Hi Alexis
the OpenCL Embedded Profile specification defines the minimum requirement for local
memory to be 1KB to pass conformance testing. Based on algorithm analysis and profiling
different image and computer vision algorithms, it was found that a 1KB local memory size
was too small to benefit those algorithms. In most cases, those algorithms actually slowed
down when using 1KB local memory. To increase performance recommended not using
local memory since it is more efficient to transfer larger chunks of data from system memory
to keep the OpenCL pipeline full.
If local memory type is CL_GLOBAL, the local memory is emulated using global memory,
and the performance will be the same as global memory. There is extra overhead on data
copy from global to local, which will slow down the performance.
Best regards
igor
-----------------------------------------------------------------------------------------------------------------------
Note: If this post answers your question, please click the Correct Answer button. Thank you!
-----------------------------------------------------------------------------------------------------------------------