• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h>
15 #include <xnnpack/math.h>
16 
xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 4);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   do {
61     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
63     int32x4_t vacc1x0123 = vacc0x0123;
64     int32x4_t vacc1x4567 = vacc0x4567;
65     int32x4_t vacc2x0123 = vacc0x0123;
66     int32x4_t vacc2x4567 = vacc0x4567;
67     int32x4_t vacc3x0123 = vacc0x0123;
68     int32x4_t vacc3x4567 = vacc0x4567;
69 
70     size_t k = kc;
71 
72 
73     while (k >= 8 * sizeof(int8_t)) {
74       const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
75       const int16x4x4_t va1 = vld4_dup_s16((const void*)a1); a1 += 8;
76       const int16x4x4_t va2 = vld4_dup_s16((const void*)a2); a2 += 8;
77       const int16x4x4_t va3 = vld4_dup_s16((const void*)a3); a3 += 8;
78 
79       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 
88       const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
89       const int8x8_t va1c0 = vreinterpret_s8_s16(va1.val[0]);
90       const int8x8_t va2c0 = vreinterpret_s8_s16(va2.val[0]);
91       const int8x8_t va3c0 = vreinterpret_s8_s16(va3.val[0]);
92 
93       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
94       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
95       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
96       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
97       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
98       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
99       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
100       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
101       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
102       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
103       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
104       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
105       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
106       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
107       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
108       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
109       const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
110       const int8x8_t va1c1 = vreinterpret_s8_s16(va1.val[1]);
111       const int8x8_t va2c1 = vreinterpret_s8_s16(va2.val[1]);
112       const int8x8_t va3c1 = vreinterpret_s8_s16(va3.val[1]);
113 
114       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
115       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
116       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
117       const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
118       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
119       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
120       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
121       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
122       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
123       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
124       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
125       const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
126       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
127       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
128       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
129       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
130       const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
131       const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]);
132       const int8x8_t va2c2 = vreinterpret_s8_s16(va2.val[2]);
133       const int8x8_t va3c2 = vreinterpret_s8_s16(va3.val[2]);
134 
135       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
136       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
137       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
138       const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
139       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
140       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
141       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
142       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
143       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
144       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
145       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
146       const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
147       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
148       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
149       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
150       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
151       const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
152       const int8x8_t va1c3 = vreinterpret_s8_s16(va1.val[3]);
153       const int8x8_t va2c3 = vreinterpret_s8_s16(va2.val[3]);
154       const int8x8_t va3c3 = vreinterpret_s8_s16(va3.val[3]);
155 
156       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
157       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
158       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
159       const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
160       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
161       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
162       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
163       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
164       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
165       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
166       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
167       const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
168       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
169       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
170       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
171       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
172 
173       k -= 8 * sizeof(int8_t);
174     }
175 
176     if XNN_UNLIKELY(k != 0) {
177       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
178       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
179       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
180       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
181 
182       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 
185       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
186       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
187       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
188       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
189       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
190       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
191       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
192       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
193       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
194       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
195       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
196       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
197       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
198       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
199       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
200       const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
201       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
202       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
203       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
204       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
205 
206       if (k > 2 * sizeof(int8_t)) {
207         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209 
210         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
211         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
212         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
213         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
214         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
215         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
216         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
217         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
218         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
219         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
220         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
221         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
222         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
223         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
224         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
225         const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
226         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
227         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
228         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
229         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
230 
231         if (k > 4 * sizeof(int8_t)) {
232           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234 
235           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
236           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
237           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
238           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
239           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
240           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
241           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
242           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
243           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
244           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
245           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
246           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
247           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
248           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
249           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
250           const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
251           const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
252           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
253           const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
254           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
255         }
256       }
257     }
258 
259     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
260     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
261     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
262 
263     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
264     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
265     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
266     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
267     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
268     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
269     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
270     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
271 
272     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
273     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
274     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
275     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
276     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
277     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
278     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
279     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
280 
281     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
282     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
283     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
284     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
285     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
286     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
287     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
288     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
289 
290     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
291 #if XNN_ARCH_ARM64
292     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
293     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
294     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
295     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
296 
297     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
298     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
299     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
300     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
301 
302     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
303     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
304 #else
305     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
306     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
307     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
308     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
309 
310     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
311     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
312     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
313     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
314 
315     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
316     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
317 #endif
318 
319     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
320     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
321     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
322 
323     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
324     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
325     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
326 
327     if (nc >= 8) {
328       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
329       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
330       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
331       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
332 
333       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
334       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
335       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
336       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
337 
338       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
339       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
340       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
341       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
342 
343       nc -= 8;
344     } else {
345       // Final case where not all of the 8 columns fit in the destination.
346       if (nc & 4) {
347         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
348         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
349         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
350         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
351         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
352         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
353       }
354       if (nc & 2) {
355         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
356         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
357         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
358         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
359         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
360         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
361       }
362       if (nc & 1) {
363         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
364         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
365         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
366         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
367       }
368 
369       nc = 0;
370     }
371   } while (nc != 0);
372 }
373