SILVIA is a set of LLVM transformation passes to automatically identify superword-level parallelism within an HLS FPGA design and exploit it by packing multiple operations, such as additions, multiplications, and multiply-and-adds, into a single DSP. It currently supports AMD Vitis HLS.

Prerequisites

Install AMD Vitis and set it up via:

source ${INSTALL_PATH}/Vitis/${VERSION}/settings64.sh

Where ${INSTALL_PATH} is the installation path of Vitis (e.g., /opt/Xilinx) and ${VERSION} is the version of Vitis (e.g., 2024.1).

Install From Precompiled Tarball

The recommended installation method is to use the provided precompiled LLVM/SILVIA archive instead of rebuilding LLVM and the SILVIA passes from source.

From the root of this repository, extract the archive into the expected llvm-project/install layout:

mkdir -p llvm-project
tar -xzf silvia-llvm31-x86_64-vitis2024.2.tar.gz -C llvm-project

After extraction, the expected toolchain path is:

${SILVIA_ROOT}/llvm-project/install

Verify the bundled tools and SILVIA passes are available:

./llvm-project/install/bin/opt --version
./llvm-project/install/bin/llvm-link --version
test -f ./llvm-project/install/lib/LLVMSILVIAAdd.so
test -f ./llvm-project/install/lib/LLVMSILVIAMuladd.so

Optional Source Build

Only use this path if the precompiled tarball is unavailable or incompatible with your system.

Build and install LLVM 3.1:

source install_llvm.sh

Build and install the SILVIA LLVM passes:

source build_pass.sh

Run

Update the Vitis HLS TCL synthesis script by:

1. Importing the SILVIA APIs with: source ${SILVIA_ROOT}/scripts/SILVIA.tcl, where ${SILVIA_ROOT} points to the root of the SILVIA repository (e.g., /home/user/SILVIA).
2. Setting SILVIA::ROOT to ${SILVIA_ROOT} and SILVIA::LLVM_ROOT to the precompiled toolchain path, normally ${SILVIA_ROOT}/llvm-project/install.
3. Populating SILVIA::PASSES with the list of passes to run. Each pass is specified as a dict with a mandatory OP key, which selects the pass ("add" or "muladd"). Optional keys include:
   • OP_SIZE, which defines the maximum size of the operands to pack (12 or 24 for "add", 4 or 8 for "muladd").
   • INST, which defines the instruction to pack ("add" or "sub" only when OP is "add").
   • MAX_CHAIN_LEN, which defines an upper limit to the number of cascaded DSPs (only when OP is "muladd").
4. Replacing the command csynth_design with the custom SILVIA::csynth_design.

The following snippet shows the changes to a synthesis script to run the "muladd" pass and the "add" pass with operands of up to 12 bits:

- csynth_design
+ set SILVIA_ROOT "/path/to/SILVIA"
+ source ${SILVIA_ROOT}/scripts/SILVIA.tcl
+ set SILVIA::ROOT ${SILVIA_ROOT}
+ set SILVIA::LLVM_ROOT ${SILVIA_ROOT}/llvm-project/install
+ set SILVIA::PASSES \
+ [list [dict create OP "muladd"] \
+ [dict create OP "add" OP_SIZE 12]]
+ SILVIA::csynth_design