python - Can I link numpy with AMD's gpu accelerated blas library -
i reconized numpy can link blas, , thought of why not using gpu accelerated blas library.
did use so?
update (2014-05-22)
amd has produced beta release of amd core math library (acml) version 6.0 can offload fft , blas functions gpu using clmath internally. announcement here: acml beta 6.0 release leverages power of heterogeneous compute. caveat here input data must transferred cpu gpu , output data returned cpu on each blas or fft call. therefore, amd has bunch of scripts tuning when problem large enough acml use gpu instead of cpu.
for sake of completeness, i'll mention nvidia supports similar functionality nvblas library relies on cublas , cuda won't work on nvidia gpus.
original answer
unfortunately, amd's gpu accelerate blas library cannot directly link numpy or other application expecting standard cpu-based blas library. reason existing gpu blas libraries require 1 first copy matrices gpu before calling blas functions. requires modify numpy copying.
edit: clyther looks can replace of numpy , converts opencl. see here: http://srossross.github.io/clyther/for_numpy_users.html
Comments
Post a Comment