torch-mlir/lib/Dialect/Torch/IR/TorchOps.cpp

1213 lines
45 KiB
C++
Raw Normal View History

2020-09-29 03:02:35 +08:00
//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Also available under a BSD-style license. See LICENSE.
2020-09-29 03:02:35 +08:00
//
//===----------------------------------------------------------------------===//
[torch-mlir earthmoving (1/N)] C/C++ code movement. This creates the `external/torch-mlir` directory as an LLVM_EXTERNAL_PROJECTS-compatible project (analogous to `iree-dialects`) and completes movement/rename of all pure MLIR C/C++ compiler code into there. The next step will be to move all the Python code / code that links/includes PyTorch C++ code (which currently lives in `frontends/pytorch`) into a subdirectory here. I call this "earthmoving" because it is mostly mechanical changes and renames. As a quick summary (we can change this down the road easily) - C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch` - CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet` - preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_` - CMake `NPCOMPFoo -> TorchMLIRFoo` The goal of this is to create a standalone project creating a center of mass for entry into the MLIR ecosystem from PyTorch, suitable in scope for eventual inclusion/ownership in PyTorch. The idea is that `external/torch-mlir` will some day be pulled out into its own repository, and then npcomp will simply pull it in as a submodule. Layering-wise, what lives in `torch-mlir` lowers code from PyTorch (currently TorchScript, but TorchFX or pytorch/xla-style tracing are possible extensions) down to what we have been calling the "Torch backend contract" which is cleaned up IR (inlining, simplifcation, conversion to value tensors, ...) entirely in the `torch` dialect. This is the branching off point for further lowering, of which npcomp takes one opinion (outside `torch-mlir` of course!), namely the `TorchConversion` dialect/transforms which lower to IR suitable for IREE and other linalg-on-tensors based lower-level compilers. Summary of changes: - move `{include,lib,test}/Dialect/Torch` into `torch-mlir` - move relevant parts of CAPI into `torch-mlir`. - leave a few things related to the `torch-mlir` Python build commented out, which should be resolved in a subsequent change.
2021-09-10 03:24:10 +08:00
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
2020-09-29 03:02:35 +08:00
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/PatternMatch.h"
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
#include "mlir/IR/TypeUtilities.h"
#include "mlir/Support/LLVM.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/Support/Casting.h"
2020-09-29 03:02:35 +08:00
using namespace mlir;
[torch-mlir earthmoving (1/N)] C/C++ code movement. This creates the `external/torch-mlir` directory as an LLVM_EXTERNAL_PROJECTS-compatible project (analogous to `iree-dialects`) and completes movement/rename of all pure MLIR C/C++ compiler code into there. The next step will be to move all the Python code / code that links/includes PyTorch C++ code (which currently lives in `frontends/pytorch`) into a subdirectory here. I call this "earthmoving" because it is mostly mechanical changes and renames. As a quick summary (we can change this down the road easily) - C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch` - CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet` - preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_` - CMake `NPCOMPFoo -> TorchMLIRFoo` The goal of this is to create a standalone project creating a center of mass for entry into the MLIR ecosystem from PyTorch, suitable in scope for eventual inclusion/ownership in PyTorch. The idea is that `external/torch-mlir` will some day be pulled out into its own repository, and then npcomp will simply pull it in as a submodule. Layering-wise, what lives in `torch-mlir` lowers code from PyTorch (currently TorchScript, but TorchFX or pytorch/xla-style tracing are possible extensions) down to what we have been calling the "Torch backend contract" which is cleaned up IR (inlining, simplifcation, conversion to value tensors, ...) entirely in the `torch` dialect. This is the branching off point for further lowering, of which npcomp takes one opinion (outside `torch-mlir` of course!), namely the `TorchConversion` dialect/transforms which lower to IR suitable for IREE and other linalg-on-tensors based lower-level compilers. Summary of changes: - move `{include,lib,test}/Dialect/Torch` into `torch-mlir` - move relevant parts of CAPI into `torch-mlir`. - leave a few things related to the `torch-mlir` Python build commented out, which should be resolved in a subsequent change.
2021-09-10 03:24:10 +08:00
using namespace mlir::torch;
using namespace mlir::torch::Torch;
// see https://github.com/pytorch/pytorch/blob/master/c10/core/ScalarType.h#L28
static int64_t getDtypeIntegerFromMlirType(Type dtype) {
if (dtype.isa<Float32Type>())
return 6;
if (auto integerType = dtype.dyn_cast<IntegerType>()) {
if (integerType.isSignedInteger(64))
return 4;
if (integerType.isSignlessInteger(1))
return 11;
}
return -1;
}
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
//===----------------------------------------------------------------------===//
// Utilities
//===----------------------------------------------------------------------===//
[torch-mlir earthmoving (1/N)] C/C++ code movement. This creates the `external/torch-mlir` directory as an LLVM_EXTERNAL_PROJECTS-compatible project (analogous to `iree-dialects`) and completes movement/rename of all pure MLIR C/C++ compiler code into there. The next step will be to move all the Python code / code that links/includes PyTorch C++ code (which currently lives in `frontends/pytorch`) into a subdirectory here. I call this "earthmoving" because it is mostly mechanical changes and renames. As a quick summary (we can change this down the road easily) - C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch` - CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet` - preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_` - CMake `NPCOMPFoo -> TorchMLIRFoo` The goal of this is to create a standalone project creating a center of mass for entry into the MLIR ecosystem from PyTorch, suitable in scope for eventual inclusion/ownership in PyTorch. The idea is that `external/torch-mlir` will some day be pulled out into its own repository, and then npcomp will simply pull it in as a submodule. Layering-wise, what lives in `torch-mlir` lowers code from PyTorch (currently TorchScript, but TorchFX or pytorch/xla-style tracing are possible extensions) down to what we have been calling the "Torch backend contract" which is cleaned up IR (inlining, simplifcation, conversion to value tensors, ...) entirely in the `torch` dialect. This is the branching off point for further lowering, of which npcomp takes one opinion (outside `torch-mlir` of course!), namely the `TorchConversion` dialect/transforms which lower to IR suitable for IREE and other linalg-on-tensors based lower-level compilers. Summary of changes: - move `{include,lib,test}/Dialect/Torch` into `torch-mlir` - move relevant parts of CAPI into `torch-mlir`. - leave a few things related to the `torch-mlir` Python build commented out, which should be resolved in a subsequent change.
2021-09-10 03:24:10 +08:00
Value mlir::torch::Torch::copyTensorToType(OpBuilder &builder, Location loc,
BaseTensorType newType,
Value tensor) {
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
auto originalType = tensor.getType().cast<BaseTensorType>();
// Adjust the static information in the type to match between the original and
// new types.
if (!originalType.hasSameSizesAndDtype(newType)) {
tensor = builder.create<TensorStaticInfoCastOp>(
loc, originalType.getWithSizesAndDtypeFrom(newType), tensor);
}
// Unless both the original and new types are both value tensors, we end
// up creating one op that converts between the value and non-value tensor
// domains. If both the original and new types are both non-value tensors,
// then we do the copy by going to a value tensor and back.
if (tensor.getType().isa<NonValueTensorType>())
tensor = builder.create<CopyToValueTensorOp>(loc, tensor);
if (newType.isa<NonValueTensorType>())
tensor = builder.create<CopyToNonValueTensorOp>(loc, tensor);
return tensor;
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
}
static IntegerAttr getI64IntegerAttr(MLIRContext *context, int64_t value) {
return IntegerAttr::get(IntegerType::get(context, 64), value);
}
//===----------------------------------------------------------------------===//
// MethodOp
//===----------------------------------------------------------------------===//
LogicalResult MethodOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
auto func =
symbolTable.lookupNearestSymbolFrom<FuncOp>(*this, functionAttr());
if (!func)
return emitError() << "'@" << function()
<< "' does not reference a valid function";
if (func.getVisibility() != SymbolTable::Visibility::Private)
return emitError() << "'@" << function()
<< "' must reference a private function";
if (func.isDeclaration())
return emitError() << "'@" << function()
<< "' must reference a function that is defined (not "
"merely declared)";
auto expectedReceiverArgType = NnModuleType::get(
getContext(), getOperation()->getParentOfType<ClassTypeOp>().getName());
if (func.getType().getNumInputs() == 0 ||
func.getType().getInput(0) != expectedReceiverArgType) {
return emitError() << "the referenced function '" << function()
<< "' must have a first argument of type "
<< expectedReceiverArgType;
}
return success();
}
//===----------------------------------------------------------------------===//
// NnModuleOp
//===----------------------------------------------------------------------===//
static LogicalResult verify(NnModuleOp op) {
for (Operation &child : *op.getBody())
if (!isa<SlotOp, NnModuleTerminatorOp>(&child))
return child.emitOpError() << "is not allowed inside 'torch.nn_module'";
return success();
}
// PyTorch has a well-developed notion of subtyping.
//
// This is a restricted subset of it.
//
// TODO: Flesh this out.
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
// TODO: Decide / properly model the distinction between PEP 483 / Python
// subtyping vs "more static information".
bool isValidSubtype(Type subtype, Type type) {
if (subtype == type)
return true;
if (auto any = type.dyn_cast<AnyType>())
return true;
if (auto number = type.dyn_cast<NumberType>())
return subtype.isa<IntType>() || subtype.isa<Torch::FloatType>();
if (auto optional = type.dyn_cast<OptionalType>())
return isValidSubtype(subtype, optional.getContainedType()) ||
subtype.isa<Torch::NoneType>();
if (auto tuple = type.dyn_cast<Torch::TupleType>()) {
if (!subtype.isa<Torch::TupleType>())
return false;
auto subtypes = subtype.cast<Torch::TupleType>().getContainedTypes();
auto types = tuple.getContainedTypes();
if (subtypes.size() != types.size())
return false;
for (auto t : llvm::zip(subtypes, types)) {
if (!isValidSubtype(std::get<0>(t), std::get<1>(t)))
return false;
}
return true;
}
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
// TODO: This is not subtyping according to PEP 483. See description
// of NonValueTensorType.
if (subtype.isa<NonValueTensorType>() && type.isa<NonValueTensorType>() &&
type ==
NonValueTensorType::getWithLeastStaticInformation(type.getContext()))
return true;
if (subtype.isa<ValueTensorType>() && type.isa<ValueTensorType>() &&
type == ValueTensorType::getWithLeastStaticInformation(type.getContext()))
return true;
return false;
}
LogicalResult NnModuleOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
auto classType = symbolTable.lookupNearestSymbolFrom<ClassTypeOp>(
*this, SymbolRefAttr::get(getContext(), getClassName()));
if (!classType)
return emitError() << "'" << getClassName()
<< "' does not reference a valid class type";
auto attrs = llvm::to_vector<6>(getBody()->getOps<SlotOp>());
auto attrDefs = llvm::to_vector<6>(classType.getBody()->getOps<AttrOp>());
if (attrs.size() != attrDefs.size())
return emitError() << "number of 'torch.slot's in a 'torch.nn_module' must "
"match number of 'torch.attr's in "
"the corresponding 'torch.class_type'";
for (int i = 0, e = attrs.size(); i != e; i++) {
SlotOp attr = attrs[i];
AttrOp attrDef = attrDefs[i];
if (!isValidSubtype(attr.value().getType(), attrDef.type()) ||
attr.name() != attrDef.name()) {
return attr.emitOpError()
.append("is expected to match type and name of '",
attrDef.getOperation(), "'")
.attachNote(attrDef.getLoc())
.append("see torch.attr at corresponding index ", i, " here");
}
}
return success();
}
//===----------------------------------------------------------------------===//
// PrimListConstructOp
//===----------------------------------------------------------------------===//
static LogicalResult verify(PrimListConstructOp op) {
auto resultType = op.getResult().getType();
auto resultElementType = resultType.dyn_cast<ListType>().getContainedType();
auto matchResultElementType = [&](Type type) {
return isValidSubtype(type, resultElementType);
};
if (!llvm::all_of(op->getOperandTypes(), matchResultElementType)) {
return op.emitError() << "operand types should have the same type as the "
"list contained type";
}
return success();
}
//===----------------------------------------------------------------------===//
// PrimDictConstructOp
//===----------------------------------------------------------------------===//
static LogicalResult verify(PrimDictConstructOp op) {
auto isValidSubTypeOf = [](Type expectedType) {
return [=](Type type) { return isValidSubtype(type, expectedType); };
};
Type keyType = op.getKeyType();
if (!llvm::all_of(op.keys().getTypes(), isValidSubTypeOf(keyType)))
return op.emitError() << "keys should be of Dict key type";
Type valueType = op.getValueType();
if (!llvm::all_of(op.values().getTypes(), isValidSubTypeOf(valueType)))
return op.emitError() << "values should be of Dict value type";
return success();
}
//===----------------------------------------------------------------------===//
// ClassTypeOp
//===----------------------------------------------------------------------===//
static LogicalResult verify(ClassTypeOp op) {
llvm::StringMap<Operation *> namesToOps;
for (Operation &child : op.getBody()->without_terminator()) {
if (!isa<AttrOp, MethodOp>(&child))
return child.emitOpError() << "is not allowed inside `torch.class_type`";
StringRef name;
if (auto attr = dyn_cast<AttrOp>(child))
name = attr.name();
else
name = cast<MethodOp>(child).name();
auto itAndWasInserted = namesToOps.insert({name, &child});
auto it = itAndWasInserted.first;
bool wasInserted = itAndWasInserted.second;
if (!wasInserted) {
auto diag = op.emitOpError().append(
"has duplicate attr/method with name '", name, "'");
diag.attachNote(it->second->getLoc())
.append("see first conflicting attr/method here");
diag.attachNote(child.getLoc())
.append("see second conflicting attr/method here");
return failure();
}
}
return success();
}
//===----------------------------------------------------------------------===//
// PrimLoopOp
//===----------------------------------------------------------------------===//
OperandRange PrimLoopOp::getSuccessorEntryOperands(unsigned index) {
assert(index == 0);
return iterArgsInit();
}
void PrimLoopOp::getSuccessorRegions(
Optional<unsigned> index, ArrayRef<Attribute> operands,
SmallVectorImpl<RegionSuccessor> &regions) {
(void)operands;
if (!index.hasValue()) {
regions.emplace_back(&region(), region().getArguments().slice(1));
return;
}
assert(*index == 0);
regions.emplace_back(&region(), region().getArguments().slice(1));
regions.emplace_back(getResults());
}
//===----------------------------------------------------------------------===//
// PrimLoopConditionOp
//===----------------------------------------------------------------------===//
MutableOperandRange
PrimLoopConditionOp::getMutableSuccessorOperands(Optional<unsigned> index) {
// Pass all operands except the condition to the successor which is the
// parent loop op.
return iterArgsMutable();
}
//===----------------------------------------------------------------------===//
// PrimIfOp
//===----------------------------------------------------------------------===//
static ParseResult parsePrimIfOp(OpAsmParser &parser, OperationState &result) {
// Create the regions.
result.regions.reserve(2);
Region *thenRegion = result.addRegion();
Region *elseRegion = result.addRegion();
auto &builder = parser.getBuilder();
OpAsmParser::OperandType cond;
Type boolType = builder.getType<Torch::BoolType>();
if (parser.parseOperand(cond) ||
parser.resolveOperand(cond, boolType, result.operands))
return failure();
// Parse results type list.
if (parser.parseArrowTypeList(result.types))
return failure();
// Parse the 'then' region.
if (parser.parseRegion(*thenRegion, /*arguments=*/{}, /*argTypes=*/{}))
return failure();
// Parse the 'else' region.
if (parser.parseKeyword("else"))
return failure();
if (parser.parseRegion(*elseRegion, /*arguments=*/{}, /*argTypes=*/{}))
return failure();
// Parse the optional attribute list.
if (parser.parseOptionalAttrDict(result.attributes))
return failure();
return success();
}
static void print(OpAsmPrinter &p, PrimIfOp op) {
p << " " << op.condition();
p << " -> (" << op.getResultTypes() << ")";
p.printRegion(op.thenRegion(), /*printEntryBlockArgs=*/false);
p << " else";
p.printRegion(op.elseRegion(), /*printEntryBlockArgs=*/false);
p.printOptionalAttrDict(op->getAttrs());
}
void PrimIfOp::getSuccessorRegions(Optional<unsigned> index,
ArrayRef<Attribute> operands,
SmallVectorImpl<RegionSuccessor> &regions) {
// The `then` and the `else` region branch back to the parent operation.
if (index.hasValue()) {
regions.push_back(RegionSuccessor(getResults()));
return;
}
// If the condition is constant, we can give a more precise answer.
if (auto condAttr = operands.front().dyn_cast_or_null<IntegerAttr>()) {
Region *executedRegion =
condAttr.getValue().isOneValue() ? &thenRegion() : &elseRegion();
regions.push_back(RegionSuccessor(executedRegion));
return;
}
// If the condition isn't constant, both regions may be executed.
regions.push_back(RegionSuccessor(&thenRegion()));
regions.push_back(RegionSuccessor(&elseRegion()));
return;
}
/// Replaces the given op with the contents of the given single-block region,
/// using the operands of the block terminator to replace operation results.
static void replaceOpWithRegion(PatternRewriter &rewriter, Operation *op,
Region &region, ValueRange blockArgs = {}) {
assert(llvm::hasSingleElement(region) && "expected single-region block");
Block *block = &region.front();
Operation *terminator = block->getTerminator();
ValueRange results = terminator->getOperands();
rewriter.mergeBlockBefore(block, op, blockArgs);
rewriter.replaceOp(op, results);
rewriter.eraseOp(terminator);
}
void PrimIfOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
MLIRContext *context) {
// If the condition is constant, delete the dead branch and inline the live
// branch.
patterns.add(+[](PrimIfOp op, PatternRewriter &rewriter) {
auto constantBool = op.condition().getDefiningOp<Torch::ConstantBoolOp>();
if (!constantBool)
return rewriter.notifyMatchFailure(op, "non-constant condition");
replaceOpWithRegion(
rewriter, op, constantBool.value() ? op.thenRegion() : op.elseRegion());
return success();
});
}
//===----------------------------------------------------------------------===//
// DerefineOp
//===----------------------------------------------------------------------===//
bool DerefineOp::areCastCompatible(mlir::TypeRange inputs,
mlir::TypeRange outputs) {
return isValidSubtype(inputs[0], outputs[0]);
}
template <typename OpTy>
static OpFoldResult atenIsOrIsNotFoldHelper(OpTy op, bool equalIsTrue) {
Type lhsType = op.self().getType();
Type rhsType = op.obj().getType();
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 either type is a NoneType, make it be the lhsType.
if (rhsType.template isa<Torch::NoneType>())
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
std::swap(lhsType, rhsType);
// TODO: Implement and use subtype infra for this.
// If neither type is a subtype of the other, then the result is false.
if (lhsType.template isa<Torch::NoneType>() &&
rhsType.template isa<Torch::NoneType>())
return IntegerAttr::get(IntegerType::get(op.getContext(), 1), equalIsTrue);
if (lhsType.template isa<Torch::NoneType>() &&
!rhsType.template isa<Torch::OptionalType>())
return IntegerAttr::get(IntegerType::get(op.getContext(), 1), !equalIsTrue);
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
return nullptr;
}
//===----------------------------------------------------------------------===//
// Aten__Is__Op
//===----------------------------------------------------------------------===//
OpFoldResult Aten__Is__Op::fold(ArrayRef<Attribute> operands) {
return atenIsOrIsNotFoldHelper(*this, /*equalIsTrue=*/true);
}
//===----------------------------------------------------------------------===//
// Aten__Isnot__Op
//===----------------------------------------------------------------------===//
OpFoldResult Aten__Isnot__Op::fold(ArrayRef<Attribute> operands) {
return atenIsOrIsNotFoldHelper(*this, /*equalIsTrue=*/false);
}
//===----------------------------------------------------------------------===//
// Aten__Not__Op
//===----------------------------------------------------------------------===//
OpFoldResult Aten__Not__Op::fold(ArrayRef<Attribute> operands) {
bool value;
if (!matchPattern(getOperand(), m_TorchConstantBool(&value)))
return nullptr;
return IntegerAttr::get(IntegerType::get(getContext(), 1), !value);
}
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
//===----------------------------------------------------------------------===//
2021-10-21 23:50:01 +08:00
// AtenNeBoolOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenNeBoolOp::fold(ArrayRef<Attribute> operands) {
if (getOperand(0) == getOperand(1))
return IntegerAttr::get(IntegerType::get(getContext(), 1), false);
bool a, b;
if (!matchPattern(getOperand(0), m_TorchConstantBool(&a)))
return nullptr;
if (!matchPattern(getOperand(1), m_TorchConstantBool(&b)))
return nullptr;
return IntegerAttr::get(IntegerType::get(getContext(), 1), a != b);
}
//===----------------------------------------------------------------------===//
// AtenSqueezeOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenSqueezeOp::fold(ArrayRef<Attribute> operands) {
if (auto tensorType = getOperand().getType().dyn_cast<BaseTensorType>()) {
if (tensorType.hasSizes() && tensorType.getSizes().size() == 0)
return getOperand();
}
return nullptr;
}
//===----------------------------------------------------------------------===//
// AtenSqueezeDimOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenSqueezeDimOp::fold(ArrayRef<Attribute> operands) {
if (auto tensorType = getOperand(0).getType().dyn_cast<BaseTensorType>()) {
if (tensorType.hasSizes() && tensorType.getSizes().size() == 0)
return getOperand(0);
}
return nullptr;
}
//===----------------------------------------------------------------------===//
// AtenToDtypeOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenToDtypeOp::fold(ArrayRef<Attribute> operands) {
bool nonBlocking, copyArg;
// The non_blocking arg must be `False`.
if (!matchPattern(non_blocking(), m_TorchConstantBool(&nonBlocking)) ||
nonBlocking)
return nullptr;
// The copy arg must be `False`.
if (!matchPattern(copy(), m_TorchConstantBool(&copyArg)) || copyArg)
return nullptr;
// The memory_format arg must be `none`.
if (!memory_format().getType().isa<Torch::NoneType>())
return nullptr;
auto inputType = getOperand(0).getType().dyn_cast<BaseTensorType>();
if (!inputType || !inputType.hasSizes())
return nullptr;
auto resType = getType().dyn_cast<BaseTensorType>();
if (!resType || !resType.hasSizes() || inputType != resType)
return nullptr;
// Fold when both the input tensor and result are of the same type.
return getOperand(0);
}
//===----------------------------------------------------------------------===//
// AtenViewOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenViewOp::fold(ArrayRef<Attribute> operands) {
auto inputType = getOperand(0).getType().dyn_cast<BaseTensorType>();
if (!inputType || !inputType.hasSizes() || inputType.getSizes().size() != 1)
return nullptr;
auto resType = getType().dyn_cast<BaseTensorType>();
if (!resType || !resType.hasSizes() || resType.getSizes().size() != 1)
return nullptr;
// Fold when both the input tensor and result are unity rank tensors.
return getOperand(0);
}
2021-10-21 23:50:01 +08:00
//===----------------------------------------------------------------------===//
// AtenDimOp
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
OpFoldResult AtenDimOp::fold(ArrayRef<Attribute> operands) {
if (auto tensorType = getOperand().getType().dyn_cast<BaseTensorType>()) {
if (tensorType.hasSizes())
return IntegerAttr::get(IntegerType::get(getContext(), 64),
tensorType.getSizes().size());
}
return nullptr;
}
//===----------------------------------------------------------------------===//
// AtenLenTOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenLenTOp::fold(ArrayRef<Attribute> operands) {
// `len([1,1,1])` -> `3`
if (auto listConstruct =
getOperand().getDefiningOp<Torch::PrimListConstructOp>()) {
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
return IntegerAttr::get(IntegerType::get(getContext(), 64),
listConstruct.getNumOperands());
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
}
return nullptr;
}
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
void AtenLenTOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
MLIRContext *context) {
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
// `len(t.size())` -> `t.ndim`
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
patterns.add(+[](AtenLenTOp op, PatternRewriter &rewriter) {
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
auto size = op.getOperand().getDefiningOp<AtenSizeOp>();
if (!size)
return rewriter.notifyMatchFailure(op, "operand not AtenSizeOp");
rewriter.replaceOpWithNewOp<AtenDimOp>(op, size.getOperand());
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
return success();
});
}
//===----------------------------------------------------------------------===//
// AtenSizeOp
//===----------------------------------------------------------------------===//
void AtenSizeOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
MLIRContext *context) {
patterns.add(+[](AtenSizeOp op, PatternRewriter &rewriter) {
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
auto type = op.getOperand().getType().dyn_cast<BaseTensorType>();
if (!type || !type.areAllSizesKnown())
return rewriter.notifyMatchFailure(op, "all sizes not known");
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
SmallVector<Value> listElements;
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
for (int64_t size : type.getSizes()) {
listElements.push_back(rewriter.create<Torch::ConstantIntOp>(
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
op->getLoc(), rewriter.getI64IntegerAttr(size)));
}
rewriter.replaceOpWithNewOp<Torch::PrimListConstructOp>(
op, Torch::ListType::get(rewriter.getType<Torch::IntType>()),
listElements);
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
return success();
});
Introduce `!torch.tensor` / `!torch.vtensor` types. This removes our reliance on the numpy dialect and avoids our off-label use of the builtin tnesor type for modeling unknown dtypes. The `!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor. The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic tensor. The new types look as follows syntactically: ``` // Least-static-information, non-value-semantic tensor. !torch.tensor // Explicit form of least-static-information variant. !torch.tensor<*,unk> // Least-static-information, value-semantic tensor. !torch.vtensor // Explicit form of least-static-information variant. !torch.vtensor<*,unk> // Fixed-set of allowable element types, with first-class support for // Torch's frontend signedness semantics. !torch.tensor<*,si32> // First-class support for unknown dtypes. !torch.tensor<[?,?,?],unk> // Standard MLIR representation of `?` for unknown dimensions. !torch.tensor<[?,2,?,4],unk> // Statically shaped / dtyped example. !torch.vtensor<[1,2,3,4],f32> ``` This required fairly significant changes throughout the compiler, but overall it is a big cleanup. We now have a much clearer layering of "the Torch frontend lowering" vs "lowering to std + linalg + etc.". At the C++ level, there is `ValueTensorType`, `NonValueTensorType`. We also have a helper `BaseTensorType` (kind of like ShapedType) which interoperates with those two. Included changes: - New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for creating torch tensor literals in the frontend. - Consistently use signedness for the types (except i1 which I didn't touch -- we need to sort out the situation with !basicpy.BoolType there anyway so will be attending to that soon) - Frontend can annotate whether an argument to the function has value semantics. We currently require this, as our backend contract does not currently allow us to even model the non-value-semantic case. Before, the value-semantic assumption was randomly injected in the middle of the pass pipeline. - Move ArrayToTensor (now called MaximizeValueSemantics) and RefinePublicReturn passes to torch dialect. - The TorchToStd and TorchToLinalg passes are now type conversions from `!torch.vtensor` to `tensor` and use the dialect conversion infra. The overall conversion pipeline is set up following the best practices of the "Type Conversions the Not-So-Hard Way" talk. This required introducing `torch-func-builtin-tensorize` and `torch-finalizing-builtin-tensorize` passes analogous to the upstream bufferization passes with the corresponding names (mostly just copypasta from there). - Misc Torch-level canonicalizations -- we now cleanly layer the lowering to std later in the pipeline, so we are gradually lessening our reliance on random std constant folding before we get to that point. Recommended review order: - New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp - New ops in TorchOps.td / TorchOps.cpp - Less important / more mechanical stuff - Frontend changes. - Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
// One-off pattern to erase if dead.
// TODO: Use the effects infra to express the semantics of this op and enable
// a centralized "erase if dead" canonicalization.
// Specifically, we need to mark the op as only MemoryEffects::Allocate
// so that `mlir::wouldOpBeTriviallyDead` does the right thing.
patterns.add(+[](AtenSizeOp op, PatternRewriter &rewriter) {
if (!op.use_empty())
return failure();
rewriter.eraseOp(op);
return failure();
});
}
//===----------------------------------------------------------------------===//
// AtenSizeIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenSizeIntOp::fold(ArrayRef<Attribute> operands) {
auto type = getOperand(0).getType().dyn_cast<BaseTensorType>();
if (!type || !type.hasSizes())
return nullptr;
int64_t inputRank = type.getSizes().size();
int64_t dim;
if (!matchPattern(this->dim(), m_TorchConstantInt(&dim)))
return nullptr;
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank))
return nullptr;
if (type.getSizes()[dim] == kUnknownSize)
return nullptr;
return IntegerAttr::get(IntegerType::get(getContext(), 64),
type.getSizes()[dim]);
}
//===----------------------------------------------------------------------===//
// AtenGtIntOp
//===----------------------------------------------------------------------===//
static IntegerAttr getI1IntegerAttr(MLIRContext *context, bool value) {
return IntegerAttr::get(IntegerType::get(context, 1),
static_cast<int64_t>(value));
}
using ConstantIntComparator = std::function<bool(int64_t, int64_t)>;
template <typename OpTy>
static OpFoldResult comparatorFoldHelper(OpTy op,
ConstantIntComparator comparator) {
if (op.getOperand(0) == op.getOperand(1))
return getI1IntegerAttr(op.getContext(), comparator(0, 0));
int64_t lhs, rhs;
if (!matchPattern(op.getOperand(0), m_TorchConstantInt(&lhs)) ||
!matchPattern(op.getOperand(1), m_TorchConstantInt(&rhs)))
return nullptr;
return getI1IntegerAttr(op.getContext(), comparator(lhs, rhs));
}
//===----------------------------------------------------------------------===//
// AtenNeIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenNeIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a != b; });
}
//===----------------------------------------------------------------------===//
// AtenEqIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenEqIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a == b; });
}
//===----------------------------------------------------------------------===//
// AtenEqFloatOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenEqFloatOp::fold(ArrayRef<Attribute> operands) {
double lhs, rhs;
if (!matchPattern(getOperand(0), m_TorchConstantFloat(&lhs)) ||
!matchPattern(getOperand(1), m_TorchConstantFloat(&rhs)))
return nullptr;
return getI1IntegerAttr(getContext(), lhs == rhs);
}
//===----------------------------------------------------------------------===//
// AtenEqStrOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenEqStrOp::fold(ArrayRef<Attribute> operands) {
if (getOperand(0) == getOperand(1))
return getI1IntegerAttr(getContext(), true);
auto aStr = a().getDefiningOp<ConstantStrOp>();
auto bStr = b().getDefiningOp<ConstantStrOp>();
if (aStr && bStr)
return getI1IntegerAttr(getContext(), aStr == bStr);
return nullptr;
}
//===----------------------------------------------------------------------===//
// AtenLtIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenLtIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a < b; });
}
//===----------------------------------------------------------------------===//
// AtenLeIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenLeIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a <= b; });
}
//===----------------------------------------------------------------------===//
// AtenGtIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenGtIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a > b; });
}
//===----------------------------------------------------------------------===//
// AtenGeIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenGeIntOp::fold(ArrayRef<Attribute> operands) {
return comparatorFoldHelper(*this,
[](int64_t a, int64_t b) { return a >= b; });
}
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
//===----------------------------------------------------------------------===//
// NonValueTensorLiteralOp
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
//===----------------------------------------------------------------------===//
LogicalResult NonValueTensorLiteralOp::inferReturnTypes(
MLIRContext *context, Optional<Location> location, ValueRange operands,
DictionaryAttr attributes, RegionRange regions,
SmallVectorImpl<Type> &inferredReturnTypes) {
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
auto attr = attributes.get("value").dyn_cast_or_null<ElementsAttr>();
if (!attr)
return failure();
RankedTensorType tensorType = attr.getType().cast<RankedTensorType>();
NonValueTensorType returnType =
NonValueTensorType::get(tensorType.getContext(), tensorType.getShape(),
tensorType.getElementType());
inferredReturnTypes.push_back(returnType);
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
return success();
}
static bool areSizesAndDtypesCompatible(BaseTensorType a, BaseTensorType b) {
if (a.hasSizes() && b.hasSizes()) {
if (failed(verifyCompatibleShape(a.getSizes(), b.getSizes())))
return false;
}
if (a.hasDtype() && b.hasDtype()) {
if (a.getDtype() != b.getDtype())
return false;
}
return true;
}
bool NonValueTensorLiteralOp::isCompatibleReturnTypes(TypeRange inferred,
TypeRange actual) {
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 (!actual[0].isa<BaseTensorType>())
return false;
return areSizesAndDtypesCompatible(inferred[0].cast<BaseTensorType>(),
actual[0].cast<BaseTensorType>());
}
//===----------------------------------------------------------------------===//
// ValueTensorLiteralOp
//===----------------------------------------------------------------------===//
LogicalResult ValueTensorLiteralOp::inferReturnTypes(
MLIRContext *context, Optional<Location> location, ValueRange operands,
DictionaryAttr attributes, RegionRange regions,
SmallVectorImpl<Type> &inferredReturnTypes) {
auto attr = attributes.get("value").dyn_cast_or_null<ElementsAttr>();
if (!attr)
return failure();
RankedTensorType tensorType = attr.getType().cast<RankedTensorType>();
ValueTensorType returnType =
ValueTensorType::get(tensorType.getContext(), tensorType.getShape(),
tensorType.getElementType());
inferredReturnTypes.push_back(returnType);
return success();
}
OpFoldResult ValueTensorLiteralOp::fold(ArrayRef<Attribute> operands) {
return valueAttr();
}
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
//----------------------------------------------------------------------------//
// TensorStaticInfoCast
//----------------------------------------------------------------------------//
bool TensorStaticInfoCastOp::areCastCompatible(mlir::TypeRange inputs,
mlir::TypeRange outputs) {
return areSizesAndDtypesCompatible(inputs[0].cast<BaseTensorType>(),
outputs[0].cast<BaseTensorType>());
}
void TensorStaticInfoCastOp::getCanonicalizationPatterns(
RewritePatternSet &patterns, MLIRContext *context) {
patterns.add(+[](TensorStaticInfoCastOp op, PatternRewriter &rewriter) {
auto reverseCast =
op.operand().getDefiningOp<Torch::TensorStaticInfoCastOp>();
if (!reverseCast || reverseCast.operand().getType() != op.getType())
return failure();
rewriter.replaceOp(op, reverseCast.operand());
return success();
});
}
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
//===----------------------------------------------------------------------===//
// CopyToNonValueTensorOp
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
//===----------------------------------------------------------------------===//
static LogicalResult verify(CopyToNonValueTensorOp op) {
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
auto resultType = op.getResult().getType().cast<BaseTensorType>();
auto operandType = op.getOperand().getType().cast<BaseTensorType>();
if (!resultType.hasSameSizesAndDtype(operandType)) {
return op.emitError()
<< "operand and result must have same sizes and dtype";
}
return success();
}
LogicalResult CopyToNonValueTensorOp::inferReturnTypes(
MLIRContext *context, Optional<Location> location, ValueRange operands,
DictionaryAttr attributes, RegionRange regions,
SmallVectorImpl<Type> &inferredReturnTypes) {
auto resultType = operands[0].getType().cast<ValueTensorType>();
inferredReturnTypes.push_back(resultType.getWithoutValueSemantics());
return success();
}
void CopyToNonValueTensorOp::getEffects(
SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
&effects) {
effects.emplace_back(MemoryEffects::Allocate::get(), getResult());
}
//===----------------------------------------------------------------------===//
// CopyToValueTensorOp
//===----------------------------------------------------------------------===//
static LogicalResult verify(CopyToValueTensorOp op) {
auto resultType = op.getResult().getType().cast<BaseTensorType>();
auto operandType = op.getOperand().getType().cast<BaseTensorType>();
if (!resultType.hasSameSizesAndDtype(operandType)) {
return op.emitError()
<< "operand and result must have same sizes and dtype";
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
}
return success();
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
}
LogicalResult CopyToValueTensorOp::inferReturnTypes(
MLIRContext *context, Optional<Location> location, ValueRange operands,
DictionaryAttr attributes, RegionRange regions,
SmallVectorImpl<Type> &inferredReturnTypes) {
auto resultType = operands[0].getType().cast<NonValueTensorType>();
inferredReturnTypes.push_back(resultType.getWithValueSemantics());
return success();
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
}
void CopyToValueTensorOp::getEffects(
SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
&effects) {
effects.emplace_back(MemoryEffects::Read::get(), getOperand());
}
//===----------------------------------------------------------------------===//
// ConstantNoneOp
//===----------------------------------------------------------------------===//
OpFoldResult ConstantNoneOp::fold(ArrayRef<Attribute> operands) {
return TypeAttr::get(Torch::NoneType::get(getContext()));
}
void ConstantNoneOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
setNameFn(getResult(), "none");
}
//===----------------------------------------------------------------------===//
// ConstantStrOp
//===----------------------------------------------------------------------===//
OpFoldResult ConstantStrOp::fold(ArrayRef<Attribute> operands) {
return valueAttr();
}
void ConstantStrOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
setNameFn(getResult(), "str");
}
//===----------------------------------------------------------------------===//
// ConstantDeviceOp
//===----------------------------------------------------------------------===//
void ConstantDeviceOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
setNameFn(getResult(), value());
}
//===----------------------------------------------------------------------===//
// ConstantIntOp
//===----------------------------------------------------------------------===//
static ParseResult parseConstantIntOp(OpAsmParser &parser,
OperationState &result) {
Builder builder(result.getContext());
result.addTypes(builder.getType<Torch::IntType>());
if (parser.parseOptionalAttrDict(result.attributes))
return failure();
int64_t value;
if (parser.parseInteger(value))
return failure();
result.addAttribute("value", builder.getI64IntegerAttr(value));
return success();
}
static void print(OpAsmPrinter &p, Torch::ConstantIntOp op) {
p << " ";
p << op.value().getSExtValue();
p.printOptionalAttrDict(op->getAttrs(), {"value"});
}
OpFoldResult Torch::ConstantIntOp::fold(ArrayRef<Attribute> operands) {
return valueAttr();
}
void Torch::ConstantIntOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
SmallVector<char> buf;
llvm::raw_svector_ostream os(buf);
os << "int" << value();
setNameFn(getResult(), os.str());
}
//===----------------------------------------------------------------------===//
// ConstantFloatOp
//===----------------------------------------------------------------------===//
OpFoldResult Torch::ConstantFloatOp::fold(ArrayRef<Attribute> operands) {
return valueAttr();
}
void Torch::ConstantFloatOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
// Calculate a stringified version of the number, compatible with MLIR
// identifier syntax. (in practice, this just removes the '+' from 'e+' in
// float string representation).
SmallVector<char> buf;
value().toString(buf, /*FormatPrecision=*/6, /*FormatMaxPadding=*/0,
/*TruncateZero=*/false);
auto isValidMLIRIdentifierChar = [](char c) {
return isalpha(c) || isdigit(c) || c == '_' || c == '$' || c == '.' ||
c == '-';
};
auto numberStr = llvm::to_vector<16>(
llvm::make_filter_range(buf, isValidMLIRIdentifierChar));
// Construct the identifier string.
buf.clear();
llvm::append_range(buf, StringRef("float"));
llvm::append_range(buf, numberStr);
setNameFn(getResult(), StringRef(buf.data(), buf.size()));
}
//===----------------------------------------------------------------------===//
// ConstantBoolOp
//===----------------------------------------------------------------------===//
OpFoldResult Torch::ConstantBoolOp::fold(ArrayRef<Attribute> operands) {
return valueAttr();
}
void Torch::ConstantBoolOp::getAsmResultNames(
function_ref<void(Value, StringRef)> setNameFn) {
setNameFn(getResult(), value() ? "true" : "false");
}
//===----------------------------------------------------------------------===//
// PrimUncheckedCastOp
//===----------------------------------------------------------------------===//
bool PrimUncheckedCastOp::areCastCompatible(mlir::TypeRange inputs,
mlir::TypeRange outputs) {
return isValidSubtype(outputs[0], inputs[0]);
}
//===----------------------------------------------------------------------===//
// Aten__Getitem__TOp
//===----------------------------------------------------------------------===//
void Aten__Getitem__TOp::getCanonicalizationPatterns(
RewritePatternSet &patterns, MLIRContext *context) {
patterns.add(+[](Aten__Getitem__TOp op, PatternRewriter &rewriter) {
auto torchList = op.getOperand(0);
// TODO: Use a proper effects interface when more operands taking a list
// are implemented.
if (!llvm::all_of(torchList.getUsers(), [](Operation *op) {
return isa<Aten__Getitem__TOp, AtenLenTOp>(op);
}))
return failure();
auto listConstruct = torchList.getDefiningOp<Torch::PrimListConstructOp>();
if (!listConstruct)
return failure();
int64_t index;
if (!matchPattern(op.getOperand(1), m_TorchConstantInt(&index)))
return failure();
rewriter.replaceOp(op, {listConstruct.getOperand(index)});
return success();
});
}
//===----------------------------------------------------------------------===//
// PrimTupleIndexOp
//===----------------------------------------------------------------------===//
void PrimTupleIndexOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
MLIRContext *context) {
patterns.add(+[](PrimTupleIndexOp op, PatternRewriter &rewriter) {
auto tupleConstruct = op.tup().getDefiningOp<Torch::PrimTupleConstructOp>();
if (!tupleConstruct)
return failure();
int64_t i;
if (!matchPattern(op.i(), m_TorchConstantInt(&i)))
return failure();
if (i >= (int64_t)tupleConstruct.elements().size())
return failure();
rewriter.replaceOp(op, tupleConstruct.elements()[i]);
return success();
});
}
//===----------------------------------------------------------------------===//
// PrimUninitializedOp
//===----------------------------------------------------------------------===//
void PrimUninitializedOp::getCanonicalizationPatterns(
RewritePatternSet &patterns, MLIRContext *context) {
patterns.add(+[](PrimUninitializedOp op, PatternRewriter &rewriter) {
if (!op.use_empty())
return failure();
rewriter.eraseOp(op);
return success();
});
}
//===----------------------------------------------------------------------===//
// PrimTupleUnpackOp
//===----------------------------------------------------------------------===//
void PrimTupleUnpackOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
MLIRContext *context) {
patterns.add(+[](PrimTupleUnpackOp op, PatternRewriter &rewriter) {
auto tupleConstruct = op.tup().getDefiningOp<Torch::PrimTupleConstructOp>();
if (!tupleConstruct)
return failure();
rewriter.replaceOp(op, tupleConstruct.elements());
return success();
});
}
static PrimDictConstructOp getDictConstructIfNotModified(Value torchDict) {
if (!llvm::all_of(torchDict.getUsers(), [](Operation *op) {
return isa<Aten__Getitem__DictStrOp, Aten__Contains__StrOp,
AtenKeysStrOp, AtenGetDefaultStrOp, PrimDictConstructOp>(op);
}))
return nullptr;
return torchDict.getDefiningOp<Torch::PrimDictConstructOp>();
}
//===----------------------------------------------------------------------===//
// Aten__Getitem__DictStrOp
//===----------------------------------------------------------------------===//
OpFoldResult Aten__Getitem__DictStrOp::fold(ArrayRef<Attribute> operands) {
auto dictConstruct = getDictConstructIfNotModified(self());
if (!dictConstruct)
return nullptr;
auto targetKey = key();
for (auto i : llvm::zip(dictConstruct.keys(), dictConstruct.values())) {
auto k = std::get<0>(i);
if (k == targetKey)
return std::get<1>(i);
}
return nullptr;
}
//===----------------------------------------------------------------------===//
// Aten__Contains__StrOp
//===----------------------------------------------------------------------===//
OpFoldResult Aten__Contains__StrOp::fold(ArrayRef<Attribute> operands) {
auto dictConstruct = getDictConstructIfNotModified(dict());
if (!dictConstruct)
return nullptr;
auto targetKey = key();
for (auto key : dictConstruct.keys()) {
if (key == targetKey)
return getI1IntegerAttr(getContext(), true);
}
return nullptr;
}
using BinaryIntOperatorFn = std::function<int64_t(int64_t, int64_t)>;
template <typename OpTy>
static OpFoldResult atenBinaryIntOperatorFoldHelper(OpTy op,
BinaryIntOperatorFn f) {
int64_t lhs, rhs;
if (!matchPattern(op.getOperand(0), m_TorchConstantInt(&lhs)) ||
!matchPattern(op.getOperand(1), m_TorchConstantInt(&rhs)))
return nullptr;
return getI64IntegerAttr(op.getContext(), f(lhs, rhs));
}
//===----------------------------------------------------------------------===//
// AtenFloordivIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenFloordivIntOp::fold(ArrayRef<Attribute> operands) {
return atenBinaryIntOperatorFoldHelper(
*this, [](int64_t a, int64_t b) { return std::floor(a / (double)b); });
}
//===----------------------------------------------------------------------===//
// AtenRemainderIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenRemainderIntOp::fold(ArrayRef<Attribute> operands) {
return atenBinaryIntOperatorFoldHelper(
*this, [](int64_t a, int64_t b) { return a % b; });
}
//===----------------------------------------------------------------------===//
// AtenAddIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenAddIntOp::fold(ArrayRef<Attribute> operands) {
return atenBinaryIntOperatorFoldHelper(
*this, [](int64_t a, int64_t b) { return a + b; });
}
//===----------------------------------------------------------------------===//
// AtenSubIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenSubIntOp::fold(ArrayRef<Attribute> operands) {
return atenBinaryIntOperatorFoldHelper(
*this, [](int64_t a, int64_t b) { return a - b; });
}
//===----------------------------------------------------------------------===//
// AtenMulIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenMulIntOp::fold(ArrayRef<Attribute> operands) {
int64_t lhs, rhs;
bool lConstant = matchPattern(getOperand(0), m_TorchConstantInt(&lhs));
bool rConstant = matchPattern(getOperand(1), m_TorchConstantInt(&rhs));
if ((lConstant && lhs == 0) || (rConstant && rhs == 0))
return getI64IntegerAttr(getContext(), 0);
if (lConstant && rConstant)
return getI64IntegerAttr(getContext(), lhs * rhs);
return nullptr;
}
//===----------------------------------------------------------------------===//
// AtenNegIntOp
//===----------------------------------------------------------------------===//
OpFoldResult AtenNegIntOp::fold(ArrayRef<Attribute> operands) {
int64_t c;
if (matchPattern(getOperand(), m_TorchConstantInt(&c)))
return getI64IntegerAttr(getContext(), -c);
return nullptr;
}
//===----------------------------------------------------------------------===//
// PrimDtypeOp
//===----------------------------------------------------------------------===//
OpFoldResult PrimDtypeOp::fold(ArrayRef<Attribute> operands) {
BaseTensorType tensorType = a().getType().cast<BaseTensorType>();
if (tensorType.hasDtype()) {
int64_t dtypeInt = getDtypeIntegerFromMlirType(tensorType.getDtype());
if (dtypeInt != -1)
return getI64IntegerAttr(getContext(), dtypeInt);
}
return nullptr;
}
OpFoldResult AtenIntTensorOp::fold(ArrayRef<Attribute> operands) {
// If an scalar number is converted to a 0-d tensor and passed on to
// aten.Int.Tensor, fold to the scalar number.
if (auto numToTensorScalar = a().getDefiningOp<PrimNumToTensorScalarOp>())
return numToTensorScalar.a();
return nullptr;
}
#define GET_OP_CLASSES
[torch-mlir earthmoving (1/N)] C/C++ code movement. This creates the `external/torch-mlir` directory as an LLVM_EXTERNAL_PROJECTS-compatible project (analogous to `iree-dialects`) and completes movement/rename of all pure MLIR C/C++ compiler code into there. The next step will be to move all the Python code / code that links/includes PyTorch C++ code (which currently lives in `frontends/pytorch`) into a subdirectory here. I call this "earthmoving" because it is mostly mechanical changes and renames. As a quick summary (we can change this down the road easily) - C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch` - CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet` - preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_` - CMake `NPCOMPFoo -> TorchMLIRFoo` The goal of this is to create a standalone project creating a center of mass for entry into the MLIR ecosystem from PyTorch, suitable in scope for eventual inclusion/ownership in PyTorch. The idea is that `external/torch-mlir` will some day be pulled out into its own repository, and then npcomp will simply pull it in as a submodule. Layering-wise, what lives in `torch-mlir` lowers code from PyTorch (currently TorchScript, but TorchFX or pytorch/xla-style tracing are possible extensions) down to what we have been calling the "Torch backend contract" which is cleaned up IR (inlining, simplifcation, conversion to value tensors, ...) entirely in the `torch` dialect. This is the branching off point for further lowering, of which npcomp takes one opinion (outside `torch-mlir` of course!), namely the `TorchConversion` dialect/transforms which lower to IR suitable for IREE and other linalg-on-tensors based lower-level compilers. Summary of changes: - move `{include,lib,test}/Dialect/Torch` into `torch-mlir` - move relevant parts of CAPI into `torch-mlir`. - leave a few things related to the `torch-mlir` Python build commented out, which should be resolved in a subsequent change.
2021-09-10 03:24:10 +08:00
#include "torch-mlir/Dialect/Torch/IR/TorchOps.cpp.inc"