An MLIR dialect for the Sparse Polyhedral Framework

By using MLIR, spf-dialect provides a more portable Sparse Polyhedral Framework (SPF) interface than the existing SPF tools it builds on. SPF is used in a wide variety of research settings. To name a few: SPF is used in synthesizing sparse tensor format conversions and contractions, and in inspector-executor compiler optimizations. With spf-dialect researchers can express SPF in MLIR, and create SPF based optimizations inside MLIR based compilers. Because of the portability of MLIR code generation, SPF based tools can now target a much broader set of hardware than previously possible.

For an MTTKRP kernel, CPU and GPU code generated using spf-dialect is competitive with code generated using the SPF C front-end (IEGenLib), and code from the PASTA benchmark suite. The SPF C front-end is not capable of generating GPU code.

Building

• Tested against LLVM 16 at sha: 570117b.

This project builds with CMake and depends on LLVM (MLIR is part of LLVM), and IEGenLib. IEGenLib will be automatically cloned and built during the spf-dialect build, however an existing source build of LLVM is required before building spf-dialect.

The LLVM docs have great instructions on how to build LLVM, though spf-dialect requires some non-standard flags. IEGenLib requires LLVM to be built with exception handling and RTTI (Run Time Type Information), which are not part of a standard LLVM build. GPU code generation also requires a few flags to be set. The spf-dialect tests require a few LLVM utilities that aren't part of a regular build. An example LLVM build with the required flags is shown here

# shallow clone of llvm to save time
git clone --depth 1 https://github.com/llvm/llvm-project.git

mkdir llvm-project/build && cd llvm-project/build

# To compile LLVM you first have to generate the build script with CMake. Taking
# a look at the flags:
# LLVM_ENABLE_PROJECTS="mlir"        - MLIR project must be enabled.
# LLVM_TARGETS_TO_BUILD="X86;NVPTX"  - NVPTX adds Nvidia GPU support.
# LLVM_ENABLE_EH=ON                  - Build llvm with exception handling. This
#                                      is disabled by default, using
#                                      fno-exceptions flag.
# LLVM_ENABLE_RTTI=ON                - Build llvm with C++ RTTI.
# LLVM_INSTALL_UTILS=ON              - Install utilities, like lit, needed for
#                                      testing spf-dialect.
# MLIR_ENABLE_CUDA_RUNNER=ON         - Required for Nvidia GPU code generation
#                                      from MLIR.
cmake -G Ninja ../llvm \
   -DLLVM_ENABLE_PROJECTS="mlir" \
   -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
   -DLLVM_REQUIRES_EH=ON \
   -DLLVM_ENABLE_RTTI=ON \
   -DLLVM_INSTALL_UTILS=ON \
   -DMLIR_ENABLE_CUDA_RUNNER=ON

# build all targets
ninja

Building spf-dalect (once LLVM is built) requires flags to set the path to the LLVM build, the path to the lit tool (built during LLVM build), and the same RTTI and execption handling flags as the LLVM build. This setup assumes that you have built LLVM (with MLIR enabled) in $BUILD_DIR. To build and launch the tests, run

mkdir build && cd build
# IEGenLib requires building with Makefiles and -j 1 for *reasons*
cmake -G "Unix Makefiles" .. \
   -DMLIR_DIR=$BUILD_DIR/lib/cmake/mlir \
   -DLLVM_EXTERNAL_LIT=$BUILD_DIR/bin/llvm-lit \
   -DLLVM_REQUIRES_EH=ON \
   -DLLVM_ENABLE_RTTI=ON
make
make check-spf