1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c8-neon-mull-padal.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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal(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_3x8c8__neon_mlal_padal(
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   kc = round_up_po2(kc, 8);
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53 
54   do {
55     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc1x0 = vacc0x0;
64     int32x4_t vacc1x1 = vacc0x1;
65     int32x4_t vacc1x2 = vacc0x2;
66     int32x4_t vacc1x3 = vacc0x3;
67     int32x4_t vacc1x4 = vacc0x4;
68     int32x4_t vacc1x5 = vacc0x5;
69     int32x4_t vacc1x6 = vacc0x6;
70     int32x4_t vacc1x7 = vacc0x7;
71     int32x4_t vacc2x0 = vacc0x0;
72     int32x4_t vacc2x1 = vacc0x1;
73     int32x4_t vacc2x2 = vacc0x2;
74     int32x4_t vacc2x3 = vacc0x3;
75     int32x4_t vacc2x4 = vacc0x4;
76     int32x4_t vacc2x5 = vacc0x5;
77     int32x4_t vacc2x6 = vacc0x6;
78     int32x4_t vacc2x7 = vacc0x7;
79 
80     size_t p = ks;
81     do {
82       const int8_t* restrict a0 = a[0];
83       if XNN_UNPREDICTABLE(a0 != zero) {
84         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85       }
86       const int8_t* restrict a1 = a[1];
87       if XNN_UNPREDICTABLE(a1 != zero) {
88         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89       }
90       const int8_t* restrict a2 = a[2];
91       if XNN_UNPREDICTABLE(a2 != zero) {
92         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93       }
94       a += 3;
95 
96       size_t k = kc;
97       // 2x partial unrolled loop to load 16 bytes at a time using MLA.
98       while (k >= 16 * sizeof(int8_t)) {
99         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
100         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
101         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
102         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
103         const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
104         const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
105 
106         const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
107         const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
108         const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
109         const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
110         const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
111         const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
112         const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
113         const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
114 
115         const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
116         int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
117         int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
118         int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
119         vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
120         vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
121         vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
122         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
123         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
124         vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
125         const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
126         int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
127         int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
128         int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
129         vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
130         vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
131         vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
132         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
133         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
134         vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
135         const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
136         int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
137         int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
138         int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
139         vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
140         vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
141         vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
142         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
143         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
144         vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
145         const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
146         int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
147         int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
148         int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
149         vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
150         vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
151         vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
152         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
153         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
154         vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
155         const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
156         int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
157         int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
158         int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
159         vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
160         vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
161         vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
162         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
163         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
164         vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
165         const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
166         int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
167         int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
168         int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
169         vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
170         vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
171         vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
172         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
173         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
174         vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
175         const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
176         int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
177         int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
178         int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
179         vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
180         vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
181         vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
182         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
183         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
184         vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
185         const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
186         int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
187         int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
188         int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
189         vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
190         vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
191         vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
192         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
193         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
194         vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
195 
196         k -= 16 * sizeof(int8_t);
197       }
198 
199       // Handle 8 bytes at a time using MUL.
200       if (k > 0) {
201         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
202         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
203         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
204 
205         const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206         const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
207         const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
208         const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
209         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
210         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
211         vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
212         const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213         const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
214         const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
215         const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
216         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
217         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
218         vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
219         const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
221         const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
222         const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
223         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
224         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
225         vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
226         const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227         const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
228         const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
229         const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
230         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
231         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
232         vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
233         const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234         const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
235         const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
236         const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
237         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
238         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
239         vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
240         const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241         const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
242         const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
243         const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
244         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
245         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
246         vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
247         const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248         const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
249         const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
250         const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
251         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
252         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
253         vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
254         const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255         const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
256         const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
257         const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
258         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
259         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
260         vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
261 
262         k -= 8 * sizeof(int8_t);
263       }
264 
265       p -= 3 * sizeof(void*);
266     } while (p != 0);
267 
268 #if XNN_ARCH_ARM64
269     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
270     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
271     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
272     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
273     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
274     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
275     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
276     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
277     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
278     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
279     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
280     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
281     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
282     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
283     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
284     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
285     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
286     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
287 #else
288     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
289     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
290     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
291     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
292     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
293     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
294     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
295     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
296     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
297     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
298     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
299     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
300     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
301     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
302     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
303     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
304     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
305     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
306     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
307     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
308     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
309     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
310     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
311     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
312     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
313     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
314     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
315     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
316     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
317     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
318     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
319     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
320     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
321     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
322     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
323     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
324     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
325     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
326     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
327     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
328     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
329     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
330 #endif
331 
332     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
333     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
334     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
335     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
336     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
337     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
338     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
339 
340     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
341     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
342     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
343     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
344     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
345     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
346     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
347     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
348 
349     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
350     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
351     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
352     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
353     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
354     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
355 
356     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
357 #if XNN_ARCH_ARM64
358     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
359     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
360     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
361     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
362     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
363 #else
364     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
365     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
366     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
367 
368     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
369     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
370 #endif
371     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
372     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
373 
374     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
375     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
376 
377     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
378     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
379 
380     if (nc >= 8) {
381       vst1_s8(c2 + 0, vout2x01234567);
382       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
383       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
384 
385       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
386       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
387       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
388 
389       a = (const int8_t**restrict) ((uintptr_t) a - ks);
390 
391       nc -= 8;
392     } else {
393       if (nc & 4) {
394         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
395         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
396         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
397         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
398         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
399       }
400       if (nc & 2) {
401         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
402         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
403         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
404         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
405         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
406       }
407       if (nc & 1) {
408         vst1_lane_s8(c2, vout2x01234567, 0);
409         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
410         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
411       }
412 
413       nc = 0;
414     }
415   } while (nc != 0);
416 }
417