319 static INLINE void idct8x8_12_pass1_bd8(const int16x4_t cospis0,  in idct8x8_12_pass1_bd8()  argument
350   t32[1] = vmull_lane_s16(step2[6], cospis0, 2);  in idct8x8_12_pass1_bd8()
351   t32[0] = vmlsl_lane_s16(t32[1], step2[5], cospis0, 2);  in idct8x8_12_pass1_bd8()
352   t32[1] = vmlal_lane_s16(t32[1], step2[5], cospis0, 2);  in idct8x8_12_pass1_bd8()
367 static INLINE void idct8x8_12_pass2_bd8(const int16x4_t cospis0,  in idct8x8_12_pass2_bd8()  argument
401   t32[2] = vmull_lane_s16(vget_low_s16(step2[6]), cospis0, 2);  in idct8x8_12_pass2_bd8()
402   t32[3] = vmull_lane_s16(vget_high_s16(step2[6]), cospis0, 2);  in idct8x8_12_pass2_bd8()
403   t32[0] = vmlsl_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2);  in idct8x8_12_pass2_bd8()
404   t32[1] = vmlsl_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2);  in idct8x8_12_pass2_bd8()
405   t32[2] = vmlal_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2);  in idct8x8_12_pass2_bd8()
406   t32[3] = vmlal_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2);  in idct8x8_12_pass2_bd8()
420 static INLINE void idct8x8_64_1d_bd8_kernel(const int16x4_t cospis0,  in idct8x8_64_1d_bd8_kernel()  argument
467   t32[2] = vmull_lane_s16(step1l[0], cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
468   t32[3] = vmull_lane_s16(step1h[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()
470   t32[5] = vmull_lane_s16(step1h[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()
472   t32[7] = vmull_lane_s16(step1h[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()
474   t32[1] = vmlal_lane_s16(t32[3], step1h[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()
476   t32[3] = vmlsl_lane_s16(t32[3], step1h[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()
478   t32[5] = vmlsl_lane_s16(t32[5], step1h[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()
480   t32[7] = vmlal_lane_s16(t32[7], step1h[3], cospis0, 3);  in idct8x8_64_1d_bd8_kernel()
495   t32[2] = vmull_lane_s16(vget_low_s16(step2[6]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
496   t32[3] = vmull_lane_s16(vget_high_s16(step2[6]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
497   t32[0] = vmlsl_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
498   t32[1] = vmlsl_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
499   t32[2] = vmlal_lane_s16(t32[2], vget_low_s16(step2[5]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
500   t32[3] = vmlal_lane_s16(t32[3], vget_high_s16(step2[5]), cospis0, 2);  in idct8x8_64_1d_bd8_kernel()
514 static INLINE void idct8x8_64_1d_bd8(const int16x4_t cospis0,  in idct8x8_64_1d_bd8()  argument
519   idct8x8_64_1d_bd8_kernel(cospis0, cospis1, io);  in idct8x8_64_1d_bd8()