Open myurkin opened 6 years ago
This becomes more important with modern powerful GPUs, which can make up to 100 iterations of the iterative solver for the same time as CPU-based preparation of Dmatrix.
The good news, however, is that is less important for orientation averaging. For the latter computations, GPU are especially effective even with the current code (potentially also using OCL_BLAS - #199) if not a lot of -ntheta
is needed (#226).
Is your feature request related to a problem? Please describe.
Right now the Fourier transform of the D-matrix is done with the FFTW even in OpenCL mode. This brings redundant library dependence and slightly affects performance.
Describe the solution you'd like
In OpenCL mode all Fourier transforms should be made with the clFFT. This package is mature enough.
Additional context
The performance impact is minor since this transform takes place only once. So this is low-priority issue.