mlir-hlo/lib/Dialect/mhlo/transforms/ral_inject_execution_contex...

140 lines
5.1 KiB
C++
Raw Permalink Normal View History

PR #50191: [MLIR][DISC] Add RAL (Runtime abstraction layer) Dialect Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/50191 DISC is a e2e flow, including both compiler side and runtime side. For runtime side, we have different targeting environments (e.g. tensorflow, pytorch, or sometimes even a standalone binary). In order to simplify the design of the compiler side, we design a Runtime Abstraction Layer (RAL) to sperate the compiler side and runtime side. Thus the compiler side only need to target RAL itself and it is the responsibility of RAL to handle the differences between different targeting environments. One of the most important functions of RAL is to manage stateful resources. To this end, it provides a context object, and hides all stateful operations behind this context, thus the compiler side itself doesn't need to care about the resource initialization. For example, a kernel must be loaded before it can be launched on GPU. However, the loading operation should only be taken once during the whole lifetime of the context in order to achieve the best performance. Based on the initialization-free interfaces provided by RAL, compiler side can focus on its core optimization logic and lets the RAL to manage the resource status. The context mentioned above is passed as a parameter to the entry function and all RAL APIs should always use the context as their first argument. This CR also provides a pass to help to ensure this property. The pass rewrites the entry function to make sure their first argument is the context. For entry function, the pass also rewrites its inputs and outputs. To be concrete, all the original inputs and outputs of the entry function are received from and sent to RAL through a sequence of RAL API calls correspondingly. The motivation behind this is to hide the implementation details of I/Os. This design may also potentially enable partial execution of the compiled module when some of the inputs are ready. Copybara import of the project: -- c4f20a89aed71181e75bcc5265723b88bde23240 by Wenyi Zhao <reyizero@gmail.com>: [MLIR][DISC] Add RAL (Runtime abstraction layer) Dialect DISC is a e2e flow, including both compiler side and runtime side. For runtime side, we have different targeting environments (e.g. tensorflow, pytorch, or sometimes even a standalone binary). In order to simplify the design of the compiler side, we design a Runtime Abstraction Layer (RAL) to sperate the compiler side and runtime side. Thus the compiler side only need to target RAL itself and it is the responsibility of RAL to handle the differences between different targeting environments. One of the most important functions of RAL is to manage stateful resources. To this end, it provides a context object, and hides all stateful operations behind this context, thus the compiler side itself doesn't need to care about the resource initialization. For example, a kernel must be loaded before it can be launched on GPU. However, the loading operation should only be taken once during the whole lifetime of the context in order to achieve the best performance. Based on the initialization-free interfaces provided by RAL, compiler side can focus on its core optimization logic and lets the RAL to manage the resource status. The context mentioned above is passed as a parameter to the entry function and all RAL APIs should always use the context as their first argument. This CR also provides a pass to help to ensure this property. The pass rewrites the entry function to make sure their first argument is the context. For entry function, the pass also rewrites its inputs and outputs. To be concrete, all the original inputs and outputs of the entry function are received from and sent to RAL through a sequence of RAL API calls correspondingly. The motivation behind this is to hide the implementation details of I/Os. This design may also potentially enable partial execution of the compiled module when some of the inputs are ready. -- 1991d4f80ab6087943956e1c0fec4940a22ab08d by Wenyi Zhao <reyizero@gmail.com>: fix PiperOrigin-RevId: 379317586
2021-06-15 02:26:41 +08:00
/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
// This file implements logic for injecting execution context to the entry
// function.
//
// Below is an example. Before Conversion:
// ```
// func @main(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>) ->
// memref<?x?xf32> {
// %0 = memref.alloc(...)
// "lmhlo.add"(%arg0, %arg1, %0) : (memref<?x?xf32>, memref<?x?xf32>,
// memref<?x?xf32>) -> memref<?x?xf32> return %0 : memref<?x?xf32>
// }
// ```
// After conversion:
// ```
// func @main(%ctx: !disc_ral.context) {
// %c0 = constant 0 : index
// %c1 = constant 1 : index
// "disc_ral.recv_input"(%ctx, %c0) : (!disc_ral.context, index) ->
// memref<?x?xf32> "disc_ral.recv_input"(%ctx, %c1) : (!disc_ral.context,
// index) -> memref<?x?xf32> %0 = memref.alloc(...) "lmhlo.add"(%arg0,
// %arg1, %0) : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) ->
// memref<?x?xf32> "disc_ral.send_output"(%ctx, %c0, %0) :
// (!disc_ral.context, index, memref<?x?xf32>) -> ()
// }
// ```
// 1. rewrite entry function (supposed that no other function directly calls the
// entry function)
// - function signature rewrite
// - return-like ops rewrite.
// 2. Currently we suppose that functions except the entry function are inlined
// to the entry function. Thus, we don't rewrite all call ops and other
// functions a.t.m. Re-visit this assumption if necessary.
#include "mlir-hlo/Dialect/mhlo/IR/disc_ral_ops.h"
#include "mlir-hlo/Dialect/mhlo/transforms/PassDetail.h"
#include "mlir/Dialect/StandardOps/IR/Ops.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/Location.h"
#include "mlir/IR/MLIRContext.h"
#include "mlir/IR/Operation.h"
#include "mlir/Pass/Pass.h"
namespace mlir {
namespace disc_ral {
namespace {
struct RalInjectExecutionContextPass
: public RalInjectExecutionContextPassBase<RalInjectExecutionContextPass> {
explicit RalInjectExecutionContextPass(const std::string& entry_func_name)
: RalInjectExecutionContextPassBase<RalInjectExecutionContextPass>::
RalInjectExecutionContextPassBase() {
this->entry_func_name_ = entry_func_name;
}
void getDependentDialects(DialectRegistry& registry) const override {
registry.insert<RalDialect>();
}
void runOnOperation() override {
ModuleOp m = getOperation();
FuncOp main = m.lookupSymbol<FuncOp>(entry_func_name_);
if (!main) {
m.emitError("entry func: " + entry_func_name_ + " not found");
signalPassFailure();
}
Location loc = main.getLoc();
FunctionType funcType = main.getType();
OpBuilder b(&main.getBody());
Block* entry_block = &main.getBody().front();
Type ctx_type = RalExecutionContextType::get(b.getContext());
// 1. Prepend context to the entry block arguments
Value ctx = entry_block->insertArgument(0u, ctx_type);
// 2. remap original arguments to recv_input ops
for (auto&& en : llvm::enumerate(
llvm::zip(funcType.getInputs(),
entry_block->getArguments().drop_front(1)))) {
Value idx = b.create<ConstantIndexOp>(loc, en.index());
Type argType = std::get<0>(en.value());
Value oldArgument = std::get<1>(en.value());
Value newInput = b.create<RecvInputOp>(loc, argType, ctx, idx);
oldArgument.replaceAllUsesWith(newInput);
}
// 3. remap all return-like ops to send_output ops
for (auto& block : main.getBody()) {
if (block.empty()) continue;
Operation& operation = block.back();
if (!operation.hasTrait<OpTrait::ReturnLike>()) continue;
b.setInsertionPoint(&operation);
for (auto& en : llvm::enumerate(operation.getOperands())) {
Value idx = b.create<ConstantIndexOp>(loc, en.index());
b.create<SendOutputOp>(loc, ctx, idx, en.value());
}
operation.eraseOperands(0, operation.getNumOperands());
}
// 4. remove unused block arguments of entry block
for (int i = 0, e = funcType.getInputs().size(); i < e; ++i) {
// continue to remove the 1st (starting from zero) argument
entry_block->eraseArgument(1);
}
// 5. set entry func to new type
main.setType(b.getFunctionType({ctx_type}, {}));
}
};
} // namespace
std::unique_ptr<OperationPass<ModuleOp>> createRalInjectExecutionContextPass(
const std::string& entry_func_name) {
return std::make_unique<RalInjectExecutionContextPass>(entry_func_name);
}
} // namespace disc_ral
} // namespace mlir