108         int16x8_t posCoef = vld1q_s16(coefsP);  in ProcessNeonIntrinsic()  local
118             posCoef1 = vsubq_s16(posCoef1, posCoef);  in ProcessNeonIntrinsic()
124             posCoef = vaddq_s16(posCoef, posCoef1);  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()
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()
216         int32x4x2_t posCoef = vld1q_s32_x2(coefsP);  in ProcessNeonIntrinsic()  local
221         int32x4x2_t posCoef;  in ProcessNeonIntrinsic()
222         posCoef.val[0] = vld1q_s32(coefsP);  in ProcessNeonIntrinsic()
224         posCoef.val[1] = vld1q_s32(coefsP);  in ProcessNeonIntrinsic()
251             posCoef1.val[0] = vsubq_s32(posCoef1.val[0], posCoef.val[0]);  in ProcessNeonIntrinsic()
252             posCoef1.val[1] = vsubq_s32(posCoef1.val[1], posCoef.val[1]);  in ProcessNeonIntrinsic()
261             posCoef.val[0] = vaddq_s32(posCoef.val[0], posCoef1.val[0]);  in ProcessNeonIntrinsic()
262             posCoef.val[1] = vaddq_s32(posCoef.val[1], posCoef1.val[1]);  in ProcessNeonIntrinsic()
279             posSamp0 = vqrdmulhq_s32(posSamp0, posCoef.val[1]); // reversed  in ProcessNeonIntrinsic()
280             posSamp1 = vqrdmulhq_s32(posSamp1, posCoef.val[0]); // reversed  in ProcessNeonIntrinsic()
305             posSamp0 = vqrdmulhq_s32(posSamp0, posCoef.val[1]); // reversed  in ProcessNeonIntrinsic()
306             posSamp1 = vqrdmulhq_s32(posSamp1, posCoef.val[0]); // reversed  in ProcessNeonIntrinsic()
322             posSamp0 = vqrdmulhq_s32(posSamp0, posCoef.val[1]); // reversed  in ProcessNeonIntrinsic()
323             posSamp1 = vqrdmulhq_s32(posSamp1, posCoef.val[0]); // reversed  in ProcessNeonIntrinsic()
391         float32x4x2_t posCoef = vld1q_f32_x2(coefsP);  in ProcessNeonIntrinsic()  local
396         float32x4x2_t posCoef;  in ProcessNeonIntrinsic()
397         posCoef.val[0] = vld1q_f32(coefsP);  in ProcessNeonIntrinsic()
399         posCoef.val[1] = vld1q_f32(coefsP);  in ProcessNeonIntrinsic()
425             posCoef1.val[0] = vsubq_f32(posCoef1.val[0], posCoef.val[0]);  in ProcessNeonIntrinsic()
426             posCoef1.val[1] = vsubq_f32(posCoef1.val[1], posCoef.val[1]);  in ProcessNeonIntrinsic()
430             posCoef.val[0] = vmlaq_lane_f32(posCoef.val[0], posCoef1.val[0], interp, 0);  in ProcessNeonIntrinsic()
431             posCoef.val[1] = vmlaq_lane_f32(posCoef.val[1], posCoef1.val[1], interp, 0);  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()
499             accum = vmlaq_f32(accum, posSamp0.val[0], posCoef.val[1]); // reversed  in ProcessNeonIntrinsic()
500             accum = vmlaq_f32(accum, posSamp1.val[0], posCoef.val[0]); // reversed  in ProcessNeonIntrinsic()
501             accum2 = vmlaq_f32(accum2, posSamp0.val[1], posCoef.val[1]); // reversed  in ProcessNeonIntrinsic()
502             accum2 = vmlaq_f32(accum2, posSamp1.val[1], posCoef.val[0]); // reversed  in ProcessNeonIntrinsic()