perf: Order OVRTX render-var reads on the consuming stream, and block the host on Linux

[pv-nvidia]

Description

Reading an OVRTX render var has to be ordered against render completion. On Linux, blocking the calling thread on the render-completion event measures faster end to end than a GPU-side wait, so that is now what Linux does.

Measured across three camera tasks at 256/1024/4096 envs, this is +14% to +73% throughput (mean +36%), and restores parity with ovrtx 0.3 on every case measured.

Other platforms order the read on the consuming Warp stream. ISAAC_LAB_OVRTX_DISABLE_LINUX_CUDA_CPU_SYNC=1 puts Linux on that ordering too — an escape hatch for when the trade-off changes; it is not currently faster.

All mapping sites now go through one helper, OVRTXRenderer._map_render_var_to_dlpack(). Camera outputs are unchanged — only the read ordering differs.

Reported in https://nvbugspro.nvidia.com/bug/6566453

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Performance improvement

Screenshots

None — no visual change.

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[greptile-apps]

Greptile Summary

The PR centralizes CUDA render-var mapping and changes Linux ordering to use a host-side render-completion wait, avoiding the reported GPU scheduling bottleneck while retaining an environment-controlled GPU-side path.

• Routes CUDA camera render-var mappings through a shared context-managed helper.
• Uses the consuming Warp stream for GPU-side synchronization and a no-sync mapping followed by a host wait on Linux.
• Adds contract tests for platform and environment selection, synchronization ordering, DLPack conversion, and mapping release.
• Documents the Linux throughput improvement and opt-out setting.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
source/isaaclab_ov/isaaclab_ov/renderers/ovrtx_renderer.py	Centralizes render-var mapping and selects host-side synchronization on Linux while preserving consuming-stream synchronization elsewhere.
source/isaaclab_ov/test/test_ovrtx_renderer_contract.py	Adds focused contract coverage for synchronization-mode selection, map arguments, wait ordering, DLPack conversion, and cleanup on failure.
source/isaaclab_ov/changelog.d/pv-ovrtx-render-var-sync.rst	Documents the Linux camera-output throughput improvement and the environment-variable opt-out.

Sequence Diagram

sequenceDiagram
    participant Caller as Camera output processing
    participant Helper as _map_render_var_to_dlpack
    participant OVRTX as RenderVarOutput
    participant Warp as Consuming Warp stream
    alt Linux host synchronization
        Helper->>OVRTX: "map(CUDA, sync_stream=0)"
        Helper->>OVRTX: mapping.wait()
        OVRTX-->>Helper: render complete
    else GPU-side synchronization
        Helper->>Warp: get consuming CUDA stream
        Helper->>OVRTX: "map(CUDA, sync_stream=stream)"
        OVRTX-->>Warp: enqueue render-completion wait
    end
    Helper->>Warp: wp.from_dlpack(mapping)
    Helper-->>Caller: mapped Warp array
    Caller->>Warp: enqueue output extraction
    Helper->>OVRTX: release mapping

Loading

_{Reviews (2): Last reviewed commit: "Bot review feedback" | Re-trigger Greptile}