4. Known Issues and Workarounds
FPGA AI Suite Version 2024.2 has the following known issues:
-
The Python benchmark when used with the software reference model can intermittently fail during inference. Typically, you can workaround this issue by rerunning inference.
- In certain circumstances, the dla_benchmark command app on the SoC may fail with the error message "Failed to prepare DMA descriptor". To resolve this, increase MSGDMA_DESC_NUM in the altera-msgdma driver and recompile with Yocto.
- When the input layout transform is enabled, the area estimates for some IP architectures can be too low.
Streaming Flow Known Issues and Workarounds
- Only single-output graphs are supported.
- The active clock cycle counter (register offset 0x240), which is normally used to make latency measurements, has been observed to not stop counting after inference finishes in streaming configurations.
- The streaming start/stop switch (register offset 0x22c ) does not work and should not be switched off because it leads to wrong results. Streaming is always active.
- MobileNet-v2 is known to fail in the streaming flow when no external memory is used.