425   int16x4_t step1l[4], step1h[4];  in idct8x8_64_1d_bd8_kernel()  local
438   step1l[0] = vget_low_s16(io[0]);  in idct8x8_64_1d_bd8_kernel()
440   step1l[1] = vget_low_s16(io[2]);  in idct8x8_64_1d_bd8_kernel()
442   step1l[2] = vget_low_s16(io[4]);  in idct8x8_64_1d_bd8_kernel()
444   step1l[3] = vget_low_s16(io[6]);  in idct8x8_64_1d_bd8_kernel()
467   t32[2] = vmull_lane_s16(step1l[0], cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
469   t32[4] = vmull_lane_s16(step1l[1], cospis0, 3);  in idct8x8_64_1d_bd8_kernel()
471   t32[6] = vmull_lane_s16(step1l[1], cospis0, 1);  in idct8x8_64_1d_bd8_kernel()
473   t32[0] = vmlal_lane_s16(t32[2], step1l[2], cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
475   t32[2] = vmlsl_lane_s16(t32[2], step1l[2], cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
477   t32[4] = vmlsl_lane_s16(t32[4], step1l[3], cospis0, 1);  in idct8x8_64_1d_bd8_kernel()
479   t32[6] = vmlal_lane_s16(t32[6], step1l[3], cospis0, 3);  in idct8x8_64_1d_bd8_kernel()