BUDDY MLIR

An MLIR-Based Ideas Landing Project.

Getting Started

LLVM/MLIR Dependencies

This project uses LLVM/MLIR as an external library. Please make sure the dependencies are available on your machine.

Clone and Initialize

$ git clone [email protected]:buddy-compiler/buddy-mlir.git
$ cd buddy-mlir
$ git submodule update --init

Build and Test LLVM/MLIR

$ cd buddy-mlir
$ mkdir llvm/build
$ cd llvm/build
$ cmake -G Ninja ../llvm \
    -DLLVM_ENABLE_PROJECTS="mlir" \
    -DLLVM_TARGETS_TO_BUILD="host;RISCV" \
    -DLLVM_ENABLE_ASSERTIONS=ON \
    -DCMAKE_BUILD_TYPE=RELEASE
$ ninja
$ ninja check-mlir

Build buddy-mlir

$ cd buddy-mlir
$ mkdir build
$ cd build
$ cmake -G Ninja .. \
    -DMLIR_DIR=$PWD/../llvm/build/lib/cmake/mlir \
    -DLLVM_DIR=$PWD/../llvm/build/lib/cmake/llvm \
    -DLLVM_ENABLE_ASSERTIONS=ON \
    -DCMAKE_BUILD_TYPE=RELEASE
$ ninja check-buddy

Dialects

Bud Dialect

Bud dialect is designed for testing and demonstrating.

DIP Dialect

DIP dialect is designed for digital image processing abstraction.

Tools

buddy-opt

The buddy-opt is the driver for dialects and optimization in buddy-mlir project.

Convolution Optimization

So far, we provide the 2D convolution vectorization pass conv-vectorization. The pass implements the Coefficients Broadcasting algorithm with Strip Mining strategy, and the strip mining size is configurable. Take the size of 256 as an example, you can use the tool with the following configuration.

$ buddy-opt <input> -conv-vectorization="strip-mining=256"

• Conversion example

We provide a function with linalg.conv_2d operation. You can use the following commands to print the conversion result.

$ cd buddy-mlir/build/bin
$ ./buddy-opt ../../examples/ConvOpt/conv2d.mlir -conv-vectorization="strip-mining=256"

• Edge detection example

We also provide an edge detection example to show the optimization. The conv-vectorization pass is responsible for lowering the linalg.conv_2d with our algorithm. And then we use mlir-translate and llc tools to generate the object file. At last, we call the MLIR convolution function in a C++ program.

Please use a Linux machine with OpenCV installed to play around. You should specify the <strip mining size> (e.g. 256) and <ISA vector extension> (e.g. avx512f).

$ cd buddy-mlir/build
$ cmake -G Ninja .. \
    -DBUDDY_EXAMPLES=ON \
    -DBUDDY_CONV_OPT_STRIP_MINING=<strip mining size> \
    -DBUDDY_CONV_OPT_ATTR=<ISA vector extension>
$ ninja edge-detection

We provide an image at buddy-mlir/examples/ConvOpt/images/YuTu.png, which is the robotic lunar rover that formed part of the Chinese Chang'e 3 mission. You can detect the edge of the image with edge-detection.

$ cd bin
$ ./edge-detection ../../examples/ConvOpt/images/YuTu.png result.png

We also provide the performance comparison between our buddy-opt tool and other state-of-the-art approaches. For more details, please see convolution comparison.

Lowering DIP Dialect

$ buddy-opt <input> -lower-dip="DIP-strip-mining=${BUDDY_DIP_OPT_STRIP_MINING}"

• Conversion example:

$ cd buddy-mlir/build/bin
$ ./buddy-opt ../../examples/DIPDialect/corr2d.mlir --lower-dip="DIP-strip-mining=${BUDDY_DIP_OPT_STRIP_MINING}"

• Edge detection example:

Build and run the example.

Note: Please use a Linux machine with OpenCV installed to play around.

$ cd buddy-mlir/build
$ cmake -G Ninja .. -DBUDDY_EXAMPLES=ON -DBUDDY_DIP_OPT_STRIP_MINING=256
$ ninja correlation2D
$ cd bin
$ ./correlation2D ../../examples/ConvOpt/images/YuTu.png result-dip.png

Note: Maximum allowed value of BUDDY_DIP_OPT_STRIP_MINING for producing correct result is equal to image width.

Lowering Bud Dialect

$ buddy-opt <input> -lower-bud

Example:

$ cd buddy-mlir/build/bin
$ ./buddy-opt ../../examples/BudDialect/TestConstant.mlir --lower-bud

Benchmarks

The benchmarks in this repo use JIT tool (mlir-cpu-runner) as the execution engine. For AOT benchmarks, please see buddy-benchmark repo.

We provide the following benchmarks:

• Conv2D

$ cd buddy-mlir/benchmark
$ make

For more features and configurations, please see the benchmark document.