129             int16x8_t posSamp = vld1q_s16(sP);  in ProcessNeonIntrinsic()  local
132             posSamp = vrev64q_s16(posSamp);  in ProcessNeonIntrinsic()
135             accum = vmlal_s16(accum, vget_low_s16(posSamp), vget_high_s16(posCoef)); // reversed  in ProcessNeonIntrinsic()
136             accum = vmlal_s16(accum, vget_high_s16(posSamp), vget_low_s16(posCoef)); // reversed  in ProcessNeonIntrinsic()
142             int16x8x2_t posSamp = vld2q_s16(sP);  in ProcessNeonIntrinsic()  local
145             posSamp.val[0] = vrev64q_s16(posSamp.val[0]);  in ProcessNeonIntrinsic()
146             posSamp.val[1] = vrev64q_s16(posSamp.val[1]);  in ProcessNeonIntrinsic()
149             accum = vmlal_s16(accum, vget_low_s16(posSamp.val[0]), vget_high_s16(posCoef)); // r  in ProcessNeonIntrinsic()
150             accum = vmlal_s16(accum, vget_high_s16(posSamp.val[0]), vget_low_s16(posCoef)); // r  in ProcessNeonIntrinsic()
151             accum2 = vmlal_s16(accum2, vget_low_s16(posSamp.val[1]), vget_high_s16(posCoef)); // r  in ProcessNeonIntrinsic()
152             accum2 = vmlal_s16(accum2, vget_high_s16(posSamp.val[1]), vget_low_s16(posCoef)); // r  in ProcessNeonIntrinsic()
268             int16x8_t posSamp = vld1q_s16(sP);  in ProcessNeonIntrinsic()  local
271             posSamp = vrev64q_s16(posSamp);  in ProcessNeonIntrinsic()
273             int32x4_t posSamp0 = vshll_n_s16(vget_low_s16(posSamp), 15);  in ProcessNeonIntrinsic()
274             int32x4_t posSamp1 = vshll_n_s16(vget_high_s16(posSamp), 15);  in ProcessNeonIntrinsic()
292             int16x8x2_t posSamp = vld2q_s16(sP);  in ProcessNeonIntrinsic()  local
295             posSamp.val[0] = vrev64q_s16(posSamp.val[0]);  in ProcessNeonIntrinsic()
296             posSamp.val[1] = vrev64q_s16(posSamp.val[1]);  in ProcessNeonIntrinsic()
299             int32x4_t posSamp0 = vshll_n_s16(vget_low_s16(posSamp.val[0]), 15);  in ProcessNeonIntrinsic()
300             int32x4_t posSamp1 = vshll_n_s16(vget_high_s16(posSamp.val[0]), 15);  in ProcessNeonIntrinsic()
316             posSamp0 = vshll_n_s16(vget_low_s16(posSamp.val[1]), 15);  in ProcessNeonIntrinsic()
317             posSamp1 = vshll_n_s16(vget_high_s16(posSamp.val[1]), 15);  in ProcessNeonIntrinsic()
438             float32x4x2_t posSamp = vld1q_f32_x2(sP);  in ProcessNeonIntrinsic()  local
443             float32x4x2_t posSamp;  in ProcessNeonIntrinsic()
444             posSamp.val[0] = vld1q_f32(sP);  in ProcessNeonIntrinsic()
446             posSamp.val[1] = vld1q_f32(sP);  in ProcessNeonIntrinsic()
455             posSamp.val[0] = vrev64q_f32(posSamp.val[0]);  in ProcessNeonIntrinsic()
456             posSamp.val[1] = vrev64q_f32(posSamp.val[1]);  in ProcessNeonIntrinsic()
457             posSamp.val[0] = vcombine_f32(  in ProcessNeonIntrinsic()
458                     vget_high_f32(posSamp.val[0]), vget_low_f32(posSamp.val[0]));  in ProcessNeonIntrinsic()
459             posSamp.val[1] = vcombine_f32(  in ProcessNeonIntrinsic()
460                     vget_high_f32(posSamp.val[1]), vget_low_f32(posSamp.val[1]));  in ProcessNeonIntrinsic()
462             accum = vmlaq_f32(accum, posSamp.val[0], posCoef.val[1]);  in ProcessNeonIntrinsic()
463             accum = vmlaq_f32(accum, posSamp.val[1], posCoef.val[0]);  in ProcessNeonIntrinsic()