PR 0058 Vulkan test a larger graphics descriptor batch

PR #58 — Vulkan: test a larger graphics descriptor batch

PR #58

Status

• GitHub state: Closed unmerged
• Disposition at review: Negative result
• Current head at review: 3a56adfac81a804e974078024cb31bd132315cb6 (research/vk-descriptor-capacity)
• Reviewed head: 3a56adfac81a804e974078024cb31bd132315cb6
• Last reviewed: 2026-09-14 UTC
• Base: main at bd1fecb93353272dda2a810991e28945de35b665
• Subsystem: Vulkan memory / synchronization

Review freshness: this page records the final/known PR head at review time. If the PR is reopened or rewritten, re-review before relying on implementation details.

At a glance

• What it changes: Vulkan: test a larger graphics descriptor batch
• Why it was attempted: Observed problem: Vulkan can exhaust its graphics descriptor-set batch and force the existing command-buffer finish/fence wait before the next draw. Profiling evidence: Diagnostic 57 recorded 665 descriptor-only capacity drains, zero uniform-only or simultaneous drains, zero unresolved other capacity producers, and 6.529970 seconds of capacity-labelled wait in a 27.007765-second setup-inclusive QPC span. Suspected cause: The 1,024-entry graphics descripto…
• Current/final result: Closed negative experiment: no repeatable benefit was established from the larger graphics descriptor batch.
• Lesson: A larger descriptor batch/capacity is not a performance win by itself; single-pair movements were insufficient and capacity changes can simply move the eventual wait.

Code impact

Implementation file	Δ	Why it matters
hw/xbox/nv2a/pgraph/vk/renderer.h	+1/-1	Renderer-owned state needed by the implementation.

Flow

flowchart LR
    subgraph Before
        A[Vulkan memory / synchronization input] --> B[Existing path] --> C[Observed limitation/cost] --> D[Guest-visible result]
    end
    subgraph PR_58[PR #58]
        E[Vulkan memory / synchronization input] --> F[PR #58 change/measurement] --> G[Qualified behavior] --> H[Guest-visible result]
    end

Loading

What changed

Change one bound in hw/xbox/nv2a/pgraph/vk/renderer.h : graphics descriptor-set array 1,024 → 2,048. - Retain ARRAY SIZE -derived graphics pool sizing and indexing. - Retain the existing capacity predicate, command-buffer finish, fence wait, reset, and teardown. - Leave the compute pool, OpenGL path, and diagnostic instrumentation unchanged/absent. Main code path: Vulkan graphics descriptor pool capacity and fence-based retirement. Commits: candidate 3a56adfac81a804e974078024cb31bd132315cb6 ; tree 463f23c13c42a2d5284152ae5879766dfa92f4d8 .

Result

Closed negative experiment: no repeatable benefit was established from the larger graphics descriptor batch.

Lesson learned

A larger descriptor batch/capacity is not a performance win by itself; single-pair movements were insufficient and capacity changes can simply move the eventual wait.

Avoid repeating

Do not increase descriptor limits and cite fewer rotations/finishes without repeated end-to-end guest-progress evidence.

Docs / evidence

Pointers only; the detailed evidence remains in its original location.

• https://github.com/Mainkill1/xemu-perf-tests/blob/393b299dc5db323c5026baa5cda4d04c76b914a0/docs/evidence/vk-capacity-20260909/REPORT.md
• https://github.com/Mainkill1/xemu-perf-tests/blob/393b299dc5db323c5026baa5cda4d04c76b914a0/docs/evidence/main-apu-load-20260909/REPORT.md
• https://github.com/Mainkill1/xemu-perf-tests/blob/bdd9effed6dc4cc30677c3d88d09b27e7b5b597c/docs/evidence/vk-descriptor-capacity-20260909/REPORT.md
• https://github.com/Mainkill1/xemu-perf-tests/blob/393b299dc5db323c5026baa5cda4d04c76b914a0/docs/performance-investigation-index.md
• https://github.com/Mainkill1/xemu-perf-tests/blob/bdd9effed6dc4cc30677c3d88d09b27e7b5b597c/docs/pr-metrics-20260909/data/xemu-58.json
• docs/evidence/vk-capacity-20260909/REPORT.md
• docs/evidence/main-apu-load-20260909/REPORT.md
• docs/evidence/vk-descriptor-capacity-20260909/REPORT.md

Related architecture

• Vulkan-Renderer
• NV2A Graphics Architecture

Related PRs referenced by this PR:

• PR #57