1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/neon-mull-addw-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/common.h>
15 #include <xnnpack/gemm.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_3x8__neon_mull_addw_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_3x8__neon_mull_addw_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32   assert(mr != 0);
33   assert(mr <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   do {
54     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc1x0123 = vacc0x0123;
57     int32x4_t vacc1x4567 = vacc0x4567;
58     int32x4_t vacc2x0123 = vacc0x0123;
59     int32x4_t vacc2x4567 = vacc0x4567;
60 
61     size_t p = ks;
62     do {
63       const int8_t* restrict a0 = a[0];
64       if XNN_UNPREDICTABLE(a0 != zero) {
65         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66       }
67       const int8_t* restrict a1 = a[1];
68       if XNN_UNPREDICTABLE(a1 != zero) {
69         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70       }
71       const int8_t* restrict a2 = a[2];
72       if XNN_UNPREDICTABLE(a2 != zero) {
73         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
74       }
75       a += 3;
76 
77       size_t k = kc;
78       while (k >= 8 * sizeof(int8_t)) {
79         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
80         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
81         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
82 
83         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 
85         const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
86         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
87         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
88         const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
89         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
90         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
91         const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
92         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
93         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
94         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 
96         const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
97         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
98         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
99         const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
100         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
101         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
102         const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
103         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
104         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
105         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 
107         const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
108         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
109         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
110         const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
111         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
112         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
113         const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
114         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
115         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
116         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 
118         const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
119         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
120         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
121         const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
122         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
123         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
124         const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
125         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
126         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
127         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 
129         const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
130         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
131         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
132         const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
133         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
134         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
135         const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
136         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
137         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
138         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 
140         const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
141         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
142         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
143         const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
144         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
145         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
146         const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
147         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
148         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
149         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 
151         const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
152         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
153         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
154         const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
155         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
156         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
157         const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
158         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
159         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
160         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161 
162         const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
163         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
164         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
165         const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
166         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
167         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
168         const int16x8_t vprod2x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va2, 7));
169         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c7));
170         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c7));
171 
172         k -= 8 * sizeof(int8_t);
173       }
174       if XNN_UNLIKELY(k != 0) {
175         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
176         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
177         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
178 
179         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180 
181         const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
182         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
183         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
184         const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
185         vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
186         vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
187         const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
188         vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
189         vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
190 
191         if (k >= 2 * sizeof(int8_t)) {
192           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193 
194           const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
195           vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
196           vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
197           const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
198           vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
199           vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
200           const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
201           vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
202           vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
203 
204           if (k > 2 * sizeof(int8_t)) {
205             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206 
207             const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
208             vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
209             vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
210             const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
211             vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
212             vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
213             const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
214             vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
215             vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
216 
217             if (k >= 4 * sizeof(int8_t)) {
218               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219 
220               const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
221               vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
222               vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
223               const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
224               vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
225               vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
226               const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
227               vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
228               vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
229 
230               if (k > 4 * sizeof(int8_t)) {
231                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 
233                 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
234                 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
235                 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
236                 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
237                 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
238                 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
239                 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
240                 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
241                 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
242 
243                 if (k >= 6 * sizeof(int8_t)) {
244                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245 
246                   const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
247                   vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
248                   vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
249                   const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
250                   vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
251                   vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
252                   const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
253                   vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
254                   vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
255 
256                   if (k > 6 * sizeof(int8_t)) {
257                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258 
259                     const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
260                     vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
261                     vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
262                     const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
263                     vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
264                     vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
265                     const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
266                     vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
267                     vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
268                   }
269                 }
270               }
271             }
272           }
273         }
274       }
275       p -= 3 * sizeof(void*);
276     } while (p != 0);
277 
278     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
279     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
280     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
281     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
282     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
283     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
284     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
285 
286     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
287     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
288     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
289     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
290     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
291     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
292     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
293     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
294 
295     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
296     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
297     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
298     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
299     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
300     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
301 
302     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
303 #if XNN_ARCH_ARM64
304     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
305     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
306     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
307 
308     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
309     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
310 #else
311     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
312     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
313     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
314 
315     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
316     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
317 #endif
318     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
319     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
320 
321     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
322     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
323 
324     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
325     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
326 
327     if (nc >= 8) {
328       vst1_s8(c2 + 0, vout2x01234567);
329       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
330       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
331 
332       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
333       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
334       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
335 
336       a = (const int8_t**restrict) ((uintptr_t) a - ks);
337 
338       nc -= 8;
339     } else {
340       if (nc & 4) {
341         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
342         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
343         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
344         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
345         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
346       }
347       if (nc & 2) {
348         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
349         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
350         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
351         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
352         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
353       }
354       if (nc & 1) {
355         vst1_lane_s8(c2, vout2x01234567, 0);
356         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
357         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
358       }
359 
360       nc = 0;
361     }
362   } while (nc != 0);
363 }
364