Avoid designing systems where one kernel writes an intermediate result to global memory and another kernel reads this data back from global memory. Instead, implement a DPC++ pipe (described in Pipes) between the producer and consumer kernels for direct data transfer. Alternatively, you can merge both kernels into a single larger kernel and use helper functions to logically separate the two original kernels.

The

Intel® oneAPI

DPC++/C++

Compiler

implements local memory in FPGAs differently than in GPUs. If your DPC++ kernel contains code to avoid GPU-specific local memory bank conflicts, remove that code because the compiler generates hardware that avoids local memory bank conflicts automatically whenever possible.