Add initial LTC backend (#610)

* Add initial LTC backend skeleton

* Disable CI build and move TorchMLIRPyTorch.cmake
pull/1125/head
Jae Hoon (Antonio) Kim 2022-02-17 20:15:37 -05:00 committed by Henry Tu
parent 8b5631d4c5
commit 2f22e2ef40
13 changed files with 732 additions and 1 deletions

View File

@ -5,6 +5,8 @@ on:
branches: branches:
- main - main
pull_request: pull_request:
branches:
- main
workflow_dispatch: workflow_dispatch:
jobs: jobs:

View File

@ -56,6 +56,12 @@ declare_mlir_python_extension(TorchMLIRPythonExtensions.Main
LLVMSupport LLVMSupport
) )
################################################################################
# Lazy Tensor Core
################################################################################
add_subdirectory(torch_mlir/csrc)
################################################################################ ################################################################################
# Optionally handle JIT IR importer. # Optionally handle JIT IR importer.
################################################################################ ################################################################################
@ -147,5 +153,8 @@ endif()
# TODO: Add after macOS builds are fixed # TODO: Add after macOS builds are fixed
#add_dependencies(TorchMLIRPythonModules torch_mlir_custom_op_example) #add_dependencies(TorchMLIRPythonModules torch_mlir_custom_op_example)
# Add Torch-MLIR LTC backend as dependency
add_dependencies(TorchMLIRPythonModules torch_mlir_ltc_backend)
add_subdirectory(test) add_subdirectory(test)

View File

@ -0,0 +1,43 @@
#-------------------------------------------------------------------------------
# Setup PyTorch/LTC
#-------------------------------------------------------------------------------
list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/python/torch_mlir/cmake/modules")
include(TorchMLIRPyTorch)
TorchMLIRProbeForPyTorchInstall()
find_package(Torch 1.11 REQUIRED)
TorchMLIRConfigurePyTorch()
include_directories(BEFORE
${TORCH_INCLUDE_DIRS}
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_CURRENT_BINARY_DIR}
${Python3_INCLUDE_DIRS}
)
link_directories("${TORCH_INSTALL_PREFIX}/lib")
add_library(torch_mlir_ltc_backend SHARED
backend/backend_impl.cc
backend/mlir_lowering_context.cc
backend/mlir_node.cc
)
target_link_libraries(torch_mlir_ltc_backend
TorchMLIRAggregateCAPI
${TORCH_LIBRARIES}
${Python3_LIBRARIES}
torch_python
)
message(STATUS "TORCH_CXXFLAGS=${TORCH_CXXFLAGS} -Wpedantic")
set_target_properties(torch_mlir_ltc_backend PROPERTIES
LIBRARY_OUTPUT_DIRECTORY "${TORCH_MLIR_PYTHON_PACKAGES_DIR}/torch_mlir/"
OUTPUT_NAME _MLIR_LTC
PREFIX "${PYTHON_MODULE_PREFIX}"
SUFFIX "${PYTHON_MODULE_EXTENSION}"
CXX_VISIBILITY_PRESET "hidden"
COMPILE_FLAGS "${TORCH_CXXFLAGS} -Wpedantic"
)

View File

@ -0,0 +1,19 @@
# Torch-MLIR Lazy Tensor Core Backend
Contained within this directory are the components that implements the
Torch-MLIR LTC backend.
The components are subclasses of the backend API interface classes found under
[torch/csrc/lazy/backend](https://github.com/pytorch/pytorch/tree/master/torch/csrc/lazy/backend).
Importantly, the subclasses are still abstract classes. Pure virtual methods
such as `Compile` were purposefully not overriden as Torch-MLIR does not know
how to compile the model for the target hardware.
The intent is that vendor hardware specific plugins will subclass the Torch-MLIR
backend classes and override the remaining pure virtual functions to complete
the backend.
The Torch-MLIR LTC backend's job is to perform the lowering from ATen to MLIR. A
hardware vendor's backend job is to take care of the actual compile and
execution of the lowered MLIR.

View File

@ -0,0 +1,159 @@
#include <torch/csrc/lazy/backend/backend_data.h>
#include <torch/csrc/lazy/backend/backend_device.h>
#include <torch/csrc/lazy/backend/lowering_context.h>
#include <torch/csrc/lazy/core/shape.h>
#include "backend_impl.h"
#include "mlir_lowering_context.h"
#include "../utils/exception.h"
namespace torch {
namespace lazy {
struct MlirBackendData::Info : public BackendData::Info {
at::Tensor tensor;
c10::optional<at::Scalar> scalar;
Info() {}
Info(const Info& other) :
tensor{other.tensor}, scalar{other.scalar} {}
Info(const at::Tensor& tensor) : tensor{tensor} {}
Info(const at::Scalar& scalar) : scalar{scalar} {}
};
MlirBackendData::MlirBackendData(BackendDevice device, Shape shape) :
BackendData(device, shape) {
auto info = std::make_shared<MlirBackendData::Info>();
SetInfo(info);
}
MlirBackendData::MlirBackendData(const at::Scalar& scalar, BackendDevice device) :
BackendData(device, torch::lazy::Shape(scalar.type(), {})) {
auto info = std::make_shared<MlirBackendData::Info>(scalar);
SetInfo(info);
}
MlirBackendData::MlirBackendData(const at::Tensor& tensor, BackendDevice device, Shape shape) :
BackendData(device, shape) {
auto info = std::make_shared<MlirBackendData::Info>(tensor);
SetInfo(info);
}
BackendData::Handle MlirBackendData::GetHandle() { return reinterpret_cast<int64_t>(this); }
void MlirBackendData::Assign(const BackendData& data) {
MlirBackendData::Info* info =
dynamic_cast<MlirBackendData::Info*>(data.info());
TORCH_CHECK(
info, "Invalid Backend Data Pointer. Expected MlirBackendData::Info."
);
auto new_info = std::make_shared<MlirBackendData::Info>(*info);
SetInfo(new_info);
}
bool MlirBackendData::HasValue() const {
return bool(info());
}
/**
* Initialization/Teardown
* */
void MlirBackendImpl::PrepareToExit() const {}
/**
* Data Transfer
* */
BackendDataPtr MlirBackendImpl::MakeComputationDataFromTensor(
const at::Tensor& tensor,
const Shape& shape,
const BackendDevice& device
) const {
return std::make_shared<MlirBackendData>(tensor, device, shape);
}
BackendDataPtr MlirBackendImpl::MakeComputationDataFromScalar(
const at::Scalar& scalar,
const torch::lazy::BackendDevice& device
) const {
return std::make_shared<MlirBackendData>(scalar, device);
}
BackendDataPtr MlirBackendImpl::CreateDataPlaceholder(
const BackendDevice& device, const Shape& shape
) const {
return std::make_shared<MlirBackendData>(device, shape);
}
at::Tensor MlirBackendImpl::MakeTensorFromComputationData(
const BackendDataPtr data,
c10::optional<at::ScalarType> logical_scalar_type
) const {
MlirBackendData::Info* info =
dynamic_cast<MlirBackendData::Info*>(data->info());
TORCH_CHECK(
info, "Invalid Backend Data Pointer. Expected MlirBackendData::Info."
);
return info->tensor;
}
/**
* Lowering, Compilation, Execution
* */
std::unique_ptr<LoweringContext> MlirBackendImpl::CreateLoweringContext(
const std::string& name,
BackendDevice device,
c10::ArrayRef<torch::lazy::Node*> post_order,
Util::EmissionMap emit_status
) const {
return std::make_unique<MlirLoweringContext>(
name,
std::forward<BackendDevice>(device),
std::forward<c10::ArrayRef<torch::lazy::Node*>>(post_order),
std::forward<Util::EmissionMap>(emit_status)
);
}
std::unique_ptr<LoweringContext> MlirBackendImpl::CreateLoweringContext(
const std::string& name, BackendDevice device
) const {
return std::make_unique<MlirLoweringContext>(
name, std::forward<BackendDevice>(device)
);
}
/**
* Device Configuration
* */
// Set or get the default device type.
// For backends used with virtual c10:: Devices, this configures what real
// device type the backend should use, and matters if the backend supports
// more than one type of real device.
// Specify which aten device should be used for eager fallback
// may change depending on current 'Default' DeviceType
at::DeviceType MlirBackendImpl::EagerFallbackDeviceType() const {
return at::DeviceType::CPU;
}
// Query all available backend devices
std::vector<BackendDevice> MlirBackendImpl::GetBackendDevices() const {
return {
GetBackendDevice(c10::Device(c10::kCPU, 0)),
GetBackendDevice(c10::Device(c10::kLazy, 0))
};
}
// Map a particular c10:: device to a concrete backend device
// Note:: c10:: devices may be virtual or concrete. xla:: and lazy:: are
// virtual devices, meaning they may map to a gpu, tpu, etc. behind the
// scenes. In the future, non-virtual c10:: devices may also use lazy tensors
// through a mode, in which case these APIs should still work, but should be
// identity mappings.
BackendDevice MlirBackendImpl::GetBackendDevice(c10::Device device) const {
return torch::lazy::BackendDevice(GetDefaultDeviceType(), device.index());
}
} // lazy
} // torch

View File

@ -0,0 +1,136 @@
#pragma once
#include <torch/csrc/lazy/backend/backend_data.h>
#include <torch/csrc/lazy/backend/backend_device.h>
#include <torch/csrc/lazy/backend/backend_interface.h>
#include <torch/csrc/lazy/core/shape.h>
namespace torch {
namespace lazy {
class MlirBackendData : public torch::lazy::BackendData {
public:
struct Info;
MlirBackendData(torch::lazy::BackendDevice device, torch::lazy::Shape shape);
MlirBackendData(const at::Scalar& scalar, torch::lazy::BackendDevice device);
MlirBackendData(const at::Tensor& tensor, torch::lazy::BackendDevice device, torch::lazy::Shape shape);
virtual torch::lazy::BackendData::Handle GetHandle() override;
virtual void Assign(const torch::lazy::BackendData& data) override;
virtual bool HasValue() const override;
};
class MlirBackendImpl : public torch::lazy::BackendImplInterface {
public:
/**
* Initialization/Teardown
* */
virtual void PrepareToExit() const override;
/**
* Configuration
* */
// virtual void SetRngSeed(size_t seed) const = 0;
/**
* Data Transfer
* */
virtual torch::lazy::BackendDataPtr MakeComputationDataFromTensor(
const at::Tensor& tensor,
const torch::lazy::Shape& shape,
const torch::lazy::BackendDevice& device
) const override;
virtual torch::lazy::BackendDataPtr MakeComputationDataFromScalar(
const at::Scalar& scalar,
const torch::lazy::BackendDevice& device
) const override;
virtual torch::lazy::BackendDataPtr CreateDataPlaceholder(
const torch::lazy::BackendDevice& device, const torch::lazy::Shape& shape
) const override;
virtual at::Tensor MakeTensorFromComputationData(
const torch::lazy::BackendDataPtr data,
c10::optional<at::ScalarType> logical_scalar_type
) const override;
/**
* Lowering, Compilation, Execution
* */
virtual std::unique_ptr<torch::lazy::LoweringContext> CreateLoweringContext(
const std::string& name,
torch::lazy::BackendDevice device,
c10::ArrayRef<torch::lazy::Node*> post_order,
torch::lazy::Util::EmissionMap emit_status
) const override;
virtual std::unique_ptr<torch::lazy::LoweringContext> CreateLoweringContext(
const std::string& name, torch::lazy::BackendDevice device
) const override;
// TODO(whc) need to keep this?
// virtual std::vector<std::string> GetCompilationDevices(
// const std::string& device, c10::ArrayRef<std::string> devices
// ) const = 0;
// virtual std::vector<torch::lazy::ComputationPtr> Compile(
// std::vector<torch::lazy::ComputationPtr> instances
// ) const = 0;
// virtual std::vector<torch::lazy::BackendDataPtr> ExecuteComputation(
// torch::lazy::Computation& computation,
// c10::ArrayRef<torch::lazy::BackendDataPtr> arguments,
// const torch::lazy::BackendDevice& device
// ) const = 0;
/**
* Device Configuration
* */
// Set or get the default device type.
// For backends used with virtual c10:: Devices, this configures what real
// device type the backend should use, and matters if the backend supports
// more than one type of real device.
// virtual std::shared_ptr<torch::lazy::BackendDeviceType> GetDefaultDeviceType() const = 0;
// virtual void SetDefaultDeviceType(std::string device_type) = 0;
// Specify which aten device should be used for eager fallback
// may change depending on current 'Default' DeviceType
virtual at::DeviceType EagerFallbackDeviceType() const override;
// Query all available backend devices
virtual std::vector<torch::lazy::BackendDevice> GetBackendDevices() const override;
// Map a particular c10:: device to a concrete backend device
// Note:: c10:: devices may be virtual or concrete. xla:: and lazy:: are
// virtual devices, meaning they may map to a gpu, tpu, etc. behind the
// scenes. In the future, non-virtual c10:: devices may also use lazy tensors
// through a mode, in which case these APIs should still work, but should be
// identity mappings.
virtual torch::lazy::BackendDevice GetBackendDevice(c10::Device device) const override;
/**
* Debug/Metrics
* */
// virtual std::map<std::string, Metric> GetMetrics() const = 0;
// virtual MemoryInfo GetMemoryInfo(const std::string& device) = 0;
// virtual std::string GetComputationBackendText(
// const torch::lazy::ComputationPtr computation
// ) const = 0;
};
} // lazy
} // torch

View File

@ -0,0 +1,93 @@
#include <iostream>
#include "mlir_lowering_context.h"
#include "../utils/exception.h"
namespace torch {
namespace lazy {
MlirLoweringContext::MlirLoweringContext(
const std::string& name, BackendDevice device
) : LoweringContext(name, std::forward<BackendDevice>(device)) {}
MlirLoweringContext::MlirLoweringContext(
const std::string& name,
BackendDevice device,
c10::ArrayRef<torch::lazy::Node*> post_order,
Util::EmissionMap emit_status
) : LoweringContext(
name,
std::forward<BackendDevice>(device),
std::forward<c10::ArrayRef<torch::lazy::Node*>>(post_order),
std::forward<Util::EmissionMap>(emit_status)
) {}
int MlirComputation::parameters_size() const {
UNIMPLEMENTED_ERROR("MlirComputation::parameters_size");
}
const std::vector<torch::lazy::Shape>& MlirComputation::parameter_shapes() const {
UNIMPLEMENTED_ERROR("MlirComputation::parameter_shapes");
}
const std::vector<std::string>& MlirComputation::parameter_names() const {
UNIMPLEMENTED_ERROR("MlirComputation::parameter_names");
}
const torch::lazy::Shape& MlirComputation::result_shape() const {
UNIMPLEMENTED_ERROR("MlirComputation::result_shape");
}
// Get the shape of the result tuple component, given by index.
torch::lazy::Shape MlirLoweringContext::GetResultShape(size_t index) const {
UNIMPLEMENTED_ERROR("MlirLoweringContext::GetResultShape( " << index << " )");
}
// Adds the given output as a component of the result tuple and returns its
// assigned position within the tuple.
size_t MlirLoweringContext::AddResult(const torch::lazy::Output& output) {
const torch::lazy::Node* node;
auto it = emitted_outputs_.find(output);
if (it == emitted_outputs_.end()) {
node = output.node;
auto post_order = Util::ComputePostOrder(node, &emit_status_);
for (auto po_node : post_order) {
// TODO: uncomment after lowering is implemented
// bool ok = lowering_->Lower(node);
// TORCH_CHECK(ok, "Failed to lower: ", node->ToString());
}
emitted_outputs_[output] = node;
} else {
node = it->second;
}
result_tuple_.emplace_back(node);
return result_tuple_.size() - 1;
}
// Associates the given output with the input parameter of the given index and
// shape. Only used for the operator-by-operator execution, mostly for
// debugging purposes.
void MlirLoweringContext::AddParameter(
const torch::lazy::Output& output,
size_t index,
const torch::lazy::Shape& shape,
const std::string& name
) {
UNIMPLEMENTED_ERROR("MlirLoweringContext::AddParameter");
}
// Build the computation capturing all the operations created with the
// embedded builder (returned by the builder() API).
ComputationPtr MlirLoweringContext::Build() {
for (const torch::lazy::Node* output : result_tuple_) {
}
return std::make_shared<MlirComputation>();
}
} // namespace lazy
} // namespace torch

View File

@ -0,0 +1,55 @@
#pragma once
#include <vector>
#include <torch/csrc/lazy/backend/lowering_context.h>
namespace torch {
namespace lazy {
class MlirComputation : public torch::lazy::Computation {
public:
int parameters_size() const override;
virtual const std::vector<torch::lazy::Shape>& parameter_shapes() const override;
virtual const std::vector<std::string>& parameter_names() const override;
virtual const torch::lazy::Shape& result_shape() const override;
};
class MlirLoweringContext : public torch::lazy::LoweringContext {
public:
MlirLoweringContext(const std::string& name, torch::lazy::BackendDevice device);
MlirLoweringContext(const std::string& name,
torch::lazy::BackendDevice device,
c10::ArrayRef<torch::lazy::Node*> post_order,
torch::lazy::Util::EmissionMap emit_status);
// Get the shape of the result tuple component, given by index.
virtual torch::lazy::Shape GetResultShape(size_t index) const override;
// Adds the given output as a component of the result tuple and returns its
// assigned position within the tuple.
virtual size_t AddResult(const torch::lazy::Output& output) override;
// Associates the given output with the input parameter of the given index and
// shape. Only used for the operator-by-operator execution, mostly for
// debugging purposes.
virtual void AddParameter(const torch::lazy::Output& output,
size_t index,
const torch::lazy::Shape& shape,
const std::string& name) override;
// Build the computation capturing all the operations created with the
// embedded builder (returned by the builder() API).
virtual torch::lazy::ComputationPtr Build() override;
private:
std::vector<const torch::lazy::Node*> result_tuple_;
torch::lazy::OutputMap<const torch::lazy::Node*> emitted_outputs_;
};
} // namespace lazy
} // namespace torch

View File

@ -0,0 +1,124 @@
#include <torch/csrc/lazy/core/cache.h>
#include "mlir_node.h"
#include "../utils/exception.h"
namespace torch {
namespace lazy {
namespace {
hash_t OperandHashes(const OpList& operands, const hash_t& seed, const bool bakeInSizes) {
hash_t hash = seed;
for (auto& operand : operands) {
if (!operand) {
hash = HashCombine(hash, static_cast<uint64_t>(kNullOpt));
continue;
}
auto operand_hash = bakeInSizes ? operand.hash_with_sizes() : operand.hash_without_sizes();
hash = HashCombine(hash, operand_hash);
}
return hash;
}
hash_t GetOpHash(OpKind op, const Shape& shape, hash_t hash_seed, const bool bakeInSizes) {
hash_t h = HashCombine(op.hash(), shape.hash(bakeInSizes));
return HashCombine(h, hash_seed);
}
} // namespace
MlirNode::MlirNode(
OpKind op, OpList operands, std::vector<Shape>&& shapes,
size_t num_outputs, hash_t hash_seed
) : Node(
op, num_outputs,
/* node_hash */ HashCombine(op.hash(), hash_seed),
/* dag_hash */
[&](bool bakeInSizes) -> hash_t {
return OperandHashes(operands, HashCombine(op.hash(), hash_seed), bakeInSizes);
}
),
shapes_(std::move(shapes)) {
for (auto& operand : operands) {
// Ideally, optional operands should be filtered by the leaf node classes,
// but it's just much easier to do it here.
if (!operand) {
continue;
}
AddOperand(operand.node, operand.index);
}
}
MlirNode::MlirNode(
OpKind op, OpList operands,
const std::function<Shape()>& shape_fn,
size_t num_outputs, hash_t hash_seed
) : MlirNode(
op, operands, std::vector<Shape>{}, num_outputs, hash_seed
) {
shapes_.push_back(GetOpShape(shape_fn));
}
MlirNode::MlirNode(
OpKind op, OpList operands, size_t num_outputs, hash_t hash_seed
) : MlirNode(op, operands, std::vector<Shape>{}, num_outputs, hash_seed) {}
void MlirNode::SetShapeDeferred(
const std::function<Shape()>& shape_fn
) {
shapes_.push_back(GetOpShape(shape_fn));
}
MlirNode::MlirNode(OpKind op, Shape shape, size_t num_outputs, hash_t hash_seed)
: Node(
op, num_outputs,
[&](bool bakeInSizes) -> hash_t {
return GetOpHash(op, shape, hash_seed, bakeInSizes);
}
) {
shapes_.push_back(std::move(shape));
}
using ShapeCache = Cache<hash_t, Shape, HashReducer>;
constexpr const int torch_lazy_shape_cache_size = 4096;
ShapeCache* GetShapeCache() {
static ShapeCache* cache = new ShapeCache(torch_lazy_shape_cache_size);
return cache;
}
Shape MlirNode::GetOpShape(const std::function<Shape()>& shape_fn) const {
ShapeCache* shape_cache = GetShapeCache();
auto shape = shape_cache->Get(hash());
if (shape == nullptr) {
shape = shape_cache->Add(
hash(), std::make_shared<Shape>(shape_fn())
);
}
return *shape;
}
const std::vector<Output>& MlirNode::operands() const {
return operands_as_outputs_;
}
const Output& MlirNode::operand(size_t i) const {
return operands_as_outputs_.at(i);
}
void MlirNode::AddOperand(NodePtr node, size_t index) {
CHECK_LT(index, node->num_outputs());
operands_.push_back(std::move(node));
operands_as_outputs_.emplace_back(operands_.back().get(), index);
}
} // namespace lazy
} // namespace torch

View File

@ -0,0 +1,71 @@
#pragma once
#include <ATen/core/interned_strings.h>
#include <torch/csrc/lazy/backend/lowering_context.h>
#include <torch/csrc/lazy/core/shape.h>
#include <torch/csrc/lazy/core/ir.h>
#include "mlir_lowering_context.h"
#include "../utils/exception.h"
namespace torch {
namespace lazy {
typedef std::vector<NodePtr> MlirOpVector;
typedef NodePtr MlirFunction;
class MlirNode : public torch::lazy::Node {
public:
MlirNode(
OpKind op, OpList operands, std::vector<Shape>&& shapes,
size_t num_outputs = 1, hash_t hash_seed = kHashSeed
);
// Same as the constructor above, but the shape is generated by a function,
// only if needed (shape cache miss).
MlirNode(
OpKind op, OpList operands,
const std::function<Shape()>& shape_fn,
size_t num_outputs = 1, hash_t hash_seed = kHashSeed
);
// The shape is set later.
MlirNode(
OpKind op, OpList operands, size_t num_outputs = 1,
hash_t hash_seed = kHashSeed
);
void SetShapeDeferred(const std::function<Shape()>& shape_fn);
// Contructor used to create leaf nodes.
MlirNode(
OpKind op, Shape shape, size_t num_outputs = 1, hash_t hash_seed = kHashSeed
);
Shape GetOpShape(const std::function<Shape()>& shape_fn) const;
const std::vector<Output>& operands() const override;
const Output& operand(size_t i) const override;
virtual MlirOpVector Lower(
MlirFunction function,
MlirLoweringContext* loctx
) const = 0;
private:
// Adds node's index output number as operand.
void AddOperand(NodePtr node, size_t index = 0);
std::vector<Shape> shapes_;
// A node holds a real reference to its operands.
std::vector<NodePtr> operands_;
// Outputs do not hold references on the nodes, and neither do the uses, since
// otherwise we get into circular reference counting.
std::vector<Output> operands_as_outputs_;
};
} // namespace lazy
} // namespace torch

View File

@ -0,0 +1,20 @@
#pragma once
#include <exception>
#include <sstream>
#include <string>
#define UNIMPLEMENTED_ERROR(msg) \
{ \
std::ostringstream err; \
err << "Unimplemented Error: " << msg; \
throw std::runtime_error(err.str()); \
}
#define UNSUPPORTED_ERROR(msg) \
{ \
std::ostringstream err; \
err << "Unsupported Error: " << msg; \
throw std::runtime_error(err.str()); \
}

View File

@ -2,7 +2,7 @@
# Setup PyTorch # Setup PyTorch
#------------------------------------------------------------------------------- #-------------------------------------------------------------------------------
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/modules") list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/python/torch_mlir/cmake/modules")
include(TorchMLIRPyTorch) include(TorchMLIRPyTorch)
TorchMLIRProbeForPyTorchInstall() TorchMLIRProbeForPyTorchInstall()