xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x8c2-minmax-rndnu-neon-mull-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_ld2r(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_ld2r(
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 int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
75       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
76       const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
77       const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
78       const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
79       const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
80       const int16x4x2_t va30 = vld2_dup_s16((const void*)a3);
81       const int16x4x2_t va31 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
82 
83       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 
92       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
93       const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
94       const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
95       const int8x8_t va3c0 = vreinterpret_s8_s16(va30.val[0]);
96 
97       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
98       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
99       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
100       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
101       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
102       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
103       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
104       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
105       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
106       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
107       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
108       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
109       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
110       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
111       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
112       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
113       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
114       const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
115       const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
116       const int8x8_t va3c1 = vreinterpret_s8_s16(va30.val[1]);
117 
118       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
119       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
120       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
121       const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
122       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
123       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
124       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
125       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
126       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
127       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
128       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
129       const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
130       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
131       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
132       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
133       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
134       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
135       const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
136       const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
137       const int8x8_t va3c2 = vreinterpret_s8_s16(va31.val[0]);
138 
139       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
140       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
141       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
142       const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
143       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
144       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
145       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
146       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
147       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
148       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
149       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
150       const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
151       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
152       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
153       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
154       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
155       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
156       const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
157       const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
158       const int8x8_t va3c3 = vreinterpret_s8_s16(va31.val[1]);
159 
160       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
161       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
162       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
163       const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
164       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
165       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
166       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
167       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
168       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
169       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
170       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
171       const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
172       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
173       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
174       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
175       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
176 
177       k -= 8 * sizeof(int8_t);
178     }
179 
180     if XNN_UNLIKELY(k != 0) {
181       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
182       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
183       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
184       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
185 
186       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188 
189       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
190       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
191       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
192       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
193       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
194       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
195       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
196       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
197       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
198       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
199       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
200       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
201       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
202       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
203       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
204       const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
205       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
206       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
207       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
208       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
209 
210       if (k > 2 * sizeof(int8_t)) {
211         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213 
214         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
215         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
216         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
217         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
218         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
219         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
220         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
221         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
222         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
223         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
224         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
225         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
226         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
227         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
228         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
229         const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
230         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
231         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
232         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
233         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
234 
235         if (k > 4 * sizeof(int8_t)) {
236           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 
239           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
240           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
241           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
242           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
243           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
244           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
245           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
246           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
247           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
248           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
249           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
250           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
251           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
252           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
253           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
254           const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
255           const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
256           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
257           const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
258           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
259         }
260       }
261     }
262 
263     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
264     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
265     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
266 
267     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
268     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
269     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
270     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
271     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
272     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
273     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
274     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
275 
276     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
277     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
278     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
279     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
280     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
281     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
282     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
283     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
284 
285     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
286     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
287     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
288     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
289     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
290     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
291     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
292     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
293 
294     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
295 #if XNN_ARCH_ARM64
296     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
297     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
298     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
299     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
300 
301     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
302     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
303     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
304     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
305 
306     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
307     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
308 #else
309     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
310     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
311     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
312     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
313 
314     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
315     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
316     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
317     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
318 
319     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
320     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
321 #endif
322 
323     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
324     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
325     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
326 
327     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
328     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
329     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
330 
331     if (nc >= 8) {
332       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
333       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
334       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
335       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
336 
337       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
338       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
339       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
340       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
341 
342       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
343       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
344       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
345       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
346 
347       nc -= 8;
348     } else {
349       // Final case where not all of the 8 columns fit in the destination.
350       if (nc & 4) {
351         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
352         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
353         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
354         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
355         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
356         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
357       }
358       if (nc & 2) {
359         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
360         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
361         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
362         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
363         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
364         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
365       }
366       if (nc & 1) {
367         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
368         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
369         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
370         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
371       }
372 
373       nc = 0;
374     }
375   } while (nc != 0);
376 }
377