Forum Discussion

DongWang-BJTU's avatar
DongWang-BJTU
Icon for Occasional Contributor rankOccasional Contributor
6 years ago

How does local memory behave when used in autorun kernels for FPGA-based OpenCL design ?

I am using local memory in autorun kernels as buffers to hold data within the kernels. The autorun kernels are replicated multiple times by using pragma "__attribute__((num_compute_units(LANE_NUM)))

".

It turns out that the local memory in the first compute unit (get_compute_id(0)==0) will also be write multiple times by other compute units (get_compute_id(0)>0) during software emulation. But for hardware, the local memory will only be write one time by its own compute unit.

This is a bit strange and there is no clear defination of the behavior of how local memory works for autorun kernels.

1) Are they shared between the multiple compute instance decleared by "__attribute__((num_compute_units(LANE_NUM)))" ?

2) Do they hold previous data when restart automatically ?

2 Replies

  • HRZ's avatar
    HRZ
    Icon for Frequent Contributor rankFrequent Contributor
    6 years ago

    Your observation with the emulator is very strange. It could be a bug in the emulator. Autorun kernels are completely separate modules with no connection between them other than on-chip channels; there is no way the same buffer could be shared between them. Regarding whether data stored in buffers local to a specific autorun copy are kept or not after the copy is restarted, there is nothing mentioned in the documents and I would assume the behavior will be undefined in this case. If you need such consistency, it is probably best if you wrap your autorun kernel in a while(1) loop to prevent it from reseting; of course this could come at the cost of significant area overhead.