Application event waits: CPU cost and latency

NVRM_EVENT_WAIT=1 enables the existing graphics-completion event wait for ordinary applications. On one RTX 5060 Laptop GPU, this reduced the CPU time spent in two synchronous workloads, at the cost of higher latency in the short compute test. This is an opt-in measurement result, not a proposed change to the shipped default.

In nvk/nvk-macos.patch, nvkmd_nvrm_wait_block() defaults to event waits for WindowServer and polling/backoff for other processes. An explicit NVRM_EVENT_WAIT=1 enables event waits after 20 microseconds, with a bounded 2-millisecond wait and completion-state rechecks. The code falls back when a suitable event is unavailable. The setting does not increase GPU clocks or its power limit, and WindowServer already takes the event path by default.

For a machine-local trial, back up /Library/GPUBundles/nvmtl/nvrm610.conf, then set exactly one NVRM_EVENT_WAIT=1 entry while preserving its other settings. Reopen the application so it reads the new configuration. Restore the previous configuration to undo the trial. No driver reinstall or kernel-cache rebuild is needed for this setting.

Release builds clear inherited NVRM_ tuning variables in the plugin constructor. An environment variable set before launching an application is therefore not a reliable A/B test. The probe below loads the installed plugin first, then sets the single wait-policy override before it creates the Metal device. The override is confined to that probe process.

• NVIDIA GeForce RTX 5060 Laptop GPU, 8 GiB, PCI ID 10de:2d19, NVK GB206-A.
• Intel Core i7-13700HX, 16 GiB RAM, AC power.
• macOS 26.7.1 (25G241), SMBIOS MacBookPro16,2.
• Built-in display at 2560 x 1600, 180 Hz, NVIDIA-driven.
• Driver source baseline: v1.13.0 (0960b2ef5dbc4a573056d4c8e042a5bd2e15d651).
• Both arms used the same locally repaired plugin and Tahoe NVAccel. The plugin included the specialized AIR symbol / sampler linking fixes from PR #79, and NVAccel had been rebuilt for the Tahoe ABI. This is not an unmodified release package result and does not qualify macOS 26 support generally.

Each workload was run in five paired rounds, reversing the policy order on alternate rounds. Every process did two warmups and 25 measured iterations: 125 measurements per workload and policy. The table reports the median of the five per-run medians, not a pooled median. The raw timing samples and per-run results are in EVENT-WAIT-RTX5060.json.

Elapsed time includes encoding, submission, execution, and synchronous waiting. It is measured with CLOCK_MONOTONIC, not GPU timestamps. Process CPU time is getrusage(RUSAGE_SELF) user plus system time across all process threads, so it can exceed elapsed time. Compilation and the final compute correctness check are outside the timed iterations. These are not whole-system CPU utilization or GPU throughput measurements.

The compute workload transforms 1,048,576 integers with 64 dependent integer iterations each; every result is compared with a CPU reference. The image test renders a 1024 x 1024 RGBA8 procedural image through a radius-10 Core Image Gaussian blur into a CPU bitmap, with intermediate caching disabled. Every image run had identical output bytes by hash and opaque alpha; the test does not establish independent image-reference accuracy.

Image timings varied considerably between runs. The five polling run medians were 5.161, 5.608, 3.047, 6.145, and 8.171 ms; event-wait medians were 3.714, 4.558, 4.517, 3.690, and 3.926 ms. No confidence interval or universal speedup is claimed. No GPU power, clock, temperature, fan-speed, or battery-life data was collected. The test was short, on one system, with its normal desktop session running.

• The compute reference checks passed for both policies.
• All paired image results had the same pixel hash (10771374122383686649).
• A cross-queue test committed a consumer's event wait before the producer's fill and signal. Both policies passed 32 rounds and verified all 128 MiB.
• Shared-to-private, private-to-shared, and private-to-private 32 MiB copies passed full byte comparisons. Their timings are not used to claim a speedup.
• No probe timeout or WindowServer restart occurred in the reported paired runs. The driver's VRAM refusal counter remained zero.
• After enabling the setting in the local configuration, new processes with no overrides logged event wait ON and passed the same functional checks. No driver binary was replaced by this configuration trial.

Existing icon and Tahoe ABI repairs are disclosed as the test baseline, not submitted as new work here. Alternative compiler flags, memory-mapping changes, and an experimental wait-threshold patch were not deployed or included.

On an x86_64 Mac with the installed NVMTL driver as the default Metal device and Apple Command Line Tools available, run from the repository root:

python3 tools/bench_event_wait.py /tmp/nvmtl-event-wait-results

Use a new output directory. The runner compiles the probe into a temporary directory, alternates both policies for five rounds, verifies the event-wait log, compares image hashes, and stops on a probe failure, timeout, or a changed WindowServer PID. It then checks cross-queue synchronization and buffer copies with both policies. --rounds 1 is a smoke test, not the five-round measurement. The tool writes local results and logs; it does not upload anything or edit the driver configuration. Normal driver shader-cache activity can still occur.

Review local stderr logs before sharing them. The checked-in fixture retains only timing samples, correctness results, policy, and round numbers; it excludes local user paths, process IDs, and raw system logs.

For an individual probe, compile and run:

xcrun clang -O2 -fobjc-arc tools/bench_event_wait.m \
  -framework Foundation -framework Metal -framework CoreGraphics \
  -framework CoreImage -o /tmp/nvmtl-event-wait-probe
/tmp/nvmtl-event-wait-probe compute NVRM_EVENT_WAIT=0
/tmp/nvmtl-event-wait-probe compute NVRM_EVENT_WAIT=1

The submitted probe preserves the measured compute/image workloads and timing boundaries, with unused experimental driver-loading paths removed and the override restricted to NVRM_EVENT_WAIT=0 or 1. Fresh measurements will vary.