32   const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neon.kernel_zero_point);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()  local
176 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
178 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
186 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
188 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
196 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
198 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
206 …const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
208 …const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
216 …const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
218 …const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
226 …const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
228 …const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
236 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
238 …const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
246 …const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
248 …const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
256 …const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
258 …const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
266 …const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
268 …const int16x8_t vk9x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
276 …const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
278 …const int16x8_t vk10x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
286 …const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
288 …const int16x8_t vk11x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
296 …const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
298 …const int16x8_t vk12x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
306 …const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
308 …const int16x8_t vk13x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
316 …const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
318 …const int16x8_t vk14x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
326 …const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
328 …const int16x8_t vk15x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
336 …const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
338 …const int16x8_t vk16x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
346 …const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
348 …const int16x8_t vk17x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
356 …const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
358 …const int16x8_t vk18x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
366 …const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
368 …const int16x8_t vk19x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
376 …const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
378 …const int16x8_t vk20x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
386 …const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
388 …const int16x8_t vk21x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
396 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
398 …const int16x8_t vk22x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
406 …const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
408 …const int16x8_t vk23x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
416 …const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
418 …const int16x8_t vk24x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
472 …t int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
477 …vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 8)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
482 …k2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
487 …k3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
492 …k4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
497 …k5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 72)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
502 …k6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
507 …7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 104)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
512 …8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
517 …9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 136)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
522 …0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 152)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
527 …1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 168)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
532 …2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
537 …3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 200)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
542 …4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 216)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
547 …5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 232)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
552 …6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 248)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
557 …7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 264)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
562 …8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 280)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
567 …9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 296)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
572 …0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 312)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
577 …1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 328)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
582 …2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 344)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
587 …3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 360)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
592 …4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 376)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()