Transition Challenges between Pascal and Ampere GPUs in Code Processing

Accelerated Computing CUDA CUDA Programming and Performance
,
1

My machine runs on a GTX 1080 (Pascal), and I’ve developed code that processes several GB/sec. However, when I try to run the compiled code on a machine using the Ampere architecture, I encounter multiple errors. Throughout the year, I’ve dedicated myself to optimizing my code to leverage the maximum speeds achievable with my GTX. The code I’ve designed specializes in lossless data compression. Regarding these errors, it feels like an endless loop—every time I solve one issue, another arises. While I could inform the client that the code is exclusively tailored for Pascal architecture, I’m inclined to explore alternative solutions rather than limiting its use. Could NVLink potentially be the cause? Any suggestions or insights from someone who has faced a similar scenario would be greatly appreciated.

The encountered errors include:

RuntimeError: CUDA Error: mapping of buffer object failed
RuntimeError: CUDA error: invalid device pointer
RuntimeError: CUDA error: misaligned address
2

The main difference between Pascal and Ampere is Independent Thread Scheduling. If your code relies on implicit warp synchronous programming, it is probably broken.

If this is not the case, there must be another bug in the code which causes the errors. compute-sanitizer should help you to identify the exact code locations of memory errors.