288                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
316                 _sum00 = vmlaq_f32(_sum00, _k40, _r40);  in convdw5x5s1_pack4_neon()
321                 _sum01 = vmlaq_f32(_sum01, _k40, _r41);  in convdw5x5s1_pack4_neon()
326                 _sum02 = vmlaq_f32(_sum02, _k40, _r42);  in convdw5x5s1_pack4_neon()
331                 _sum03 = vmlaq_f32(_sum03, _k40, _r43);  in convdw5x5s1_pack4_neon()
346                 _sum10 = vmlaq_f32(_sum10, _k40, _r50);  in convdw5x5s1_pack4_neon()
351                 _sum11 = vmlaq_f32(_sum11, _k40, _r51);  in convdw5x5s1_pack4_neon()
356                 _sum12 = vmlaq_f32(_sum12, _k40, _r52);  in convdw5x5s1_pack4_neon()
361                 _sum13 = vmlaq_f32(_sum13, _k40, _r53);  in convdw5x5s1_pack4_neon()
532                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
550                 _sum00 = vmlaq_f32(_sum00, _k40, _r40);  in convdw5x5s1_pack4_neon()
555                 _sum01 = vmlaq_f32(_sum01, _k40, _r41);  in convdw5x5s1_pack4_neon()
568                 _sum10 = vmlaq_f32(_sum10, _k40, _r50);  in convdw5x5s1_pack4_neon()
573                 _sum11 = vmlaq_f32(_sum11, _k40, _r51);  in convdw5x5s1_pack4_neon()
698                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
711                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s1_pack4_neon()
723                 _sum1 = vmlaq_f32(_sum1, _k40, _r50);  in convdw5x5s1_pack4_neon()
921                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
928                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s1_pack4_neon()
933                 _sum1 = vmlaq_f32(_sum1, _k40, _r41);  in convdw5x5s1_pack4_neon()
938                 _sum2 = vmlaq_f32(_sum2, _k40, _r42);  in convdw5x5s1_pack4_neon()
943                 _sum3 = vmlaq_f32(_sum3, _k40, _r43);  in convdw5x5s1_pack4_neon()
1073                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
1080                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s1_pack4_neon()
1085                 _sum1 = vmlaq_f32(_sum1, _k40, _r41);  in convdw5x5s1_pack4_neon()
1187                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s1_pack4_neon()  local
1194                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s1_pack4_neon()
1436                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s2_pack4_neon()  local
1443                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s2_pack4_neon()
1448                 _sum1 = vmlaq_f32(_sum1, _k40, _r42);  in convdw5x5s2_pack4_neon()
1453                 _sum2 = vmlaq_f32(_sum2, _k40, _r44);  in convdw5x5s2_pack4_neon()
1458                 _sum3 = vmlaq_f32(_sum3, _k40, _r46);  in convdw5x5s2_pack4_neon()
1593                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s2_pack4_neon()  local
1600                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s2_pack4_neon()
1605                 _sum1 = vmlaq_f32(_sum1, _k40, _r42);  in convdw5x5s2_pack4_neon()
1707                 float32x4_t _k40 = vld1q_f32(k0);  in convdw5x5s2_pack4_neon()  local
1714                 _sum0 = vmlaq_f32(_sum0, _k40, _r40);  in convdw5x5s2_pack4_neon()