Skip to content

Latest commit

 

History

History
44 lines (30 loc) · 1.35 KB

generate_code.md

File metadata and controls

44 lines (30 loc) · 1.35 KB
Raw

Generate Code

Generation Process

The code generation (aka building the specifications) is done via a separate script, generate.py. The script is a wrapper of Cava, which configures all necessary Cava arguments for building different supported specifications.

generate.py downloads the customized AvA-LLVM module into $AVA_ROOT/llvm/build and unpacks the tarball. It then executes Cava to generate codes from the provided specifications.

Generator Script

To see the supported specifications, please run:

./generate.py -h

Here is a full list of the supported specifications, input after -s:

./generate.py -s cudadrv cudart demo gti ncsdk onnx_dump onnx_opt opencl pt_dump pt_opt qat test tf_c tf_dump tf_opt

AvA Optimizations

We implement common optimization techniques in the AvA libraries and generate necessary dependent codes using Cava, instead of injecting those optimization codes from specifications. To enable those optimizations, run generate.py with -O:

./generate.py -s tf_opt -O batching

Cava will generate a macro AVA_OPT_{CAPITALIZED_OPTIMIZATION_NAME} and pass it to both guestlib and API server at the compile time. In the above example, the guestlib's and API server's compile option will contain -DAVA_OPT_BATCHING in the generated CMakeLists.txt.

Batching

TO BE COMPLETED