gpu-compute: Skip LDS reservation for inactive waves

[Basemism]

When a local-memory instruction has no active lanes, gem5 skips issuing it to the LDS pipeline. However, the scheduler had already reserved an LDS request for it. That reservation is normally released by the pipeline, so it remained counted and later caused a panic when gem5 reused the wavefront.

Only reserve LDS resources when the instruction has active lanes. This leaves the existing handling for active LDS and flat-memory instructions unchanged.

[Copilot]

🟢 Approval recommended

The change is small, localized, and directly prevents the confirmed counter-leak path for inactive-lane local-memory instructions without altering active-lane behavior.

Pull request overview

This PR fixes a GPU compute scheduling/resource-accounting bug where a local-memory (LDS) instruction with no active lanes could still reserve LDS pipeline request resources, then get skipped before the LDS pipeline had a chance to release them—eventually triggering a panic when the wavefront is reused.

Changes:

• In Wavefront::reserveResources(), only call reserveLmResource() for local-memory instructions when exec_mask has at least one active lane.
• For local-memory instructions with an empty exec_mask, avoid incrementing LDS in-pipe request counters while still returning localMem as the execution unit ID for scheduling bookkeeping.

File summaries

File	Description
src/gpu-compute/wavefront.cc	Skips LDS request reservation for local-memory instructions with no active lanes to prevent leaked per-wave LDS request counters.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.