Skip to content

Performance claims aggregator (May 2026, historical)

This page is a frozen record of 11 performance claims copied from the descriptions of PRs shipped in May 2026. It is not current guidance.

Historical, unverified

The numbers below are claims made in PR descriptions. No benchmark artifact, hardware, or baseline file is cited for any of them, and the page is not refreshed. Entries 1 to 5 describe the Vulkan backend, which was removed (ADR-0726); the code they describe no longer exists. For current measurement workflow read perf.md.

Summary

Item Count
Total PRs 11
Measured claims (as stated in the PR) 8
Infrastructure, no direct claim 3

Vulkan submit-pool migration (backend removed)

  1. PR #561: Vulkan submit-pool infrastructure. Claim: "Foundation for pooled command buffer reuse; no direct user-visible perf delta (infrastructure)." Status: Infrastructure; measurement pending on real workloads.

  2. PR #562: Vulkan submit-pool refactor. Claim: "Consolidates queue submission; latency reduction via batch submission (measured: ~5–8% on feature extraction)." Status: Measured.

  3. PR #563: Vulkan submit-pool optimization. Claim: "Further batch reduction; memory traffic optimization (measured: ~2–3% on some kernels)." Status: Measured.

  4. PR #564: Vulkan buffer pool lifecycle. Claim: "Reduces per-frame allocation churn; no direct perf claim (memory hygiene)." Status: Infrastructure; measurement pending.

  5. PR #565: Vulkan submit-pool finalization. Claim: "Consolidates pooling across all features; cumulative effect expected (measured: validated against baseline suite)." Status: Measured.

CUDA improvements

  1. PR #569: CUDA kernel optimization. Claim: "Improved occupancy + register pressure; perf gain varies by feature (measured: 3–12% on motion/cambi kernels)." Status: Measured.

  2. PR #571: CUDA memory layout. Claim: "Coalescing improvements; latency reduction (measured: ~6% on bandwidth-bound features)." Status: Measured.

Motion v2 AVX2 fix

  1. PR #587: Motion v2 AVX2 correctness fix. Claim: "No perf claim (bug fix to bit-exactness vs. scalar)." Status: Correctness; no perf delta expected.

HIP kernels

  1. PR #612: HIP motion kernel. Claim: "Initial HIP motion implementation; equivalent to CUDA baseline (measured: parity with CUDA on AMD hardware)." Status: Measured.

  2. PR #675: HIP CAMBI kernel. Claim: "CAMBI on HIP; GPU parity (measured: validated against CUDA baseline)." Status: Measured.

  3. PR #686: HIP integer CAMBI. Claim: "Integer variant reduces register pressure; minor improvement expected (measured: ~1–2% on integer CAMBI)." Status: Measured.