htm-community / htm.cuda Goto Github PK

with aim to run cudaSP seamlessly from C++. Ideally a drop-in replacement.

• integrate HelloSP & SpatialPooler.cu into 1 file
  • there should be SpatialPooler.cu & a main.cu (main only calls initialization & compute of SP.cu)
• cmake detect CUDA, nvcc & nVidia HW availability
  • cmake compile using nvcc & run tests
• C++ wrapper
  • wrap all CUDA code with ifdef HAS_CUDA
  • SpatialPoolerCuda.hpp that wraps implemented cu code to c++ API (minimal initialize & compute)
  • SpatialPoolerCuda.cpp implements simple buffering
    • as c++ compute calls with a singleton data, while cu calls with block of data
  • cmake builds c++ wrapper

Add to Readme

• requirements
• steps how to build
• high level doc of this work
  • differences to c++ version/limitations (only local inhibition), forced separation of cols into "blocks"(how many,...)
  • hi-lev What does it do - CUDA-wise (loads data in batch, execute in parallel each "block",...)
• comments in SpatialPooler.cu, HelloSP.cu - why there are "duplicit" methods (getPermanences), document cuda impl a bit
• mention performance gains
• Notes on Future work (what else can be optimized (in SP, in HTM,..)

• review/compare existing HTM GPU projects:
  • https://github.com/calclavia/htm-tensorflow
  • https://github.com/JonnoFTW/htm-cl
• experiments with the GPU version
  • create & impact of huge SDRs
  • speed analysis of the optimizations - py/c++/c++optimized/cuda comparison