[QDP] Fix async amplitude pipeline out-of-bounds write

[rich7420]

Summary

When processing large data in chunks, launch_amplitude_encode() is called with state_len (full state vector size) instead of chunk_len (actual chunk size). This causes the kernel to write state_len elements starting from the chunk offset, which exceeds bounds for chunks after the first one.

[Rutuja123-dos]

Hi @rich7420,
I was reading through this issue and wanted to check before starting any work.

From what I understand, 'launch_amplitude_encode()' is currently being called with 'state_len' instead of 'chunk_len', which could lead to out-of-bounds writes when processing later chunks.

Please let me know if someone is already working on this, or if it’s okay for me to take it up.
Thanks!

[rich7420]

@Rutuja123-dos thanks for your asking!
yes, I think you got the point. It seems nobody doing this feel free to take it.
Meanwhile plz add a test for this part.

[alisha-1000]

Hi @rich7420,

I noticed that this issue was previously linked to PRs #773 and #844 related to regression tests for async amplitude chunk bounds.

Could you please clarify whether the core out-of-bounds write in launch_amplitude_encode() is still unresolved, or if only test coverage remains to be improved?

If the main fix is still pending, I’d be happy to investigate the kernel launch parameter mismatch (state_len vs chunk_len) and propose a fix along with additional regression tests.

Thanks!