ScaleHLS Project (scalehls)

This project aims to create a framework that ultimately converts an algorithm written in a high level language into an efficient hardware implementation. With multiple levels of intermediate representations (IRs), MLIR appears to be the ideal tool for exploring ways to optimize the eventual design at various levels of abstraction (e.g. various levels of parallelism). Our framework will be based on MLIR, it will incorporate a backend for high level synthesis (HLS) C/C++ code. However, the key contribution will be our parametrization and optimization of a tremendously large design space.

Quick Start

1. Install LLVM and MLIR

IMPORTANT This step assumes that you have cloned LLVM from (https://github.com/circt/llvm/tree/main) to $LLVM_DIR and checked out the main branch. To build LLVM and MLIR, run:

$ mkdir $LLVM_DIR/build
$ cd $LLVM_DIR/build
$ cmake -G Ninja ../llvm \
    -DLLVM_ENABLE_PROJECTS="mlir" \
    -DLLVM_TARGETS_TO_BUILD="X86;RISCV" \
    -DLLVM_ENABLE_ASSERTIONS=ON \
    -DCMAKE_BUILD_TYPE=DEBUG
$ ninja
$ ninja check-mlir

2. Install ScaleHLS

This step assumes this repository is cloned to $SCALEHLS_DIR. To build and launch the tests, run:

$ mkdir $SCALEHLS_DIR/build
$ cd $SCALEHLS_DIR/build
$ cmake -G Ninja .. \
    -DMLIR_DIR=$LLVM_DIR/build/lib/cmake/mlir \
    -DLLVM_DIR=$LLVM_DIR/build/lib/cmake/llvm \
    -DLLVM_ENABLE_ASSERTIONS=ON \
    -DCMAKE_BUILD_TYPE=DEBUG
$ ninja check-scalehls

3. Try ScaleHLS

After the installation and test successfully completed, you should be able to play with:

$ export PATH=$SCALEHLS_DIR/build/bin:$PATH
$ cd $SCALEHLS_DIR

$ # Loop and pragma-level optimizations, performance estimation, and C++ code generation.
$ scalehls-opt samples/polybench/syrk.mlir \
    -affine-loop-perfection -remove-variable-bound -affine-loop-normalize \
    -partial-affine-loop-tile="tile-level=1 tile-size=4" \
    -legalize-to-hlscpp="top-function=test_syrk" -loop-pipelining="pipeline-level=1" \
    -affine-store-forward -simplify-memref-access -array-partition -cse -canonicalize \
    -qor-estimation="target-spec=config/target-spec.ini" \
    | scalehls-translate -emit-hlscpp

$ # Benchmark generation, dataflow-level optimization, HLSKernel lowering and bufferization.
$ benchmark-gen -type "cnn" -config "config/cnn-config.ini" -number 1 \
    | scalehls-opt -legalize-dataflow -split-function \
    -hlskernel-bufferize -hlskernel-to-affine -func-bufferize -canonicalize

$ # Put them together.
$ benchmark-gen -type "cnn" -config "config/cnn-config.ini" -number 1 \
    | scalehls-opt -legalize-dataflow -split-function \
    -hlskernel-bufferize -hlskernel-to-affine -func-bufferize \
    -affine-loop-perfection -affine-loop-normalize \
    -legalize-to-hlscpp="top-function=auto_gen_cnn" \
    -affine-store-forward -simplify-memref-access -cse -canonicalize \
    -qor-estimation="target-spec=config/target-spec.ini" \
    | scalehls-translate -emit-hlscpp

Integration with ONNX-MLIR

If you have installed ONNX-MLIR or established ONNX-MLIR docker to $ONNXMLIR_DIR following the instruction from (https://github.com/onnx/onnx-mlir), you should be able to run the following integration test:

$ cd $SCALEHLS_DIR/sample/onnx-mlir/resnet18

$ # Export PyTorch model to ONNX.
$ python export_resnet18.py

$ # Parse ONNX model to MLIR.
$ $ONNXMLIR_DIR/build/bin/onnx-mlir -EmitONNXIR resnet18.onnx

$ # Lower from ONNX dialect to Affine dialect.
$ $ONNXMLIR_DIR/build/bin/onnx-mlir-opt resnet18.onnx.mlir \
    -shape-inference -convert-onnx-to-krnl -pack-krnl-constants \
    -convert-krnl-to-affine > resnet18.mlir

$ # (Optional) Print model graph.
$ scalehls-opt resnet18.tmp -print-op-graph 2> resnet18.gv
$ dot -Tpng resnet18.gv > resnet18.png

$ # Legalize the output of ONNX-MLIR, optimize and emit C++ code.
$ scalehls-opt resnet18.mlir -legalize-onnx -affine-loop-normalize -canonicalize \
    -legalize-dataflow="min-gran=2 insert-copy=false" -split-function \
    -convert-linalg-to-affine-loops -affine-loop-fusion \
    -legalize-to-hlscpp="top-function=main_graph" \
    | scalehls-translate -emit-hlscpp

Ablation Study (Deprecated)

If Vivado HLS (2019.1 tested) is installed on your machine, running the following script will report the HLS results for some benchmarks (around 8 hours on AMD Ryzen7 3800X for all 33 tests).

For the ablation_test_run.sh script, -n determines the number of tests to be processed, the maximum supported value of which is 33; -c determines from which test to begin to rerun the C++ synthesis. The generated C++ source code will be written to sample/cpp_src; the Vivado HLS project will be established in sample/hls_proj; the collected report will be written to sample/test_results; the test summary will be generated to sample.

$ cd $SCALEHLS_DIR/sample
$ ./ablation_test_run.sh -n 33 -c 0

References

1. MLIR documents
2. mlir-npcomp github
3. onnx-mlir github
4. circt github
5. comba github
6. dahlia github