• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_4x16c4__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_4x16c4__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 <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (4 * 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   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53   uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr != 4) {
55     c3 = c2;
56   }
57 
58   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
59 
60   do {
61     // Initialize accumulators with bias. 16 bias values are loaded from the
62     // weight matrix, at the start of the group of 16 columns.
63     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
64     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
65     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
66     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
67     uint32x4_t vpacc1x0123 = vpacc0x0123;
68     uint32x4_t vpacc1x4567 = vpacc0x4567;
69     uint32x4_t vpacc1x89AB = vpacc0x89AB;
70     uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
71     uint32x4_t vpacc2x0123 = vpacc0x0123;
72     uint32x4_t vpacc2x4567 = vpacc0x4567;
73     uint32x4_t vpacc2x89AB = vpacc0x89AB;
74     uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
75     uint32x4_t vpacc3x0123 = vpacc0x0123;
76     uint32x4_t vpacc3x4567 = vpacc0x4567;
77     uint32x4_t vpacc3x89AB = vpacc0x89AB;
78     uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
79     uint32x2_t vnacc0 = vmov_n_u32(0);
80     uint32x2_t vnacc1 = vmov_n_u32(0);
81     uint32x2_t vnacc2 = vmov_n_u32(0);
82     uint32x2_t vnacc3 = vmov_n_u32(0);
83 
84     size_t p = ks;
85     do {
86       const uint8_t* restrict a0 = a[0];
87       if XNN_UNPREDICTABLE(a0 != zero) {
88         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
89       }
90       const uint8_t* restrict a1 = a[1];
91       if XNN_UNPREDICTABLE(a1 != zero) {
92         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
93       }
94       const uint8_t* restrict a2 = a[2];
95       if XNN_UNPREDICTABLE(a2 != zero) {
96         a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
97       }
98       const uint8_t* restrict a3 = a[3];
99       if XNN_UNPREDICTABLE(a3 != zero) {
100         a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
101       }
102       a += 4;
103 
104       // Inner accumulation loop along the 16 columns.
105       size_t k = kc;
106       // 2x partial unrolled loop to load 8 bytes at a time.
107       while (k >= 8 * sizeof(uint8_t)) {
108         // Load a 4x8 block of activations.
109         const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
110         const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
111         const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
112         const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
113 
114         // Load a 8x16 block of weights.
115         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
116         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
118         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
119         const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
120         const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
121         const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
122         const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
123 
124         // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
125         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
126         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
127         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
128         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
129         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
130         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
131         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
132         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
133         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
134         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
135         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
136         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
137         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
138         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, 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         vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
144         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
145         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
146         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
147         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
148         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
149         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
150         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
151         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
152         vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
153         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
154         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
155         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
156         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
157         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
158         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
159         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
160         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
161 
162         k -= 8 * sizeof(uint8_t);
163       }
164       // Handle up to 4 final positions of `k`
165       if XNN_UNLIKELY(k != 0) {
166         // Load a 4x4 block of activations.
167         const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
168         const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
169         const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
170         const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
171 
172         // Load a 4x16 block of weights.
173         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
174         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
175         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
176         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
177 
178         // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
179         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
180         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
181         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
182         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
183         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
184         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
185         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
186         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
187         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
188         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
189         vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
190         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
191         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
192         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
193         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
194         vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
195         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
196         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
197         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
198         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
199       }
200       p -= 4 * sizeof(void*);
201     } while (p != 0);
202 
203     // Subtract zero point from accumulators.
204     vnacc0 = vpadd_u32(vnacc0, vnacc0);
205     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
206     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
207     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
208     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
209     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
210     vnacc1 = vpadd_u32(vnacc1, vnacc1);
211     const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
212     int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
213     int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
214     int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
215     int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
216     vnacc2 = vpadd_u32(vnacc2, vnacc2);
217     const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
218     int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
219     int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
220     int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
221     int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
222     vnacc3 = vpadd_u32(vnacc3, vnacc3);
223     const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
224     int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
225     int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
226     int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
227     int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
228 
229     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
230     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
231     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
232 
233     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
234     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
235     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
236     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
237     vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
238     vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
239     vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
240     vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
241     vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
242     vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
243     vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
244     vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
245     vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
246     vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
247     vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
248     vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
249 
250     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
251     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
252     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
253     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
254     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
255     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
256     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
257     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
258     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
259     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
260     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
261     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
262     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
263     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
264     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
265     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
266 
267     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
268     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
269     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
270     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
271     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
272     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
273     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
274     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
275     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
276     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
277     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
278     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
279     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
280     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
281     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
282     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
283 
284     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
285 #if XNN_ARCH_ARM64
286     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
287     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
288     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
289     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
290     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
291     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
292     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
293     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
294 
295     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
296     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
297     uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
298     uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
299 #else
300     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
301     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
302     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
303     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
304     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
305     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
306     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
307     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
308 
309     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
310     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
311     uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
312     uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
313 #endif
314     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
315     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
316 
317     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
318     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
319     vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
320     vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
321 
322     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
323     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
324     vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
325     vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
326 
327     if (nc >= 16) {
328       vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
329       vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
330       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
331       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
332 
333       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
334       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
335       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
336       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
337 
338       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
339 
340       nc -= 16;
341     } else {
342       uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
343       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
344       if (nc & 8) {
345         vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
346         vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
347         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
348         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
349         vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
350         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
351       }
352       if (nc & 4) {
353         vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
354         vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
355         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
356         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
357         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
358         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
359       }
360       if (nc & 2) {
361         vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
362         vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
363         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
364         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
365         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
366         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
367       }
368       if (nc & 1) {
369         vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
370         vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
371         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
372         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
373       }
374 
375       nc = 0;
376     }
377   } while (nc != 0);
378 }
379