Fix ring buffer indexing in AIE AQL queue SubmitPackets

[aamarnat]

Motivation

Fix a ring buffer overflow bug in the AIE AQL queue that causes out-of-bounds memory access when queue indices wrap around. The SubmitPackets function was using absolute packet indices (cur_id and peak_pkt_id) directly to access the ring buffer, without applying modulo. This caused crashes after the queue size (64 packets) was exceeded.

Technical Details

In AieAqlQueue::SubmitPackets():

• Added queue_size variable to cache the queue size
• Changed packet access from queue_base + cur_id to queue_base + (cur_id % queue_size)
• Changed lookahead packet access from queue_base + peak_pkt_id to queue_base + (peak_pkt_id % queue_size)

The producer side (eg. code in GGML https://github.com/ypapadop-amd/ggml/blob/68e0094019edabe519d9f50240d4aa5378bd01b3/src/ggml-hsa/aie-kernel.cpp#L29-L31) already correctly uses modulo when writing packets, but the consumer side (in ROCr-runtime: here and here) was not using modulo when reading packets, causing the mismatch.

Test Plan

• Ran a test that performs matrix multiplication operations that exceed 64 AIE kernel dispatches
• Verified that dispatches beyond queue size 64 complete successfully
• Previously, the test would crash after approximately 64 dispatches due to out-of-bounds memory access

Test Result

Test passes with 100+ dispatches completing successfully. Queue IDs increment past the original crash point (64) and continue to work correctly with proper ring buffer wrap-around.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.