torch-mlir/lib/RefBackend/LowerToRefbackrtABI.cpp

222 lines
8.5 KiB
C++
Raw Normal View History

Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +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
//
//===----------------------------------------------------------------------===//
#include "PassDetail.h"
#include "npcomp/RefBackend/RefBackend.h"
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
#include "mlir/Dialect/StandardOps/IR/Ops.h"
#include "mlir/IR/BuiltinTypes.h"
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
#include "mlir/IR/Verifier.h"
#include "mlir/Transforms/DialectConversion.h"
#include "npcomp/Dialect/Refback/IR/RefbackOps.h"
#include "npcomp/Dialect/Refbackrt/IR/RefbackrtDialect.h"
#include "npcomp/Dialect/Refbackrt/IR/RefbackrtOps.h"
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
using namespace mlir;
using namespace mlir::NPCOMP;
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
// Get the type used to represent MemRefType `type` on ABI boundaries.
// For convenience we do a cast to MemRefType internally.
static Type getABIMemrefType(Type type) {
return UnrankedMemRefType::get(type.cast<MemRefType>().getElementType(),
/*memorySpace=*/0);
}
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
//===----------------------------------------------------------------------===//
// Creating module metadata.
//===----------------------------------------------------------------------===//
// Returns true if the function signature can be expressed with the refbackrt
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
// ABI.
static bool expressibleWithRefbackrtABI(FunctionType type) {
// Currently, only memref types can be exposed at refbackrt ABI boundaries.
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
return llvm::all_of(
llvm::concat<const Type>(type.getInputs(), type.getResults()),
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
[](Type t) { return t.isa<MemRefType>(); });
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
}
static LogicalResult createModuleMetadata(ModuleOp module) {
auto moduleMetadata =
OpBuilder::atBlockBegin(module.getBody())
.create<refbackrt::ModuleMetadataOp>(module.getLoc());
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
moduleMetadata.metadatas().push_back(new Block);
Block &metadatas = moduleMetadata.metadatas().front();
OpBuilder::atBlockEnd(&metadatas)
.create<refbackrt::ModuleMetadataTerminatorOp>(module.getLoc());
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
SymbolTable symbolTable(module);
auto builder = OpBuilder::atBlockBegin(&metadatas);
for (auto func : module.getOps<FuncOp>()) {
if (symbolTable.getSymbolVisibility(func) !=
SymbolTable::Visibility::Public) {
continue;
}
// TODO: Add richer information here such as expected shapes and element
// types.
builder.create<refbackrt::FuncMetadataOp>(
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
func.getLoc(), builder.getSymbolRefAttr(func.getName()),
builder.getI32IntegerAttr(func.getNumArguments()),
builder.getI32IntegerAttr(func.getNumResults()));
if (!expressibleWithRefbackrtABI(func.getType()))
return func.emitError() << "func not expressible with refbackrt ABI";
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
}
return success();
}
//===----------------------------------------------------------------------===//
// Dialect conversion.
//===----------------------------------------------------------------------===//
namespace {
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
class LowerAssertOp : public OpConversionPattern<AssertOp> {
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
matchAndRewrite(AssertOp op, ArrayRef<Value> operands,
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
ConversionPatternRewriter &rewriter) const override {
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
AssertOp::Adaptor adaptor(operands);
// The refbackrt runtime function aborts if the argument is true, rather
// than when it is false as an `assert` does. So negate the predicate (by
// xor'ing with 1).
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
auto c1 = rewriter.create<ConstantOp>(
op.getLoc(), rewriter.getIntegerAttr(rewriter.getI1Type(),
APInt(/*numBits=*/1, /*val=*/1)));
Value assertFailed = rewriter.create<XOrOp>(op.getLoc(), adaptor.arg(), c1);
rewriter.replaceOpWithNewOp<refbackrt::AbortIfOp>(op, assertFailed,
op.msgAttr());
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
return success();
}
};
} // namespace
namespace {
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
// At ABI boundaries, convert all memrefs to unranked memrefs so that they have
// a fixed ABI.
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
class FuncOpSignatureConversion : public OpConversionPattern<FuncOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
matchAndRewrite(FuncOp op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const override {
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
FunctionType type = op.getType();
TypeConverter::SignatureConversion entryConversion(type.getNumInputs());
if (failed(typeConverter->convertSignatureArgs(type.getInputs(),
entryConversion)))
return rewriter.notifyMatchFailure(op, "could not convert inputs");
SmallVector<Type, 1> newResultTypes;
if (failed(typeConverter->convertTypes(type.getResults(), newResultTypes)))
return rewriter.notifyMatchFailure(op, "could not convert outputs");
rewriter.updateRootInPlace(op, [&] {
// Update the function type.
op.setType(FunctionType::get(op.getContext(),
entryConversion.getConvertedTypes(),
newResultTypes));
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
// Rewrite the entry block.
Block &oldEntry = op.getBody().front();
Block &newEntry =
*rewriter.applySignatureConversion(&op.getBody(), entryConversion);
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&newEntry);
BlockArgument newArg, oldArg;
for (auto newAndOldArg :
llvm::zip(newEntry.getArguments(), oldEntry.getArguments())) {
std::tie(newArg, oldArg) = newAndOldArg;
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
auto memref = rewriter.create<MemRefCastOp>(op.getLoc(), newArg,
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
oldArg.getType());
rewriter.replaceUsesOfBlockArgument(oldArg, memref);
}
});
return success();
}
};
} // namespace
namespace {
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
// At the return ABI boundaries, convert to the ABI type.
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
// This pattern is needed to trigger the type conversion mechanics to do a
// target materialization.
class RewriteReturnOp : public OpConversionPattern<ReturnOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
matchAndRewrite(ReturnOp op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const override {
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
rewriter.replaceOpWithNewOp<ReturnOp>(op, operands);
return success();
}
};
} // namespace
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
static LogicalResult doDialectConversion(ModuleOp module) {
auto *context = module.getContext();
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
TypeConverter typeConverter;
typeConverter.addConversion([](Type type) { return type; });
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
typeConverter.addConversion(
[](MemRefType type) { return getABIMemrefType(type); });
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
typeConverter.addTargetMaterialization(
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
[](OpBuilder &builder, UnrankedMemRefType type, ValueRange inputs,
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
Location loc) -> Value {
assert(inputs.size() == 1);
[RefBackend] Fix leaks related to ABI boundaries. Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks except for those due to buffers created internally to the codegenned code itself (up next I'll add the buffer deallocation pass to fix those). The main change is that instead of attempting to pass `refbackrt::Tensor` to the codegenned function directly, we make all the ABI types be UnrankedMemRef which gets passed awkwardly (but workably) as a `{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why refbackrt::Tensor wasn't workable is that is that MLIR doesn't really have a way to deal with the lifetime of unranked memref descriptors that happen inside the function, which is inevitably what would happen in the old code that would emit runtime calls to `refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to `refbackrt::Tensor` inside the codegenned code. So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no real sound basis for valid lifetime management, we now have a lovely piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely seems to be sound. We rely on the codegenned code having these properties, which it seems to have: - it won't free memref descriptors or their backing buffer for arguments of UnrankedMemRef type. - it will allocate a separate memref descriptor for each result UnrankedMemRef (which is ensured by having a separate memref_cast for each) - we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers) to avoid double-freeing in the case of aliasing of the backing buffer (including backing buffers for arguments feeding into results) - to catch the case of statically allocated data (which we need to avoid passing to `free`) , check if the `allocatedPtr` is (no joke) equal to `0xDEADBEEF`, because there is otherwise no way to distinguish statically allocated from malloc'ed data... (std.global_memref lowering to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`, presumably mainly as a debugging thing) Even with all this, we *still* need to (internally to refbackrt::invoke) make copies of all inputs/outputs! And the details of how the LLVM-level ABI gets laid out for e.g. function arguments/returns is still super tricky. This really highlights how deficient memref is as the general runtime type for our use case. It's stewing in my mind how best to improve the situation. My general gut feeling is that IREE's abstractions for this are "right", but I need to think more how to distill those aspects of IREE's design in a "reference" way for RefBackend. Some implementation notes: - In terms of how this is implemented, this did catch a bug in our ABI wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to work before through some combination of npcomprt::Tensor being passed as a single pointer + probably me infinite-monkey-ing it until it worked) - This actually removes 2 out of the 3 compiler runtime functions (the only one left is "abort_if". (most of the memref descriptor code moved from CopmilerRuntime.cpp to Runtime.cpp) - this also means deleting `refbackrt.from_memref` and `refbackrt.to_memref`
2020-11-25 09:18:57 +08:00
return builder.create<MemRefCastOp>(
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
loc, inputs[0], getABIMemrefType(inputs[0].getType()));
});
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
OwningRewritePatternList patterns;
ConversionTarget target(*context);
target.addLegalDialect<refbackrt::RefbackrtDialect>();
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
target.addLegalDialect<StandardOpsDialect>();
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
patterns.insert<FuncOpSignatureConversion>(typeConverter, context);
target.addDynamicallyLegalOp<FuncOp>(
[&](FuncOp op) { return typeConverter.isSignatureLegal(op.getType()); });
patterns.insert<RewriteReturnOp>(typeConverter, context);
target.addDynamicallyLegalOp<ReturnOp>(
[&](ReturnOp op) { return typeConverter.isLegal(op); });
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
Totally rework RefE2E tensor to memref flow. (#42) This now gets the overall "RefE2E" compilation stack to a point that I'm fairly happy with. We simplify it by mostly embracing the "descriptor" view of the world. The overall flow is best understood by reading through the createE2ELoweringPipeline function in lib/E2E/E2E.cpp That function creates a pass pipeline that lowers from "TCF" (which is ~numpy level of abstraction) down to LLVM IR. A brief high-level summary of what happens there: 1. TCF to TCP conversion. This involves reifying error handling in the form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir 2. Lowering shape constraints. This converts shape constraints into eager error-handling code. See test/E2E/lower-shape-constraints.mlir This pass will soon go upstream. Because this lowers to std.assert, some later passes like LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this through e2e. See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test that properly aborts in case of an error. 3. Lowering tensors to memrefs. This is done via a series of passes rather than an single mega conversion. Unlike the previous code that mixed in the npcomprt ABI stuff here, it's now a very clean "pure memref" conversion. See test/E2E/lower-*-to-memref.mlir and lib/E2E/TensorToMemref/ Most of the changes are concentrated here. 4. As part of the above, we use the upstream ConvertShapeToStandard for lowering shapes. 5. We lower linalg to loops and lower loops to CFG using upstream passes. 6. Rewrite the "ABI" boundaries of the program to npcomprt data structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and how global tensor constants are represented. One of the major improvements in this commit is that now it's a very clean rewrite that just replaces memrefs on ABI boundaries with !npcomprt.tensor (before there was a get_extent function that is not needed). See test/E2E/lower-to-npcomprt-abi.mlir 7. Lower to LLVM with upstream mlir patterns + some patterns for the npcomprt lowerings. One aspect here that is still a remnant of a non-descriptor-based tensor to memref flow is the BypassShapes + LowerShapedResultsToMemref. BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results (basically a "tie_shape" kind of op), and then LowerShapedResultsToMemref uses those annotations to allocate output buffers while lowering the "tensor compute ops". Note that there are very few "tensor compute" ops currently supported (tcp.add + tcp.broadcast_to), so we just hardcode them in both passes. Realistically, I expect this to go away as we fully embrace the descriptor-based approach for simplicity, so don't look too deep into it.
2020-09-17 08:31:40 +08:00
patterns.insert<LowerAssertOp>(context);
target.addIllegalOp<AssertOp>();
return applyPartialConversion(module, target, std::move(patterns));
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
}
namespace {
// This pass lowers the public ABI of the module to the primitives exposed by
// the refbackrt dialect.
class LowerToRefbackrtABI
: public LowerToRefbackrtABIBase<LowerToRefbackrtABI> {
2020-09-10 15:23:46 +08:00
void getDependentDialects(DialectRegistry &registry) const override {
registry.insert<refbackrt::RefbackrtDialect>();
2020-09-10 15:23:46 +08:00
}
void runOnOperation() override {
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
ModuleOp module = getOperation();
// Before we lower anything, capture any needed metadata about the argument
// lists that will be needed for safely invoking the raw runtime functions
// later. (for example, number of expected arguments/results, types,
// etc.)
if (failed(createModuleMetadata(module)))
return signalPassFailure();
// Now do the actual conversion / lowering.
if (failed(doDialectConversion(module)))
return signalPassFailure();
}
};
} // namespace
std::unique_ptr<OperationPass<ModuleOp>>
mlir::NPCOMP::createLowerToRefbackrtABIPass() {
return std::make_unique<LowerToRefbackrtABI>();
Rework e2e flow to use new "npcomprt" This ~totally reworks the existing "runtime" stuff to be more principled and usable, such as from Python. It's still not fully production-quality, mainly in the department of memory management (e.g. it currently leaks memory; we need to figure out "who frees memrefs" + the analysis and transformation needed to do that (maybe use upstream buffer allocation pass?)). The user API is in include/npcomp/runtime/UserAPI.h, though include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper. The stuff under {include,lib}/runtime is totally firewalled from the compiler and tiny (<6kB, though no attention has gone into optimizing that size). For example, we don't link in libSupport into the runtime, instead having our own bare bones replacements for basics like ArrayRef (the JITRuntime helps with bridging that gap, since it *can* depend on all common LLVM utilities). The overall features of npcomprt is that it exposes a module that with multiple function entry points. Each function has arguments and results that are tensor-valued, and npcomprt::Tensor is the runtime type that is used to interact with that (and a npcomprt::Ref<T> reference-counting wrapper is provided to wrap npcomprt::Tensor in the common case). From an implementation perspective, an npcomprt module at the LLVM/object/binary level exposes a single module descriptor struct that has pointers to other metadata (currently just a list of function metadata descriptors). All interactions with the npcomp runtime are keyed off of that module descriptor, including function lookups and dispatching. This is done to dodge platform ABI issues and also allow enough reflection to e.g. verify provided arguments. Most of the compiler-side work here was in LowerToNpcomprtABI and LowerToLLVM. Also, - Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting annoying to type the underscores/caps. - misc improvements to bash_helpers.sh
2020-07-09 08:15:40 +08:00
}