33         float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);  in convdw3x3s1_pack4_neon()
45         float32x4_t _k00 = vld1q_f32(k0);  in convdw3x3s1_pack4_neon()
46         float32x4_t _k01 = vld1q_f32(k0 + 4);  in convdw3x3s1_pack4_neon()
47         float32x4_t _k02 = vld1q_f32(k0 + 8);  in convdw3x3s1_pack4_neon()
48         float32x4_t _k10 = vld1q_f32(k0 + 12);  in convdw3x3s1_pack4_neon()
49         float32x4_t _k11 = vld1q_f32(k0 + 16);  in convdw3x3s1_pack4_neon()
50         float32x4_t _k12 = vld1q_f32(k0 + 20);  in convdw3x3s1_pack4_neon()
51         float32x4_t _k20 = vld1q_f32(k0 + 24);  in convdw3x3s1_pack4_neon()
52         float32x4_t _k21 = vld1q_f32(k0 + 28);  in convdw3x3s1_pack4_neon()
53         float32x4_t _k22 = vld1q_f32(k0 + 32);  in convdw3x3s1_pack4_neon()
769                 float32x4_t _r00 = vld1q_f32(r0);  in convdw3x3s1_pack4_neon()
770                 float32x4_t _r01 = vld1q_f32(r0 + 4);  in convdw3x3s1_pack4_neon()
771                 float32x4_t _r02 = vld1q_f32(r0 + 8);  in convdw3x3s1_pack4_neon()
772                 float32x4_t _r10 = vld1q_f32(r1);  in convdw3x3s1_pack4_neon()
773                 float32x4_t _r11 = vld1q_f32(r1 + 4);  in convdw3x3s1_pack4_neon()
774                 float32x4_t _r12 = vld1q_f32(r1 + 8);  in convdw3x3s1_pack4_neon()
775                 float32x4_t _r20 = vld1q_f32(r2);  in convdw3x3s1_pack4_neon()
776                 float32x4_t _r21 = vld1q_f32(r2 + 4);  in convdw3x3s1_pack4_neon()
777                 float32x4_t _r22 = vld1q_f32(r2 + 8);  in convdw3x3s1_pack4_neon()
822         float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);  in convdw3x3s2_pack4_neon()
834         float32x4_t _k00 = vld1q_f32(k0);  in convdw3x3s2_pack4_neon()
835         float32x4_t _k01 = vld1q_f32(k0 + 4);  in convdw3x3s2_pack4_neon()
836         float32x4_t _k02 = vld1q_f32(k0 + 8);  in convdw3x3s2_pack4_neon()
837         float32x4_t _k10 = vld1q_f32(k0 + 12);  in convdw3x3s2_pack4_neon()
838         float32x4_t _k11 = vld1q_f32(k0 + 16);  in convdw3x3s2_pack4_neon()
839         float32x4_t _k12 = vld1q_f32(k0 + 20);  in convdw3x3s2_pack4_neon()
840         float32x4_t _k20 = vld1q_f32(k0 + 24);  in convdw3x3s2_pack4_neon()
841         float32x4_t _k21 = vld1q_f32(k0 + 28);  in convdw3x3s2_pack4_neon()
842         float32x4_t _k22 = vld1q_f32(k0 + 32);  in convdw3x3s2_pack4_neon()
1257                 float32x4_t _r00 = vld1q_f32(r0);  in convdw3x3s2_pack4_neon()
1258                 float32x4_t _r01 = vld1q_f32(r0 + 4);  in convdw3x3s2_pack4_neon()
1259                 float32x4_t _r02 = vld1q_f32(r0 + 8);  in convdw3x3s2_pack4_neon()
1260                 float32x4_t _r10 = vld1q_f32(r1);  in convdw3x3s2_pack4_neon()
1261                 float32x4_t _r11 = vld1q_f32(r1 + 4);  in convdw3x3s2_pack4_neon()
1262                 float32x4_t _r12 = vld1q_f32(r1 + 8);  in convdw3x3s2_pack4_neon()
1263                 float32x4_t _r20 = vld1q_f32(r2);  in convdw3x3s2_pack4_neon()
1264                 float32x4_t _r21 = vld1q_f32(r2 + 4);  in convdw3x3s2_pack4_neon()
1265                 float32x4_t _r22 = vld1q_f32(r2 + 8);  in convdw3x3s2_pack4_neon()