The FuseFlow Compiler

Prerequisites

• Initialize submodules in external directory (llvm-project, or-tools)

git submodule update --init

• Install cmake or use pre-installed cmake.
• Build MLIR under external/llvm-project using instructions provided here. For Linux (Ubuntu) optimized for memory usage. May remove --parallel 8 if you want to build single core (less memory)

cmake -G Ninja ../llvm \           
   -DLLVM_ENABLE_PROJECTS=mlir \
   -DLLVM_TARGETS_TO_BUILD="Native;" \
   -DCMAKE_BUILD_TYPE=RelWithDebInfo \
   -DLLVM_USE_SPLIT_DWARF=ON -DLLVM_ENABLE_ASSERTIONS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DLLVM_ENABLE_LLD=ON && cmake --build . --target tools/mlir/test/check-mlir --parallel 8

• Build or-tools for C++ under external/or-tools using instructions provided here. For Linux (Ubuntu) machines to reduce time and memory usage. May remove -j for single-core build.

cmake -S . -B build -DBUILD_DEPS=ON -DUSE_SCIP=OFF -DUSE_HIGHS=OFF -DUSE_COINOR=OFF && 
cmake --build build --config Release --target all -j -v

• Install protobuf compiler and c++ runtime

Building The Project And Testing

• Build using cmake

mkdir build && cd build
cmake .. && cmake --build . --parallel 8

• Run compiler flow using sam-opt and sam-translate binaries under tools/ passing in input MLIR test to compile.
• Executing sam-opt binary compiles and lowers from a linalg/sparse_tensor program to the samml dialect
• Executing sam-translate binary with --emit-proto flag emits protobuf binary of samml output program
• Example mlir input programs provided under tests directory.

# Generated protobuf output program binary is stored under /tmp/op.bin 
./tools/sam-opt --linalg-to-sam ${MLIR_TEST} | ./tools/sam-translate --emit-proto

• Apply optimizations/transformations (parallelization/vectorization). (optional)

# Using parallelization pass, stream-level parameter specifies level to be parallelized and par-factor specifies the parallelization factor
# Can parallelize multiple levels by adding more levels separated by a space
./tools/sam-opt --linalg-to-sam "--stream-parallelizer="stream-level=0 par-factor=1"" ${MLIR_TEST} | ./tools/sam-translate --emit-proto
# Using vectorization pass, stream-shape parameter specifies vector length
./tools/sam-opt --linalg-to-sam "--stream-vectorizer="stream-shape=16"" ${MLIR_TEST} | ./tools/sam-translate --emit-proto
# Using vectorization pass with block sparse enabled
./tools/sam-opt --linalg-to-sam "--stream-vectorizer="stream-shape=16 enable-block-sparse"" ${MLIR_TEST} | ./tools/sam-translate --emit-proto

Building and Testing in CLion

• Reload cmake project after installing all prerequisites
• In right drop down should see sam-opt and sam-translate. Click or search for "Edit Configurations..."
• For the sam-opt "Program arguments:" add --linalg-to-sam along with the MLIR file you want to run. For example, --linalg-to-sam /home/username/Documents/repos/samml/tests/nested_matmuls.mlir
• For the sam-translate "Program arguments:" add --emit-proto with the generated SAMML MLIR file you want to run. For example, --emit-proto /home/user/Documents/repos/samml/tests/nested_matmuls_samml.mlir
• Now you can bpress the Build sam-opt icon in CLion