[nihui]

Add tiled INT8 Winograd F(2,3) and F(4,3) for RISC-V Convolution.

Implementation:

• INT16 kernel/input transforms and widening INT32 GEMM, using eight m2 accumulators with explicit 8/4/2/1 spatial blocks and scalar 2x2/tail kernels. No fractional LMUL or microkernel templates.
• Fuse input transformation with packing, reuse input panels across output-channel tiles, and retain cache-sized split-K partial sums until the final K tile.
• Normalize once after the full reduction: /4 for F23 and /576 for F43, following the ARM/x86 arithmetic organization. No bounded normalization groups, special modular recovery, or additional input-channel safety limits.
• Use the x86 INT8 entry predicate (input channels > 8 or output channels > 8) consistently in pipeline creation and execution. Prefer F43 when enabled, otherwise F23; retain the existing geometry, options, packing and fallback behavior.
• Use an explicit RTZ conversion on XTHeadVector, preserving the caller's rounding mode. Standard RVV retains its dedicated RTZ instruction.

Related correctness fixes:

• Align both K tile selection and balancing to eight in the SSE2, MSA, LSX and ARM32 NEON Winograd helpers, matching pack8 input addressing. The ARM32 case is source-proven; it was not a reproduced ARM CI failure.
• Restore the enclosing thread ID/team size in both GNU and Clang SIMPLEOMP paths after a serialized inner region, and keep Clang's pending thread-count request separate from active team state. Cover this in the existing test_cpu regression.

Tests and validation:

• Run the unmodified pre-PR convolution tests against the new implementation before selecting additions. Preserve all 134 original cases and their data-generation sequence, ranges and tolerances. Keep only one targeted F23 case (7x3, K384, out9, unpacked with threading) in test_convolution_3a; remove the obsolete F43 grouping cases. No new test file.
• The measured standard-RVV/scalar executable-line union improves from 672/749 to 748/749 with that one case; only the nT==0 default-thread fallback is not exercised.
• Final optimized existing-suite matrix: 30/30 passes for convolution_3a, convolution_3b and convolution_oom across VLEN128/256/512/1024, runtime dispatch ON/OFF, scalar and V-disabled execution. Native x86 convolution_3a also passes.
• 6,636 bounded-data production pipeline/forward cases and 491,520 selector combinations pass, covering options, actual packing, tails, lightmode, thread changes and normal/balanced AVL. These do not claim universal overflow safety or resolve existing heterogeneous-scale postprocessing issues.
• Compile the actual runtime-ON RVV/XTHeadVector and runtime-OFF XTHeadVector convolution translation units. Check generated normalization instructions and rounding-mode restoration; standard-RVV preprocessing is unchanged by the XTHead compatibility branch. Local XTHead validation is compile/assembly-only.

Latest-head CI for 35ede47 is still running at this update. No Vulkan shader or Quantize/Dequantize changes, and no RVV hardware-performance claim.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

❌ Patch coverage is 98.01829% with 13 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.28%. Comparing base (b0f97f4) to head (1bbee07).

Files with missing lines	Patch %	Lines
src/layer/riscv/convolution_3x3_winograd_int8.h	98.24%	11 Missing ⚠️
src/layer/riscv/convolution_riscv.cpp	90.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #7093      +/-   ##
==========================================
+ Coverage   95.26%   95.28%   +0.01%     
==========================================
  Files        1476     1477       +1     
  Lines      503457   504776    +1319     
==========================================
+ Hits       479643   480999    +1356     
+ Misses      23814    23777      -37     

☔ 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.

[github-actions]

Binary size comparison

The binary size change of libncnn.so (bytes)

architecture	base size	pr size	difference
x86_64	27751048	27751048	0 😘
armhf	7710292	7710292	0 😘
aarch64	11849384	11849384	0 😘

Commit: 1bbee071a1c446673748ba19495e8dd5936d6274 · Workflow run

[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 2026-10-14T04:53:06.170092Z	1bbee07	Draft marked ready

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: 1bbee071a1

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".