33   const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()  local
180 …const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
182 …const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
184 …const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
186 …const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
198 …const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
200 …const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
202 …const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
204 …const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
216 …const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
218 …const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
220 …const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
222 …const int16x8_t vk2xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
234 …const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
236 …const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
238 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
240 …const int16x8_t vk3xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
252 …const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
254 …const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
256 …const int16x8_t vk4xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
258 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
270 …const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
272 …const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
274 …const int16x8_t vk5xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
276 …const int16x8_t vk5xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
288 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
290 …const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
292 …const int16x8_t vk6xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
294 …const int16x8_t vk6xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
306 …const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
308 …const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
310 …const int16x8_t vk7xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
312 …const int16x8_t vk7xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
324 …const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
326 …const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
328 …const int16x8_t vk8xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
330 …const int16x8_t vk8xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
342 …const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
344 …const int16x8_t vk9x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
346 …const int16x8_t vk9xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
348 …const int16x8_t vk9xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w …  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
360 …const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
362 …const int16x8_t vk10x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
364 …const int16x8_t vk10xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
366 …const int16x8_t vk10xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
378 …const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
380 …const int16x8_t vk11x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
382 …const int16x8_t vk11xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
384 …const int16x8_t vk11xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
396 …const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
398 …const int16x8_t vk12x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
400 …const int16x8_t vk12xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
402 …const int16x8_t vk12xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
414 …const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
416 …const int16x8_t vk13x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
418 …const int16x8_t vk13xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
420 …const int16x8_t vk13xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
432 …const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
434 …const int16x8_t vk14x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
436 …const int16x8_t vk14xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
438 …const int16x8_t vk14xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
450 …const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
452 …const int16x8_t vk15x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
454 …const int16x8_t vk15xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
456 …const int16x8_t vk15xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
468 …const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
470 …const int16x8_t vk16x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
472 …const int16x8_t vk16xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
474 …const int16x8_t vk16xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
486 …const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
488 …const int16x8_t vk17x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
490 …const int16x8_t vk17xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
492 …const int16x8_t vk17xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
504 …const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
506 …const int16x8_t vk18x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
508 …const int16x8_t vk18xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
510 …const int16x8_t vk18xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
522 …const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
524 …const int16x8_t vk19x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
526 …const int16x8_t vk19xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
528 …const int16x8_t vk19xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
540 …const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
542 …const int16x8_t vk20x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
544 …const int16x8_t vk20xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
546 …const int16x8_t vk20xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
558 …const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
560 …const int16x8_t vk21x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
562 …const int16x8_t vk21xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
564 …const int16x8_t vk21xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
576 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
578 …const int16x8_t vk22x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
580 …const int16x8_t vk22xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
582 …const int16x8_t vk22xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
594 …const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
596 …const int16x8_t vk23x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
598 …const int16x8_t vk23xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
600 …const int16x8_t vk23xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
612 …const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
614 …const int16x8_t vk24x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
616 …const int16x8_t vk24xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
618 …const int16x8_t vk24xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w…  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
700 …t int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
705 …k1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
710 …k2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
715 …k3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
720 …4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
725 …5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 152)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
730 …6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
735 …7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 216)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
740 …8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 248)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
745 …9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 280)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
750 …0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 312)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
755 …1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 344)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
760 …2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 376)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
765 …3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 408)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
770 …4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 440)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
775 …5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 472)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
780 …6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 504)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
785 …7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 536)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
790 …8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 568)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
795 …9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 600)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
800 …0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 632)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
805 …1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 664)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
810 …2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 696)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
815 …3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 728)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
820 …4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 760)), vkernel_zero_point));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()