RVV1.0 Dequantize Layer

[Deepdive543443]

Road to int8 optimization Episode 2: Dequantize Layer

[codecov-commenter]

Codecov Report

❌ Patch coverage is 97.29730% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.85%. Comparing base (9ee760d) to head (8cc4a44).
⚠️ Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/riscv/dequantize_riscv_zfh.cpp	96.00%	3 Missing ⚠️
src/layer/riscv/dequantize_riscv.cpp	98.63%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           master    #6658    +/-   ##
========================================
  Coverage   95.85%   95.85%            
========================================
  Files         934      936     +2     
  Lines      312692   312319   -373     
========================================
- Hits       299720   299371   -349     
+ Misses      12972    12948    -24     

☔ View full report in Codecov by Sentry.
📢 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.

[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.

✅ Deepdive543443
❌ nihui
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[nihui]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 01fa3a135c

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".

[Copilot]

Pull request overview

Adds a RISC-V (RVV 1.0) optimized implementation of the Dequantize layer as part of the int8 optimization work, including optional fp16-storage support, and adjusts the unit test suite behavior on RISC-V.

Changes:

• Added RVV-accelerated Dequantize_riscv forward path for fp32 output.
• Added NCNN_ZFH fp16-storage forward path for Dequantize_riscv.
• Disabled the pack8 dequantize test case on RISC-V builds.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
tests/test_dequantize.cpp	Skips the pack8 dequantize test on __riscv.
src/layer/riscv/dequantize_riscv.h	Declares the RISC-V optimized Dequantize layer class and fp16-storage hook.
src/layer/riscv/dequantize_riscv.cpp	Implements RVV-based dequantization to fp32 and dispatch to fp16-storage when enabled.
src/layer/riscv/dequantize_riscv_zfh.cpp	Implements fp16-storage output path (optionally using RVV).

Comments suppressed due to low confidence (2)

src/layer/riscv/dequantize_riscv.cpp:94

• Similarly, _bias is only initialized for bias_data.w == 1 or elempack == vlm1. If bias_data.w > 1 and elempack is not vlm1, _bias is used uninitialized in vfmacc, leading to undefined behavior and wrong outputs for per-channel bias with other pack widths. Add initialization + pack-width handling consistent with the _scale logic.

#if __riscv_vector
        vfloat32m8_t _bias;
        if (bias_data.w == 1)
        {
            _bias = __riscv_vfmv_v_f_f32m8(bias, __riscv_vsetvlmax_e32m8());
        }
        else if (elempack == vlm1)
        {
            vfloat32m1_t _b = __riscv_vle32_v_f32m1(bias_data, vlm1);
            _bias = __riscv_vcreate_v_f32m1_f32m8(_b, _b, _b, _b, _b, _b, _b, _b);
        }

        int n = size;
        while (n > 0)
        {
            size_t vl = __riscv_vsetvl_e32m8(n);
            vfloat32m8_t _v = __riscv_vfcvt_f_x_v_f32m8(__riscv_vle32_v_i32m8(intptr, vl), vl);
            _v = __riscv_vfmacc_vv_f32m8(_bias, _v, _scale, vl);
            __riscv_vse32_v_f32m8(ptr, _v, vl);

src/layer/riscv/dequantize_riscv_zfh.cpp:80

• In the RVV fp16-storage path, _bias is also potentially used uninitialized when bias_data.w > 1 and elempack is not vlm1. Add initialization and support for other pack widths so per-channel bias works for all packed layouts that can reach this code.

#if __riscv_vector
        vfloat32m8_t _bias;
        if (bias_data.w == 1)
        {
            _bias = __riscv_vfmv_v_f_f32m8(bias, __riscv_vsetvlmax_e32m8());
        }
        else if (elempack == vlm1)
        {
            vfloat32m1_t _b = __riscv_vle32_v_f32m1(bias_data, vlm1);
            _bias = __riscv_vcreate_v_f32m1_f32m8(_b, _b, _b, _b, _b, _b, _b, _b);
        }

        int n = size;
        while (n > 0)
        {
            size_t vl = __riscv_vsetvl_e16m4(n);
            vfloat32m8_t _v = __riscv_vfcvt_f_x_v_f32m8(__riscv_vle32_v_i32m8(intptr, vl), vl);
            _v = __riscv_vfmacc_vv_f32m8(_bias, _v, _scale, vl);
            __riscv_vse16_v_f16m4(ptr, __riscv_vfncvt_f_f_w_f16m4(_v, vl), vl);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ee2084d509

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 27a556049e

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".

[nihui]

Thanks for your contribution !