg1 : !torch.vtensor<[?],f32>, !torch.vtensor<[?],f32> -> !torch.vtensor<[?,?],f32>
return %0 : !torch.vtensor<[?,?],f32>
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
}
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// -----
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// If the result is missing dtype, we cannot lower it.
func.func @torch.aten.mm$no_convert$result_missing_dtype(%arg0: !torch.vtensor<[?,?],f32>, %arg1: !torch.vtensor<[?,?],f32>) -> !torch.vtensor {
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// expected-error@+1 {{failed to legalize}}
%0 = torch.aten.mm %arg0, %arg1 : !torch.vtensor<[?,?],f32>, !torch.vtensor<[?,?],f32> -> !torch.vtensor
return %0 : !torch.vtensor
Significantly restructure torch/aten import design. This is a really major and invasive restructuring of the way we get torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into MLIR. Please forgive the challenging review, but due to the sheer invasiveness, it wasn't really practical do do it in sane smaller pieces. This fully replaces everything that was already working on the TorchScript path (actually, more -- we added tanh support to TorchToLinalg in order to delete the older code paths). Additionally, I've kept the lights on for the acap path too, including what little e2e stuff was working before (for expediency I made a few tiny compromises along the way that will be easy to undo when we give that path proper attention). Overview of the new design: - The torch operator `somens::someunqualname.someoverloadname` is imported as `torch.somens.someunqualname.someoverloadname` (skip the last dotted part if the overload name is empty), OR, if we don't have such an op registered, it is imported as `torch.operator "somens.someunqualname.someoverloadname" (...) : ...`. - The addition of the "overload name" is a critical element here, as the `(ns,unqual,overload)` triple is unique, which solves a lot of problems we were having. - This involves having separate MLIR ops for the `trailing_` and `.out` variants and all the different overloads. This seemed necessary, because the set of overloads is so wild and varied and unstructured. The previous design was leaning into some underlying structure that just isn't there -- the default situation is the "random overload that we want to manage on the MLIR side", rather than that being an exception. E.g. `aten::ne` (not-equal) has 21 overloads, only 4 of which are c10 dispatcher ops see [gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1), and the "out" variant is really called `.Tensor_out` instead of `.out` as it frequently is for other ops. - Rationale for all being in `torch` namespace: the set of operators are so varied and unstructured that "dialect per namespace" doesn't result in anything resembling the typical MLIR dialect boundary expectations. We could maybe draw the boundary at dispatcher ops vs non-dispatcher ops, but that doesn't seem to really result in very much useful structure at this point in time. - Note: within the torch operator registry, we effectively have a mini-basicpy subdialect (already type-resolved), which is reasonably structured. - The existing Torch op interfaces are also removed -- now that we track the overload name, we can losslessly find the original operator. - Instead of `ATenRecognizeKernelsPass`, we now have a `ReduceOpVariantsPass` that keys off certain traits (and perhaps eventually interfaces) to reduce variants of ops to a smaller set, ideally operating on immutable tensors and using surrounding ops to model the mutability/aliasing aspects. - Note: `torch.ns.unqual.overload` ops allow both immutable and mutable tensors (unlike the previous hard distinction in the common case). This is a premonition for a future change that will introduce a bona fide `!torch.tensor` type that will clean up a bunch of stuff. - `TorchToLinalg` / `TorchToStd` supercede the existing "ATen->TCF->TCP->Linalg" path. - The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`. It should look somewhat familiar, but the benefit of hindsight has allowed a lot of simplifications. The overall trend seems to be to make the `torch` dialect a nice layer independent of anything else. It feels like as a natural result of various future changes we will be removing the reliance on basicpy+numpy dialects and have a nice self-contained type system too that properly models the TorchScript type system (including proper subtyping, mutable/immutable tensors, optional dtype, etc.). Recommended review order: - Start at some of the new import IR, e.g. in `frontends/pytorch/test/node_import/prim.py`, `frontends/pytorch/test/acap_export/test_export_add3.py`, and other tests. - `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py` and associated generated files: - `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td` - `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td` - Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h` - Various code changes in the import path in `frontends/pytorch/csrc/builder`. Probably most interesting is the new code in `torch_to_mlir_utils.cpp` that has the logic to create the `torch.operator` ops or `torch.ns.unqual.overload` ops. This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe), just to be able to look at a substantial sample of IR in the new style.
2021-05-05 05:42:50 +08:00
}
// -----
// CHECK-LABEL: func.func @torch.aten.Int.Tensor$zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[],si64>) -> !torch.int {
// CHECK: %[[I:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[],si64> -> tensor<i64>
// CHECK: %[[EXT:.*]] = tensor.extract %[[I]][] : tensor<i64>
// CHECK: %[[RET:.*]] = torch_c.from_i64 %[[EXT]]
// CHECK: return %[[RET]] : !torch.int
func.func @torch.aten.Int.Tensor$zero_rank(%arg0: !torch.vtensor<[],si64>) -> !torch.int {
%0 = torch.aten.Int.Tensor %arg0 : !torch.vtensor<[],si64> -> !torch.int
return %0 : !torch.int
}
// -----
// CHECK-LABEL: func.func @torch.aten.Int.Tensor$non_zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[?,?],si64>) -> !torch.int {
// CHECK: %[[I:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[?,?],si64> -> tensor<?x?xi64>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM0:.*]] = tensor.dim %[[I]], %[[C0]] : tensor<?x?xi64>
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[DIM1:.*]] = tensor.dim %[[I]], %[[C1]] : tensor<?x?xi64>
// CHECK: %[[ONE:.*]] = arith.constant 1 : i64
// CHECK: %[[DIM0_INDEX:.*]] = arith.index_cast %[[DIM0]] : index to i64
// CHECK: %[[PRED0:.*]] = arith.cmpi eq, %[[DIM0_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED0]], "mismatching contracting dimension"
// CHECK: %[[DIM1_INDEX:.*]] = arith.index_cast %[[DIM1]] : index to i64
// CHECK: %[[PRED1:.*]] = arith.cmpi eq, %[[DIM1_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED1]], "mismatching contracting dimension"
// CHECK: %[[ZERO:.*]] = arith.constant 0 : index
// CHECK: %[[EXT:.*]] = tensor.extract %[[I]][%[[ZERO]], %[[ZERO]]] : tensor<?x?xi64>
// CHECK: %[[RET:.*]] = torch_c.from_i64 %[[EXT]]
// CHECK: return %[[RET]] : !torch.int
func.func @torch.aten.Int.Tensor$non_zero_rank(%arg0: !torch.vtensor<[?,?],si64>) -> !torch.int {
%0 = torch.aten.Int.Tensor %arg0 : !torch.vtensor<[?,?],si64> -> !torch.int
return %0 : !torch.int
}
// -----
// CHECK-LABEL: func.func @torch.aten.Float.Tensor$zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[],f64>) -> !torch.float {
// CHECK: %[[F:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[],f64> -> tensor<f64>
// CHECK: %[[EXT:.*]] = tensor.extract %[[F]][] : tensor<f64>
// CHECK: %[[RET:.*]] = torch_c.from_f64 %[[EXT]]
// CHECK: return %[[RET]] : !torch.float
func.func @torch.aten.Float.Tensor$zero_rank(%arg0: !torch.vtensor<[],f64>) -> !torch.float {
%0 = torch.aten.Float.Tensor %arg0 : !torch.vtensor<[],f64> -> !torch.float
return %0 : !torch.float
}
// -----
// CHECK-LABEL: func.func @torch.aten.Float.Tensor$non_zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[?,?],f64>) -> !torch.float {
// CHECK: %[[F:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[?,?],f64> -> tensor<?x?xf64>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM0:.*]] = tensor.dim %[[F]], %[[C0]] : tensor<?x?xf64>
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[DIM1:.*]] = tensor.dim %[[F]], %[[C1]] : tensor<?x?xf64>
// CHECK: %[[ONE:.*]] = arith.constant 1 : i64
// CHECK: %[[DIM0_INDEX:.*]] = arith.index_cast %[[DIM0]] : index to i64
// CHECK: %[[PRED0:.*]] = arith.cmpi eq, %[[DIM0_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED0]], "mismatching contracting dimension"
// CHECK: %[[DIM1_INDEX:.*]] = arith.index_cast %[[DIM1]] : index to i64
// CHECK: %[[PRED1:.*]] = arith.cmpi eq, %[[DIM1_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED1]], "mismatching contracting dimension"
// CHECK: %[[ZERO:.*]] = arith.constant 0 : index
// CHECK: %[[EXT:.*]] = tensor.extract %[[F]][%[[ZERO]], %[[ZERO]]] : tensor<?x?xf64>
// CHECK: %[[RET:.*]] = torch_c.from_f64 %[[EXT]]
// CHECK: return %[[RET]] : !torch.float
func.func @torch.aten.Float.Tensor$non_zero_rank(%arg0: !torch.vtensor<[?,?],f64>) -> !torch.float {
%0 = torch.aten.Float.Tensor %arg0 : !torch.vtensor<[?,?],f64> -> !torch.float
return %0 : !torch.float
}
// -----
// CHECK-LABEL: func.func @torch.aten.Bool.Tensor$zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[],i1>) -> !torch.bool {
// CHECK: %[[B:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[],i1> -> tensor<i1>
// CHECK: %[[EXT:.*]] = tensor.extract %[[B]][] : tensor<i1>
// CHECK: %[[RES:.*]] = torch_c.from_i1 %[[EXT]]
// CHECK: return %[[RES]] : !torch.bool
func.func @torch.aten.Bool.Tensor$zero_rank(%arg0: !torch.vtensor<[],i1>) -> !torch.bool {
%0 = torch.aten.Bool.Tensor %arg0 : !torch.vtensor<[],i1> -> !torch.bool
return %0 : !torch.bool
}
// -----
// CHECK-LABEL: func.func @torch.aten.Bool.Tensor$non_zero_rank
// CHECK-SAME: (%[[ARG:.*]]: !torch.vtensor<[?,?],i1>) -> !torch.bool {
// CHECK: %[[B:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<[?,?],i1> -> tensor<?x?xi1>
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM0:.*]] = tensor.dim %[[B]], %[[C0]] : tensor<?x?xi1>
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[DIM1:.*]] = tensor.dim %[[B]], %[[C1]] : tensor<?x?xi1>
// CHECK: %[[ONE:.*]] = arith.constant 1 : i64
// CHECK: %[[DIM0_INDEX:.*]] = arith.index_cast %[[DIM0]] : index to i64
// CHECK: %[[PRED0:.*]] = arith.cmpi eq, %[[DIM0_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED0]], "mismatching contracting dimension"
// CHECK: %[[DIM1_INDEX:.*]] = arith.index_cast %[[DIM1]] : index to i64
// CHECK: %[[PRED1:.*]] = arith.cmpi eq, %[[DIM1_INDEX]], %[[ONE]] : i64
// CHECK: assert %[[PRED1]], "mismatching contracting dimension"
// CHECK: %[[ZERO:.*]] = arith.constant 0 : index
// CHECK: %[[EXT:.*]] = tensor.extract %[[I]][%[[ZERO]], %[[ZERO]]] : tensor<?x?xi1>
// CHECK: %[[RET:.*]] = torch_c.from_i1 %[[EXT]]
// CHECK: return %[[RET]] : !torch.bool
func.func @torch.aten.Bool.Tensor$non_zero_rank(%arg0: !torch.vtensor<[?,?],i1>) -> !torch.bool {
%0 = torch.aten.Bool.Tensor %arg0 : !torch.vtensor<[?,?],i1> -> !torch.bool
return %0 : !torch.bool
}
2021-11-09 04:28:51 +08:00
// -----
// CHECK: func.func @torch.prim.NumToTensor.Scalar$basic(%[[IN:.*]]: !torch.int) -> !torch.vtensor<[],si64> {
2021-11-09 04:28:51 +08:00
// CHECK: %[[INI64:.*]] = torch_c.to_i64 %[[IN]]
// CHECK: %[[NEWVEC:.*]] = linalg.init_tensor [] : tensor<i64>
// CHECK: %[[FILLVEC:.*]] = linalg.fill ins(%[[INI64]] : i64) outs(%[[NEWVEC]] : tensor<i64>) -> tensor<i64>
2021-11-09 04:28:51 +08:00
// CHECK: %[[OUTVEC:.*]] = torch_c.from_builtin_tensor %[[FILLVEC]] : tensor<i64> -> !torch.vtensor<[],si64>
// CHECK: return %[[OUTVEC]] : !torch.vtensor<[],si64>
func.func @torch.prim.NumToTensor.Scalar$basic(%arg0: !torch.int) -> !torch.vtensor<[],si64> {
2021-11-09 04:28:51 +08:00
%0 = torch.prim.NumToTensor.Scalar %arg0 : !torch.int -> !torch.vtensor<[],si64>
return %0 : !torch.vtensor<[],si64>
}
// -----
// CHECK-LABEL: func.func @torch.tensor_static_info_cast$basic(
// CHECK-SAME: %[[VALUE_T:.*]]: !torch.vtensor<[?],f32>) -> !torch.vtensor<[4],f32> {
// CHECK: %[[T:.*]] = torch_c.to_builtin_tensor %[[VALUE_T]] : !torch.vtensor<[?],f32> -> tensor<?xf32>
// CHECK: %[[T_CAST:.*]] = tensor.cast %[[T]] : tensor<?xf32> to tensor<4xf32>
// CHECK: %[[VALUE_T_CAST:.*]] = torch_c.from_builtin_tensor %[[T_CAST]] : tensor<4xf32> -> !torch.vtensor<[4],f32>
// CHECK: return %[[VALUE_T_CAST]] : !torch.vtensor<[4],f32>
func.func @torch.tensor_static_info_cast$basic(%t: !torch.vtensor<[?],f32>) -> !torch.vtensor<[4],f32> {
%t_cast = torch.tensor_static_info_cast %t : !torch.vtensor<[?],f32> to !torch.vtensor<[4],f32>
return %t_cast : !torch.vtensor<[4],f32>
}
// -----
// CHECK-LABEL: func.func @torch.aten.neg
// CHECK: linalg.generic {{.*}} {
// CHECK-NEXT: ^bb0(%[[LHS:.*]]: f32, %{{.*}}: f32):
// CHECK-NEXT: %[[NEG:.*]] = arith.negf %[[LHS]] : f32
// CHECK-NEXT: linalg.yield %[[NEG]] : f32
// CHECK-NEXT: } -> tensor<?x?xf32>
func.func @torch.aten.neg(%arg0: !torch.vtensor<[?,?],f32>) -> !torch.vtensor<[?,?],f32> {
%0 = torch.aten.neg %arg0 : !torch.vtensor<[?,?],f32> -> !torch.vtensor<[?,?],f32>
return %0 : !torch.vtensor<[?,?],f32>
}
// -----
// CHECK-LABEL: func.func @torch.aten.neg.bf16
// CHECK: linalg.generic {{.*}} {
// CHECK-NEXT: ^bb0(%[[LHS:.*]]: bf16, %{{.*}}: bf16):
// CHECK-NEXT: %[[NEG:.*]] = arith.negf %[[LHS]] : bf16
// CHECK-NEXT: linalg.yield %[[NEG]] : bf16
// CHECK-NEXT: } -> tensor<?x?xbf16>
func.func @torch.aten.neg.bf16(%arg0: !torch.vtensor<[?,?],bf16>) -> !torch.vtensor<[?,?],bf16> {
%0 = torch.aten.neg %arg0 : !torch.vtensor<[?,?],bf16> -> !torch.vtensor<[?,?],bf16>
return %0 : !torch.vtensor<[?,?],bf16>
}