xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c2s4-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-shuffle.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 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull(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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
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 
54   kc = round_up_po2(kc, 8 * sizeof(int8_t));
55   do {
56     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
58     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
59     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
60     int32x4_t vacc1x0123 = vacc0x0123;
61     int32x4_t vacc1x4567 = vacc0x4567;
62     int32x4_t vacc1x89AB = vacc0x89AB;
63     int32x4_t vacc1xCDEF = vacc0xCDEF;
64     int32x4_t vacc2x0123 = vacc0x0123;
65     int32x4_t vacc2x4567 = vacc0x4567;
66     int32x4_t vacc2x89AB = vacc0x89AB;
67     int32x4_t vacc2xCDEF = vacc0xCDEF;
68 
69     size_t k = kc;
70     do {
71       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
72       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
73       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
74 
75       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 
92       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
93       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
94       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
95       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
96       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
97       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
98       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
99       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
100       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
101       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
102       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
103       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
104       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
105       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
106       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
107       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
108       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
109       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
110       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
111       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
112       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
113       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
114       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
115       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
116       va0x0 = vext_s8(va0x0, va0x0, 2);
117       va1x0 = vext_s8(va1x0, va1x0, 2);
118       va2x0 = vext_s8(va2x0, va2x0, 2);
119       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
120       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
121       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
122       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
123       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
124       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
125       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
126       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
127       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
128       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
129       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
130       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
131       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
132       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
133       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
134       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
135       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
136       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
137       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
138       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
139       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
140       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
141       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
142       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
143       va0x0 = vext_s8(va0x0, va0x0, 2);
144       va1x0 = vext_s8(va1x0, va1x0, 2);
145       va2x0 = vext_s8(va2x0, va2x0, 2);
146       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
147       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
148       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
149       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
150       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
151       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
152       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
153       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
154       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
155       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
156       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
157       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
158       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
159       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
160       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
161       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
162       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
163       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
164       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
165       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
166       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
167       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
168       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
169       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
170       va0x0 = vext_s8(va0x0, va0x0, 2);
171       va1x0 = vext_s8(va1x0, va1x0, 2);
172       va2x0 = vext_s8(va2x0, va2x0, 2);
173       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
174       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
175       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
176       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
177       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
178       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
179       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
180       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
181       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
182       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
183       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
184       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
185       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
186       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
187       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
188       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
189       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
190       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
191       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
192       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
193       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
194       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
195       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
196       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
197 
198       k -= 8 * sizeof(int8_t);
199     } while (k != 0);
200 
201     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
202     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
203     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
204 
205     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
206     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
207     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
208     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
209     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
210     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
211     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
212     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
213     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
214     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
215     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
216     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
217 
218     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
219     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
220     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
221     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
222     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
223     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
224     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
225     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
226     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
227     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
228     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
229     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
230 
231     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
232     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
233     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
234     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
235     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
236     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
237     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
238     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
239     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
240     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
241     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
242     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
243 
244     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
245 #if XNN_ARCH_ARM64
246     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
247     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
248     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
249     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
250     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
251     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
252 
253     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
254     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
255     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
256     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
257     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
258     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
259 
260     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
261     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
262     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
263 #else
264     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
265     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
266     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
267     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
268     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
269     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
270 
271     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
272     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
273     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
274     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
275     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
276     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
277 
278     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
279     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
280     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
281 #endif
282 
283     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
284     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
285     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
286     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
287 
288     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
289     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
290     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
291     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
292 
293     if (nc >= 16) {
294       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
295       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
296       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
297 
298       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
299       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
300       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
301 
302       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
303       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
304       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
305 
306       nc -= 16;
307     } else {
308       // Final case where not all of the 16 columns fit in the destination.
309       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
310       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
311       if (nc & 8) {
312         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
313         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
314         vst1_s8(c2, vout2x01234567); c2 += 8;
315         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
316         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
317       }
318       if (nc & 4) {
319         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
320         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
321         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
322         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
323         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
324       }
325       if (nc & 2) {
326         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
327         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
328         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
329         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
330         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
331       }
332       if (nc & 1) {
333         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
334         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
335         vst1_lane_s8(c2, vout2x01234567, 0);
336       }
337 
338       nc = 0;
339     }
340   } while (nc != 0);
341 }
342