maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/mlir
History
Alex Zinenko 3183394328 Enable EDSC API test running through lit 
EDSC subsystem contains an API test which is a .cpp file calling the API in
    question and producing IR.  This IR is further checked using FileCheck and
    should plug into lit.  Provide a CMakeLists.txt to build the test and modify
    the lit configuration to process the source file.

--

PiperOrigin-RevId: 248794443
 		2019-05-20 13:46:09 -07:00
..
bindings/python Cleanup linalg integration test 2019-05-20 13:43:13 -07:00
examples Update cmake dependencies. 2019-05-20 13:45:08 -07:00
g3doc Unify the 'constantFold' and 'fold' hooks on an operation into just 'fold'. This new unified fold hook will take constant attributes as operands, and may return an existing 'Value *' or a constant 'Attribute' when folding. This removes the awkward situation where a simple canonicalization like "sub(x,x)->0" had to be written as a canonicalization pattern as opposed to a fold. 2019-05-20 13:44:24 -07:00
include Refactor Attribute and Type to use 'classof' instead of 'kindof' internally. If a 'classof' method is not defined, a default implementation will invoke 'kindof' on a derived type. This allows for defining supplementary Attribute/Type classes that expose additional functionality, but do not have a specific kind value. An example of this can be seen in the 'Constant(Float|Index|Int)Ops' that derive from 'ConstantOp'. 2019-05-20 13:45:51 -07:00
lib Allow for the case where ShapedType is a MemRef in fixed point math kernel utils 2019-05-20 13:46:00 -07:00
test Enable EDSC API test running through lit 2019-05-20 13:46:09 -07:00
tools Unify the 'constantFold' and 'fold' hooks on an operation into just 'fold'. This new unified fold hook will take constant attributes as operands, and may return an existing 'Value *' or a constant 'Attribute' when folding. This removes the awkward situation where a simple canonicalization like "sub(x,x)->0" had to be written as a canonicalization pattern as opposed to a fold. 2019-05-20 13:44:24 -07:00
unittests Overload arithmetic operators for SDBM expressions 2019-05-20 13:45:34 -07:00
utils Update header notices. 2019-03-29 17:43:20 -07:00
.clang-format [mlir] add .clang-format 2019-03-29 12:41:43 -07:00
CMakeLists.txt Use -force_load instead of -all_load on MacOS 2019-04-23 22:03:23 -07:00
CONTRIBUTING.md Add contributing file. 2019-03-29 17:27:11 -07:00
LICENSE.TXT Continue sketching out basic infrastructure, including an input and output 2019-03-29 12:23:51 -07:00
README.md Add references to the EuroLLVM talks in the README 2019-04-23 22:01:26 -07:00

README.md

Multi-Level Intermediate Representation Overview

The MLIR project aims to define a common intermediate representation (IR) that will unify the infrastructure required to execute high performance machine learning models in TensorFlow and similar ML frameworks. This project will include the application of HPC techniques, along with integration of search algorithms like reinforcement learning. This project aims to reduce the cost to bring up new hardware, and improve usability for existing TensorFlow users.

Note that this repository contains the core of the MLIR framework. The TensorFlow compilers we are building on top of MLIR will be part of the main TensorFlow repository soon.

How to Contribute

We'd love to accept your patches and contributions to this project soon. But we are not yet ready to accept community contributions at this time.

More resources

For more information on MLIR, please see:

Join the MLIR mailing list to hear about announcements and discussions. Please be mindful of the TensorFlow Code of Conduct, which pledges to foster an open and welcoming environment.

What is MLIR for?

MLIR is intended to be a hybrid IR which can support multiple different requirements in a unified infrastructure. For example, this includes:

  • The ability to represent all TensorFlow graphs, including dynamic shapes, the user-extensible op ecosystem, TensorFlow variables, etc.
  • Optimizations and transformations typically done on a TensorFlow graph, e.g. in Grappler.
  • Quantization and other graph transformations done on a TensorFlow graph or the TF Lite representation.
  • Representation of kernels for ML operations in a form suitable for optimization.
  • Ability to host high-performance-computing-style loop optimizations across kernels (fusion, loop interchange, tiling, etc) and to transform memory layouts of data.
  • Code generation "lowering" transformations such as DMA insertion, explicit cache management, memory tiling, and vectorization for 1D and 2D register architectures.
  • Ability to represent target-specific operations, e.g. the MXU on TPUs.

MLIR is a common IR that also supports hardware specific operations. Thus, any investment into the infrastructure surrounding MLIR (e.g. the compiler passes that work on it) should yield good returns; many targets can use that infrastructure and will benefit from it.

MLIR is a powerful representation, but it also has non-goals. We do not try to support low level machine code generation algorithms (like register allocation and instruction scheduling). They are a better fit for lower level optimizers (such as LLVM). Also, we do not intend MLIR to be a source language that end-users would themselves write kernels in (analogous to CUDA C++). While we would love to see a kernel language happen someday, that will be an independent project that compiles down to MLIR.

Compiler infrastructure

We benefitted from experience gained from building other IRs (HLO, LLVM and SIL) when building MLIR. We will directly adopt existing best practices, e.g. writing and maintaining an IR spec, building an IR verifier, providing the ability to dump and parse MLIR files to text, writing extensive unit tests with the FileCheck tool, and building the infrastructure as a set of modular libraries that can be combined in new ways. We plan to use the infrastructure developed by the XLA team for performance analysis and benchmarking.

Other lessons have been incorporated and integrated into the design in subtle ways. For example, LLVM has non-obvious design mistakes that prevent a multithreaded compiler from working on multiple functions in an LLVM module at the same time. MLIR solves these problems by having per-function constant pools and by making references explicit with function_ref.

Getting started with MLIR

The following instructions assume that you have git, ninja, and a working C++ toolchain. In the future, we aim to align on the same level of platform support as LLVM. For now, MLIR has been tested on Linux and macOS, with recent versions of clang and with gcc 7.

git clone https://github.com/llvm/llvm-project.git
git clone https://github.com/tensorflow/mlir llvm-project/llvm/projects/mlir
mkdir llvm-project/build
cd llvm-project/build
cmake -G Ninja ../llvm -DLLVM_BUILD_EXAMPLES=ON -DLLVM_ENABLE_CXX1Y=Y -DLLVM_TARGETS_TO_BUILD="host"
cmake --build . --target check-mlir

As a starter, you may try the tutorial on building a compiler for a Toy language.

MLIR talks