graph-engine

branches

• dev : Latest development should be merged to dev branch. This is the default branch.
• phx-dev : Phoenix project release branch

If you create branches for pull requests, and they become stale, please delete them.

documentation

see doxygen

continuous integration

When adding code to this repo, please create your own fork/branch and submit a merge request.

Merge requests will trigger a pipeline, that builds, and runs sanity tests.

build

scl enable devtoolset-9 bash
source /opt/xilinx/xrt/setup.sh
mkdir build
cd build
cmake ..
make -j

test

cd build/tests
ctest # Optionally use --verbose to see all test outputs

novel ideas

• Bring your own Kernel Runner
  • Just need to define your own MyKernelRunner.hpp
  • Can support pre/postprocess kernels
• Kernel Runners will not manage internal buffers
  • No need for get_inputs() at the subgraph level (Only needed at graph level)
• Single Thread Pool can execute all Kernel Runners execute_async function
• Superclasses of vart::TensorBuffer allow direct sharing of xrt::bo objects
• GraphEngine provides a GraphRunner that can
  • Globally allocate buffers based on the sequence of runners
  • Implement client facing VART APIs for THE WHOLE GRAPH not a subgraph

use cases

• Whole App Acceleration
• Inference Server