Correlator-Bench

A benchmarking and validation framework for the X-engine of FX telescope arrays.

Early development. Only targeting TCC and mwax-xGPU so far.

Build

• git clone --recursive to obtain xGPU and TCC
• follow instructions to build xGPU and TCC
• ensure LD_LIBRARY_PATH points to shared libraries
  • tensor-core-correlator/libtcc/libtcc.so
  • tensor-core-correlator/external/cuda-wrappers/libcu.so
  • mwax-xGPU/src/libxgpu.so
• make

Software	Minimum version
CUDA	10.0 or later
CMake	3.17 or later
gcc	9.3 or later
OS	Linux distro (amd64)

Roadmap

Critical

• complex normally distributed input values
• xGPU
  • input reorder to xGPU
  • execution of xGPU
  • output reorder xGPU to MWAX
• TCC
  • float to half conversion
  • input reorder to TCC
  • execution of TCC
  • output reorder xGPU to MWAX
• Serial
• benchmarking analysis
• output validation
• result precision comparison (TCC requires downconversion to FP16)
• change tcc reorder kernel to support > 1024 channels

Non-critical

• use -Wl and -rpath in Makefile so user doesn't have to manually change LD_LIBRARY_PATH
• generalise interface
  • change main.cu to main.c/main.cpp
  • pass arrays as void* and let user typecast to custom complex format
  • store size of input/output type in parameters
  • compile flags for AMD/CUDA
  • save output to disk to compare accuracy and performance across AMD/CUDA