xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/2x32c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-igemm/c4-neondot.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const uint8_t** restrict a,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const uint8_t* zero,
30     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(uint8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44   uint8_t* c0 = c;
45   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     c1 = c0;
48   }
49 
50   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
51 
52   do {
53     // Initialize accumulators with bias. 32 bias values are loaded from the
54     // weight matrix, at the start of the group of 32 columns.
55     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
56     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
58     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
59     uint32x4_t vpacc0xGHIJ = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
60     uint32x4_t vpacc0xKLMN = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
61     uint32x4_t vpacc0xOPQR = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
62     uint32x4_t vpacc0xSTUV = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
63     uint32x4_t vpacc1x0123 = vpacc0x0123;
64     uint32x4_t vpacc1x4567 = vpacc0x4567;
65     uint32x4_t vpacc1x89AB = vpacc0x89AB;
66     uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
67     uint32x4_t vpacc1xGHIJ = vpacc0xGHIJ;
68     uint32x4_t vpacc1xKLMN = vpacc0xKLMN;
69     uint32x4_t vpacc1xOPQR = vpacc0xOPQR;
70     uint32x4_t vpacc1xSTUV = vpacc0xSTUV;
71     uint32x2_t vnacc0 = vmov_n_u32(0);
72     uint32x2_t vnacc1 = vmov_n_u32(0);
73 
74     size_t p = ks;
75     do {
76       const uint8_t* restrict a0 = a[0];
77       if XNN_UNPREDICTABLE(a0 != zero) {
78         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
79       }
80       const uint8_t* restrict a1 = a[1];
81       if XNN_UNPREDICTABLE(a1 != zero) {
82         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
83       }
84       a += 2;
85 
86       // Inner accumulation loop along the 32 columns.
87       size_t k = kc;
88       // 2x partial unrolled loop to load 8 bytes at a time.
89       while (k >= 8 * sizeof(uint8_t)) {
90         // Load a 2x8 block of activations.
91         const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
92         const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
93 
94         // Load a 8x32 block of weights.
95         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
96         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
97         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
98         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
99         const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
100         const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
101         const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
102         const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
103         const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
104         const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
105         const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
106         const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
107         const uint8x16_t vb4567xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
108         const uint8x16_t vb4567xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
109         const uint8x16_t vb4567xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
110         const uint8x16_t vb4567xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
111 
112         // Multiply-accumulate: 2x8 * 8x32 --> 2x32.
113         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
114         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
115         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
116         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
117         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
118         vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
119         vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
120         vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
121         vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
122         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
123         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
124         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
125         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
126         vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb4567xGHIJ, va0x01234567, 1);
127         vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb4567xKLMN, va0x01234567, 1);
128         vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb4567xOPQR, va0x01234567, 1);
129         vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb4567xSTUV, va0x01234567, 1);
130         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
131         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
132         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
133         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
134         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
135         vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb0123xGHIJ, va1x01234567, 0);
136         vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb0123xKLMN, va1x01234567, 0);
137         vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb0123xOPQR, va1x01234567, 0);
138         vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb0123xSTUV, va1x01234567, 0);
139         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
140         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
141         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
142         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
143         vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb4567xGHIJ, va1x01234567, 1);
144         vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb4567xKLMN, va1x01234567, 1);
145         vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb4567xOPQR, va1x01234567, 1);
146         vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb4567xSTUV, va1x01234567, 1);
147 
148         k -= 8 * sizeof(uint8_t);
149       }
150       // Handle up to 4 final positions of `k`
151       if XNN_UNLIKELY(k != 0) {
152         // Load a 2x4 block of activations.
153         const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
154         const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
155 
156         // Load a 4x32 block of weights.
157         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
158         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
159         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
160         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
161         const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
162         const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
163         const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
164         const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
165 
166         // Multiply-accumulate: 2x4 * 4x32 --> 2x32.
167         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
168         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
169         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
170         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
171         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
172         vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
173         vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
174         vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
175         vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
176         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
177         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
178         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
179         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
180         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
181         vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb0123xGHIJ, va1x01234567, 0);
182         vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb0123xKLMN, va1x01234567, 0);
183         vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb0123xOPQR, va1x01234567, 0);
184         vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb0123xSTUV, va1x01234567, 0);
185       }
186       p -= 2 * sizeof(void*);
187     } while (p != 0);
188 
189     // Subtract zero point from accumulators.
190     vnacc0 = vpadd_u32(vnacc0, vnacc0);
191     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
192     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
193     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
194     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
195     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
196     int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123));
197     int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123));
198     int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123));
199     int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123));
200     vnacc1 = vpadd_u32(vnacc1, vnacc1);
201     const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
202     int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
203     int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
204     int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
205     int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
206     int32x4_t vacc1xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc1xGHIJ, vnacc1x0123));
207     int32x4_t vacc1xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc1xKLMN, vnacc1x0123));
208     int32x4_t vacc1xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc1xOPQR, vnacc1x0123));
209     int32x4_t vacc1xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc1xSTUV, vnacc1x0123));
210 
211     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
212     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
213     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
214 
215     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
216     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
217     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
218     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
219     vacc0xGHIJ = vshlq_s32(vacc0xGHIJ, vright_pre_shift);
220     vacc0xKLMN = vshlq_s32(vacc0xKLMN, vright_pre_shift);
221     vacc0xOPQR = vshlq_s32(vacc0xOPQR, vright_pre_shift);
222     vacc0xSTUV = vshlq_s32(vacc0xSTUV, vright_pre_shift);
223     vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
224     vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
225     vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
226     vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
227     vacc1xGHIJ = vshlq_s32(vacc1xGHIJ, vright_pre_shift);
228     vacc1xKLMN = vshlq_s32(vacc1xKLMN, vright_pre_shift);
229     vacc1xOPQR = vshlq_s32(vacc1xOPQR, vright_pre_shift);
230     vacc1xSTUV = vshlq_s32(vacc1xSTUV, vright_pre_shift);
231 
232     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
233     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
234     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
235     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
236     vacc0xGHIJ = vqdmulhq_s32(vacc0xGHIJ, vmultiplier);
237     vacc0xKLMN = vqdmulhq_s32(vacc0xKLMN, vmultiplier);
238     vacc0xOPQR = vqdmulhq_s32(vacc0xOPQR, vmultiplier);
239     vacc0xSTUV = vqdmulhq_s32(vacc0xSTUV, vmultiplier);
240     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
241     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
242     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
243     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
244     vacc1xGHIJ = vqdmulhq_s32(vacc1xGHIJ, vmultiplier);
245     vacc1xKLMN = vqdmulhq_s32(vacc1xKLMN, vmultiplier);
246     vacc1xOPQR = vqdmulhq_s32(vacc1xOPQR, vmultiplier);
247     vacc1xSTUV = vqdmulhq_s32(vacc1xSTUV, vmultiplier);
248 
249     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
250     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
251     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
252     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
253     vacc0xGHIJ = vrshlq_s32(vacc0xGHIJ, vright_post_shift);
254     vacc0xKLMN = vrshlq_s32(vacc0xKLMN, vright_post_shift);
255     vacc0xOPQR = vrshlq_s32(vacc0xOPQR, vright_post_shift);
256     vacc0xSTUV = vrshlq_s32(vacc0xSTUV, vright_post_shift);
257     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
258     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
259     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
260     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
261     vacc1xGHIJ = vrshlq_s32(vacc1xGHIJ, vright_post_shift);
262     vacc1xKLMN = vrshlq_s32(vacc1xKLMN, vright_post_shift);
263     vacc1xOPQR = vrshlq_s32(vacc1xOPQR, vright_post_shift);
264     vacc1xSTUV = vrshlq_s32(vacc1xSTUV, vright_post_shift);
265 
266     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
267 #if XNN_ARCH_ARM64
268     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
269     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
270     const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xGHIJ), vacc0xKLMN), voutput_zero_point);
271     const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xOPQR), vacc0xSTUV), voutput_zero_point);
272     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
273     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
274     const int16x8_t vacc1xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1xGHIJ), vacc1xKLMN), voutput_zero_point);
275     const int16x8_t vacc1xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1xOPQR), vacc1xSTUV), voutput_zero_point);
276 
277     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
278     uint8x16_t vout0xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc0xGHIJKLMN), vacc0xOPQRSTUV);
279     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
280     uint8x16_t vout1xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc1xGHIJKLMN), vacc1xOPQRSTUV);
281 #else
282     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
283     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
284     const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xGHIJ), vqmovn_s32(vacc0xKLMN)), voutput_zero_point);
285     const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xOPQR), vqmovn_s32(vacc0xSTUV)), voutput_zero_point);
286     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
287     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
288     const int16x8_t vacc1xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1xGHIJ), vqmovn_s32(vacc1xKLMN)), voutput_zero_point);
289     const int16x8_t vacc1xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1xOPQR), vqmovn_s32(vacc1xSTUV)), voutput_zero_point);
290 
291     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
292     uint8x16_t vout0xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc0xGHIJKLMN), vqmovun_s16(vacc0xOPQRSTUV));
293     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
294     uint8x16_t vout1xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc1xGHIJKLMN), vqmovun_s16(vacc1xOPQRSTUV));
295 #endif
296     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
297     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
298 
299     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
300     vout0xGHIJKLMNOPQRSTUV = vmaxq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
301     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
302     vout1xGHIJKLMNOPQRSTUV = vmaxq_u8(vout1xGHIJKLMNOPQRSTUV, voutput_min);
303 
304     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
305     vout0xGHIJKLMNOPQRSTUV = vminq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_max);
306     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
307     vout1xGHIJKLMNOPQRSTUV = vminq_u8(vout1xGHIJKLMNOPQRSTUV, voutput_max);
308 
309     if (nc >= 32) {
310       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
311       vst1q_u8(c1 + 16, vout1xGHIJKLMNOPQRSTUV);
312       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
313       vst1q_u8(c0 + 16, vout0xGHIJKLMNOPQRSTUV);
314 
315       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
316       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
317 
318       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
319 
320       nc -= 32;
321     } else {
322       if (nc & 16) {
323         vst1q_u8(c1, vout1x0123456789ABCDEF);  c1 += 16;
324         vst1q_u8(c0, vout0x0123456789ABCDEF);  c0 += 16;
325 
326         vout1x0123456789ABCDEF = vout1xGHIJKLMNOPQRSTUV;
327         vout0x0123456789ABCDEF = vout0xGHIJKLMNOPQRSTUV;
328       }
329       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
330       if (nc & 8) {
331         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
332         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
333         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
334       }
335       if (nc & 4) {
336         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
337         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
338         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
339       }
340       if (nc & 2) {
341         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
342         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
343         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
344       }
345       if (nc & 1) {
346         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
347         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
348       }
349 
350       nc = 0;
351     }
352   } while (nc != 0);
353 }
354