Forum Discussion

Björne2's avatar
Björne2
Icon for New Contributor rankNew Contributor
2 years ago
Solved

What causes OpenCL to insert arbitration for local memory accesses?

I know FPGA OpenCL is deprecated in favor of OneAPI, but I hope you can help me anyway. I've created a MWE of a kernel for which the compiler inserts arbitration: __attribute__((uses_global_work_o...
Board Debug
High-level Design Tools
aikeu's avatar
aikeu
Icon for Regular Contributor rankRegular Contributor
2 years ago

Hi Björne2,


Maybe can refer to this link for OpenCL optimization related.

https://www.youtube.com/watch?v=1zGpN28mXN4

I will try to consult the team if there is any further info which may help.


Thanks.

Regards,

Aik Eu


Björne2's avatar
Björne2
Icon for New Contributor rankNew Contributor
2 years ago

Oh, the video gave me a hint. The compiler inserts redundant arbitration because the next iteration of the n_tics loop can start before the previous one has finished. So tmp = frontier[i] could be concurrent with frontier[dsts[src]] += 50... So the fix was to add #pragma disable_loop_pipelining to the n_tics loop to tell aoc that iterations of the outer loop can't be concurrent.