Hi all,
yesterday during the meeting on IRC someone tell about https://github.
com/guruofquality/gras/wiki/Opencl.
At the beginning, thank you for your advice, is impossible know
everything
(:
It could be a good way to write different GNU Radio blocks which have
openCL
code using __global memory in order to pass data between block, avoiding
the
transfer host /device.
But there aren’t way in openCL to pass data between kernel using __local
(called __shared in CUDA!!).
_local(which is on-chip and don’t require lot of cycles for access,if I
don’t
have banks conflict) is much faster then _global(it is off-chip,uncached
and
require hundred of cycles), hence I think that the best way is to use
only one
block to perform gpgpu calculations.
By guys,
Marco Ribero