The binary size change of libncnn.so (bytes)

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.17%. Comparing base (80da741) to head (287f9ce).
Report is 3 commits behind head on master.

@@             Coverage Diff             @@
##           master    #6005       +/-   ##
===========================================
+ Coverage   95.40%   96.17%    +0.77%     
===========================================
  Files         821      589      -232     
  Lines      268820   142476   -126344     
===========================================
- Hits       256465   137026   -119439     
+ Misses      12355     5450     -6905     

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

After trying to update the compiler and researching, I found these restrictions under the xtheadvector extension

__riscv_vfsgnjn_vv_f32m8_mu will generate illegal instruction, to be reported

__riscv_vfsgnjn_vv_f32m8_mu and __riscv_vfadd_vv_f32m8_mu u policy in xtheadvector is cleared to 0 instead of maintaining src values, see XUANTIE-RV/xuantie-gnu-toolchain#26

• workaround 1, neg(abs(x))

#if __riscv_xtheadvector
            vfloat32m8_t _comm = __riscv_vfsgnjx_vv_f32m8(_p, _p, vl);
            _comm = __riscv_vfsgnjn_vv_f32m8(_comm, _comm, vl);
#else
            vfloat32m8_t _comm = __riscv_vfsgnjn_vv_f32m8_mu(_mask, _p, _p, _p, vl);
#endif

• workaround 2, use vmerge

#if __riscv_xtheadvector
            vfloat32m8_t _res = __riscv_vfadd_vv_f32m8(_comm, _p, vl);
            _res = __riscv_vmerge_vvm_f32m8(_comm, _res, _mask, vl);
#else
            vfloat32m8_t _res = __riscv_vfadd_vv_f32m8_mu(_mask, _comm, _comm, _p, vl);
#endif

Thanks for your contribution !