[nihui]

Summary

Replacement for #6758. This branch preserves futz12's original InstanceNorm subgroup optimization commits and authorship, merges current master, and appends the correctness fixes below. The original PR is left unchanged.

• follow current LayerNorm Vulkan centered, precise compensated mean accumulation for pack1 and pack4, retaining two-pass variance with precise compensated square-sum accumulation
• merge subgroup partials through shared memory and broadcast the complete mean before variance accumulation
• split channel workgroups across x/y with a bounds guard, avoiding the reduced z-dispatch limit
• add focused black-box GPU regressions with double-precision references for long constant/nonconstant inputs, small deviations after large outliers, and 65,537 channels
• skip unused fallback reduction pipelines when subgroup operations are selected
• share a pure double-precision reference between numerical and dispatch regressions
• use direct buffer transfers for the large-channel regression so Packing's independent dispatch limit does not interfere

Validation

• unchanged existing InstanceNorm tests passed before editing
• old shaders reproduced a constant-input error on SwiftShader: 0.133789 instead of zero for 65,536 copies of 10.1f, for both pack1 and pack4
• fixed InstanceNorm regressions pass with subgroup operations enabled and disabled
• independently reproduced and fixed ~0.7777% variance error for a long balanced-outlier input in both packs; the variance-specific regression uses 0.0001 tolerance
• InstanceNorm, LayerNorm, and RMSNorm CTests pass against the final combined tree (3/3)
• a separate validation-only harness exercised eight actual software Vulkan subgroups per workgroup; old pack1/pack4 statistics failed, fixed statistics matched double references
• independent review of the final fix found no blocking issues

Validation used Chromium SwiftShader (native subgroup size 4), not a physical GPU. The multi-subgroup harness deliberately adjusted its local size/stride to exercise the fallback merge; it is not included in the production changes. No new hardware performance claims or fp16 storage/arithmetic coverage.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ futz12
❌ nihui
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

❌ Patch coverage is 96.09929% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.29%. Comparing base (21eb6ca) to head (95dedf2).
⚠️ Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/vulkan/instancenorm_vulkan.cpp	95.41%	10 Missing ⚠️
...kan/shader/instancenorm_reduce_subgroup_pack4.comp	96.87%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #7136      +/-   ##
==========================================
+ Coverage   94.97%   95.29%   +0.31%     
==========================================
  Files        1464     1487      +23     
  Lines      497166   504725    +7559     
==========================================
+ Hits       472178   480964    +8786     
+ Misses      24988    23761    -1227     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-18T13:38:24.194556Z	95dedf2	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1a2eb31833

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

Binary size comparison

The binary size change of libncnn.so (bytes)

architecture	base size	pr size	difference
x86_64	27763760	27772104	+8344 ⚠️
armhf	7723152	7727388	+4236 ⚠️
aarch64	11849720	11916024	+66304 ⚠️

Commit: 01e135b73b503ffc99ea1f540bbd5791defd3625 · Workflow run