Lines Matching full:s3
24 const int16x4_t s3, const int16x4_t s4, const int16x4_t s5, in convolve12_4_2d_v() argument
34 sum = vmlal_lane_s16(sum, s3, y_filter_0_3, 3); in convolve12_4_2d_v()
49 const int16x8_t s3, const int16x8_t s4, const int16x8_t s5, in convolve12_8_2d_v() argument
60 sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), y_filter_0_3, 3); in convolve12_8_2d_v()
73 sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), y_filter_0_3, 3); in convolve12_8_2d_v()
98 int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10; in convolve_2d_sr_vert_12tap_neon() local
99 load_s16_4x11(src_ptr, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, in convolve_2d_sr_vert_12tap_neon()
107 int32x4_t d0 = convolve12_4_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, in convolve_2d_sr_vert_12tap_neon()
109 int32x4_t d1 = convolve12_4_2d_v(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, in convolve_2d_sr_vert_12tap_neon()
111 int32x4_t d2 = convolve12_4_2d_v(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, in convolve_2d_sr_vert_12tap_neon()
114 convolve12_4_2d_v(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, in convolve_2d_sr_vert_12tap_neon()
136 s3 = s7; in convolve_2d_sr_vert_12tap_neon()
155 int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10; in convolve_2d_sr_vert_12tap_neon() local
156 load_s16_8x11(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, &s8, in convolve_2d_sr_vert_12tap_neon()
165 convolve12_8_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, in convolve_2d_sr_vert_12tap_neon()
168 convolve12_8_2d_v(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, in convolve_2d_sr_vert_12tap_neon()
171 convolve12_8_2d_v(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, in convolve_2d_sr_vert_12tap_neon()
174 convolve12_8_2d_v(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, in convolve_2d_sr_vert_12tap_neon()
182 s3 = s7; in convolve_2d_sr_vert_12tap_neon()
202 const int16x4_t s2, const int16x4_t s3, in convolve8_4_2d_v() argument
212 sum = vmlal_lane_s16(sum, s3, y_filter_lo, 3); in convolve8_4_2d_v()
222 const int16x8_t s2, const int16x8_t s3, in convolve8_8_2d_v() argument
233 sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), y_filter_lo, 3); in convolve8_8_2d_v()
242 sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), y_filter_lo, 3); in convolve8_8_2d_v()
265 int16x4_t s0, s1, s2, s3, s4, s5, s6; in convolve_2d_sr_vert_8tap_neon() local
266 load_s16_4x7(src_ptr, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); in convolve_2d_sr_vert_8tap_neon()
274 int16x4_t d0 = convolve8_4_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, y_filter); in convolve_2d_sr_vert_8tap_neon()
275 int16x4_t d1 = convolve8_4_2d_v(s1, s2, s3, s4, s5, s6, s7, s8, y_filter); in convolve_2d_sr_vert_8tap_neon()
276 int16x4_t d2 = convolve8_4_2d_v(s2, s3, s4, s5, s6, s7, s8, s9, y_filter); in convolve_2d_sr_vert_8tap_neon()
278 convolve8_4_2d_v(s3, s4, s5, s6, s7, s8, s9, s10, y_filter); in convolve_2d_sr_vert_8tap_neon()
289 s3 = s7; in convolve_2d_sr_vert_8tap_neon()
298 int16x4_t d0 = convolve8_4_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, y_filter); in convolve_2d_sr_vert_8tap_neon()
306 s2 = s3; in convolve_2d_sr_vert_8tap_neon()
307 s3 = s4; in convolve_2d_sr_vert_8tap_neon()
323 int16x8_t s0, s1, s2, s3, s4, s5, s6; in convolve_2d_sr_vert_8tap_neon() local
324 load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); in convolve_2d_sr_vert_8tap_neon()
332 uint8x8_t d0 = convolve8_8_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, in convolve_2d_sr_vert_8tap_neon()
334 uint8x8_t d1 = convolve8_8_2d_v(s1, s2, s3, s4, s5, s6, s7, s8, in convolve_2d_sr_vert_8tap_neon()
336 uint8x8_t d2 = convolve8_8_2d_v(s2, s3, s4, s5, s6, s7, s8, s9, in convolve_2d_sr_vert_8tap_neon()
338 uint8x8_t d3 = convolve8_8_2d_v(s3, s4, s5, s6, s7, s8, s9, s10, in convolve_2d_sr_vert_8tap_neon()
346 s3 = s7; in convolve_2d_sr_vert_8tap_neon()
355 uint8x8_t d0 = convolve8_8_2d_v(s0, s1, s2, s3, s4, s5, s6, s7, in convolve_2d_sr_vert_8tap_neon()
361 s2 = s3; in convolve_2d_sr_vert_8tap_neon()
362 s3 = s4; in convolve_2d_sr_vert_8tap_neon()
379 const int16x4_t s2, const int16x4_t s3, in convolve6_4_2d_v() argument
388 sum = vmlal_lane_s16(sum, s3, y_filter_hi, 0); in convolve6_4_2d_v()
396 const int16x8_t s2, const int16x8_t s3, in convolve6_8_2d_v() argument
406 sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), y_filter_hi, 0); in convolve6_8_2d_v()
413 sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), y_filter_hi, 0); in convolve6_8_2d_v()
434 int16x4_t s0, s1, s2, s3, s4; in convolve_2d_sr_vert_6tap_neon() local
435 load_s16_4x5(src_ptr, src_stride, &s0, &s1, &s2, &s3, &s4); in convolve_2d_sr_vert_6tap_neon()
443 int16x4_t d0 = convolve6_4_2d_v(s0, s1, s2, s3, s4, s5, y_filter); in convolve_2d_sr_vert_6tap_neon()
444 int16x4_t d1 = convolve6_4_2d_v(s1, s2, s3, s4, s5, s6, y_filter); in convolve_2d_sr_vert_6tap_neon()
445 int16x4_t d2 = convolve6_4_2d_v(s2, s3, s4, s5, s6, s7, y_filter); in convolve_2d_sr_vert_6tap_neon()
446 int16x4_t d3 = convolve6_4_2d_v(s3, s4, s5, s6, s7, s8, y_filter); in convolve_2d_sr_vert_6tap_neon()
457 s3 = s7; in convolve_2d_sr_vert_6tap_neon()
464 int16x4_t d0 = convolve6_4_2d_v(s0, s1, s2, s3, s4, s5, y_filter); in convolve_2d_sr_vert_6tap_neon()
472 s2 = s3; in convolve_2d_sr_vert_6tap_neon()
473 s3 = s4; in convolve_2d_sr_vert_6tap_neon()
487 int16x8_t s0, s1, s2, s3, s4; in convolve_2d_sr_vert_6tap_neon() local
488 load_s16_8x5(s, src_stride, &s0, &s1, &s2, &s3, &s4); in convolve_2d_sr_vert_6tap_neon()
497 convolve6_8_2d_v(s0, s1, s2, s3, s4, s5, y_filter, sub_const); in convolve_2d_sr_vert_6tap_neon()
499 convolve6_8_2d_v(s1, s2, s3, s4, s5, s6, y_filter, sub_const); in convolve_2d_sr_vert_6tap_neon()
501 convolve6_8_2d_v(s2, s3, s4, s5, s6, s7, y_filter, sub_const); in convolve_2d_sr_vert_6tap_neon()
503 convolve6_8_2d_v(s3, s4, s5, s6, s7, s8, y_filter, sub_const); in convolve_2d_sr_vert_6tap_neon()
510 s3 = s7; in convolve_2d_sr_vert_6tap_neon()
518 convolve6_8_2d_v(s0, s1, s2, s3, s4, s5, y_filter, sub_const); in convolve_2d_sr_vert_6tap_neon()
523 s2 = s3; in convolve_2d_sr_vert_6tap_neon()
524 s3 = s4; in convolve_2d_sr_vert_6tap_neon()