Feature/rvv support

[Sherlockzhangjinge]

Description

Develop the RVV accelerated version of the methods below:
MNNAvgPoolInt8
MNNFloat2Int8
MNNInt8ScaleToFloat
MNNLineDepthWiseInt8AddBiasScaleUnit
MNNMaxPoolInt8
MNNReluWithSlopeChannelInt8

Module

CPU

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

[wangzhaode]

Code Review

🔴 Blocking Issues

1. 代码风格检查不通过（CRLF 换行 / 缩进混用 / 未使用代码）

多处违反项目 clang-format 规范（.clang-format: UseTab: Never, IndentWidth: 4）：

• MNNAvgPoolInt8.cpp, MNNInt8ScaleToFloat.cpp, MNNLineDepthWiseInt8AddBiasScaleUnit.cpp, MNNMaxPoolInt8.cpp, MNNReluWithSlopeChannelInt8.cpp — 全部使用 CRLF (\r\n) 换行，与项目其余文件的 LF 换行不一致，会导致 clang-format 和 diff 工具报错
• Int8FunctionsOpt.cpp 第 2586 行 — 使用 tab 缩进而非 4 空格，与同文件其他行风格不一致
• MNNReluWithSlopeChannelInt8.cpp 第 13-18 行 — 定义了 ALIMAX/ALIMIN 宏但全文未使用，属于残留代码
• MNNFloat2Int8.cpp 第 44 行 — 行尾存在 trailing whitespace

修复建议：运行 clang-format -i -style=file 统一格式化，并将 5 个新文件的换行符转为 LF。移除未使用的宏定义。

---

2. vfcvt.x.f 截断取整 vs 参考实现 roundf() 四舍五入不一致

文件: MNNFloat2Int8.cpp:68, MNNLineDepthWiseInt8AddBiasScaleUnit.cpp:63, MNNReluWithSlopeChannelInt8.cpp:67

RVV 的 vfcvt.x.f.v 指令是向零截断取整，而参考实现（Int8FunctionsOpt.cpp:1710, 1669）使用 roundf() 做四舍五入。例如 2.7f 参考得 3，RVV 得 2；-2.7f 参考得 -3，RVV 得 -2。这会导致量化精度偏差，推理结果与 CPU 参考路径不一致。

修复建议：在 vfcvt 前加 +/-0.5f 偏移模拟四舍五入。正数路径：vfcvt(v + 0.5f)；负数路径：vfcvt(v - 0.5f)。可借助 vmfgt mask + vmerge 实现分支。

---

3. MNNFloat2Int8.cpp — vsetvlmax 返回值不确定，小 VLEN 设备可能越界

文件: MNNFloat2Int8.cpp, 第 34 行

vsetvlmax 取决于硬件 VLEN。VLEN=128 的最小配置下，e32m2 的 vlmax 仅为 2。后续用 vid % 4 构造 channel index 并通过 vloxei 从 scale[4] gather，当 vl_template < 4 时无法覆盖 4 个 channel，scale/zero 模板不完整。

修复建议：添加 vl_template < 4 的 fallback 标量路径，或改用 e32m1 类型并确保 vlmax >= 4，否则回退到参考实现。

---

Reviewed by qwen3.6-plus-preview

[wangzhaode]

Review Summary

This PR adds RVV (RISC-V Vector) accelerated implementations for 6 Int8 quantization functions: MNNAvgPoolInt8, MNNFloat2Int8, MNNInt8ScaleToFloat, MNNLineDepthWiseInt8AddBiasScaleUnit, MNNMaxPoolInt8, and MNNReluWithSlopeChannelInt8. The RVV intrinsics usage looks correct overall.

Blocking Issues

Two CI checks are currently failing and must be fixed before merge:

1. Commit Message Format: Two commits do not follow the required [Module:Type] Description format:

   • 670284be — implement some methods in Int8FunctionsOpt.cpp with RVV
   • 5f6bce67 — Fix duplicate assignment for MNNMaxPoolInt8
   • Please squash/reword to match the format, e.g. [CPU:Feature] Add RVV accelerated Int8 functions

2. Changed Lines Format (clang-format): Multiple files have formatting issues detected by CI:

   • source/backend/cpu/compute/Int8FunctionsOpt.cpp: extern declarations exceed line width and need wrapping; line 2586 uses a tab character instead of spaces for indentation
   • source/backend/cpu/riscv/rvv/MNNReluWithSlopeChannelInt8.cpp: macro definitions and function signature formatting do not conform to clang-format
   • Other new .cpp files likely have similar issues
   • Fix: run clang-format -i -style=file on all changed files

Minor Suggestions

• The Chinese comment // ---- 循环外预构造周期向量 ---- and // ---- 主循环：纯向量运算，无标量介入 ---- in MNNFloat2Int8.cpp — consider using English comments for consistency with the rest of the codebase.

@Sherlockzhangjinge Could you please fix the above formatting and commit message issues? Thanks!

Reviewed by qwen3.6-plus-preview

[wangzhaode]

Summary

This PR adds RVV (RISC-V Vector Extension) accelerated implementations for 6 Int8 quantization operators: MNNAvgPoolInt8, MNNFloat2Int8, MNNInt8ScaleToFloat, MNNLineDepthWiseInt8AddBiasScaleUnit, MNNMaxPoolInt8, and MNNReluWithSlopeChannelInt8. It also applies clang-format to Int8FunctionsOpt.cpp.

Blocking issues

1. MNNInt8ScaleToFloat_RVV — scale/zero vector length mismatch bug

• File: source/backend/cpu/riscv/rvv/MNNInt8ScaleToFloat.cpp, lines 15–37
• v_scale and v_zero are initialized with vl=4 (only 4 elements written), but the main loop uses a much larger vl (e.g., VLMAX_e32m4 = 16 when VLEN=128). Elements beyond index 3 are undefined, producing incorrect results.
• Suggested fix: Use the same vloxei32 + index-modulo pattern as MNNFloat2Int8_RVV to replicate the 4-element scale/zero pattern across the full vector length.

2. Missing #include for struct type definitions — compile error

• MNNLineDepthWiseInt8AddBiasScaleUnit.cpp uses QuanPostTreatParameters without including its definition.
• MNNReluWithSlopeChannelInt8.cpp uses QuanPrePostParameters without including its definition.
• The existing MNNGemmInt8AddBiasScale_16x4_Unit_RVV.cpp correctly includes "../../compute/Int8FunctionsOpt.h" — the new files should do the same.

Minor suggestions

• Rounding mode: RVV vfcvt_x_f defaults to round-to-nearest-even, while the reference C code uses roundf() (round-half-away-from-zero). Consider using __riscv_vfcvt_x_f_v_i32m4_rm(v, __RISCV_FRM_RMM, vl) if exact parity is needed.
• MNNFloat2Int8.cpp is missing a trailing newline.
• The bulk of Int8FunctionsOpt.cpp changes (~500 deleted lines) are clang-format reformatting — consider splitting into a separate commit for easier review.

Reviewed by qwen3.6-plus-preview