Very long latency of clSetKernelArg function on embedded platforms ?

[doonny]

Hi, I am runing vitis acceleration applications on a ZCU102 boards, It was found that the operation clsetKernelArgs has very long latency. The following pic are from vitis analyzer. When I perform around ten times of clsetKernelArgs operations，it can take 3ms, which is too slow compared to the kernel execution time. There is any way to shorten the latency of the function clsetKernelArgs ?

Is this a problem of the arm processor or it is slow when using the XRT in linux mode ?

[stsoe]

@doonny I haven't heard clSetKernelArg being an issue before. Can you keep clSetKernelArg out of the critical path? E.g. associate a kernel object with a set of arguments that do not need to be changed other than content of global buffers. You can have as many kernel objects as you have buffer pools.