Lines Matching refs:_sum
439 float32x4_t _sum = vmulq_f32(_r00, _k012x); in convdw3x3s1_neon() local
440 _sum = vmlaq_f32(_sum, _r10, _k345x); in convdw3x3s1_neon()
441 _sum = vmlaq_f32(_sum, _r20, _k678x); in convdw3x3s1_neon()
447 _sum = vsetq_lane_f32(bias0, _sum, 3); in convdw3x3s1_neon()
450 *outptr = vaddvq_f32(_sum); in convdw3x3s1_neon()
453 float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum)); in convdw3x3s1_neon()
787 float32x4_t _sum = vmulq_f32(_r00, _k012x); in convdw3x3s1_neon() local
788 _sum = vmlaq_f32(_sum, _r10, _k345x); in convdw3x3s1_neon()
789 _sum = vmlaq_f32(_sum, _r20, _k678x); in convdw3x3s1_neon()
791 _sum = vsetq_lane_f32(bias0, _sum, 3); in convdw3x3s1_neon()
793 *outptr = vaddvq_f32(_sum); in convdw3x3s1_neon()
795 float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum)); in convdw3x3s1_neon()
1037 float32x4_t _sum = vmulq_f32(_r00, _k012x); in convdw3x3s2_neon() local
1038 _sum = vmlaq_f32(_sum, _r10, _k345x); in convdw3x3s2_neon()
1039 _sum = vmlaq_f32(_sum, _r20, _k678x); in convdw3x3s2_neon()
1041 _sum = vsetq_lane_f32(bias0, _sum, 3); in convdw3x3s2_neon()
1043 *outptr = vaddvq_f32(_sum); in convdw3x3s2_neon()
1045 float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum)); in convdw3x3s2_neon()