xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c16-minmax-rndnu-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c16-neon-mlal.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_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal(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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal(
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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * 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, 16 * sizeof(int8_t));
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 
50   do {
51     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67     int32x4_t vacc1x0 = vacc0x0;
68     int32x4_t vacc1x1 = vacc0x1;
69     int32x4_t vacc1x2 = vacc0x2;
70     int32x4_t vacc1x3 = vacc0x3;
71     int32x4_t vacc1x4 = vacc0x4;
72     int32x4_t vacc1x5 = vacc0x5;
73     int32x4_t vacc1x6 = vacc0x6;
74     int32x4_t vacc1x7 = vacc0x7;
75     int32x4_t vacc1x8 = vacc0x8;
76     int32x4_t vacc1x9 = vacc0x9;
77     int32x4_t vacc1x10 = vacc0x10;
78     int32x4_t vacc1x11 = vacc0x11;
79     int32x4_t vacc1x12 = vacc0x12;
80     int32x4_t vacc1x13 = vacc0x13;
81     int32x4_t vacc1x14 = vacc0x14;
82     int32x4_t vacc1x15 = vacc0x15;
83 
84     size_t p = ks;
85     do {
86       const int8_t* restrict a0 = a[0];
87       if XNN_UNPREDICTABLE(a0 != zero) {
88         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89       }
90       const int8_t* restrict a1 = a[1];
91       if XNN_UNPREDICTABLE(a1 != zero) {
92         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93       }
94       a += 2;
95 
96       // KC loop of 16 with up to 15 remainder
97       size_t k = kc;
98       while (k != 0) {
99         const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
100         const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
101 
102         const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
103         const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
104         const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
105         const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
106         const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
107         const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
108         const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
109         const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
110         const int8x16_t vb8 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
111         const int8x16_t vb9 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
112         const int8x16_t vb10 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
113         const int8x16_t vb11 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
114         const int8x16_t vb12 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
115         const int8x16_t vb13 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
116         const int8x16_t vb14 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
117         const int8x16_t vb15 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
118 
119         int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
120         int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
121         vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
122         vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
123         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
124         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
125         int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
126         int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
127         vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
128         vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
129         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
130         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
131         int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
132         int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
133         vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
134         vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
135         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
136         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
137         int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
138         int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
139         vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
140         vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
141         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
142         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
143         int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
144         int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
145         vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
146         vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
147         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
148         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
149         int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
150         int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
151         vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
152         vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
153         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
154         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
155         int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
156         int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
157         vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
158         vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
159         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
160         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
161         int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
162         int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
163         vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
164         vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
165         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
166         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
167         int16x8_t vprod0x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va0));
168         int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1));
169         vprod0x8 = vmlal_s8(vprod0x8, vget_high_s8(vb8), vget_high_s8(va0));
170         vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1));
171         vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
172         vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
173         int16x8_t vprod0x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va0));
174         int16x8_t vprod1x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va1));
175         vprod0x9 = vmlal_s8(vprod0x9, vget_high_s8(vb9), vget_high_s8(va0));
176         vprod1x9 = vmlal_s8(vprod1x9, vget_high_s8(vb9), vget_high_s8(va1));
177         vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
178         vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
179         int16x8_t vprod0x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va0));
180         int16x8_t vprod1x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va1));
181         vprod0x10 = vmlal_s8(vprod0x10, vget_high_s8(vb10), vget_high_s8(va0));
182         vprod1x10 = vmlal_s8(vprod1x10, vget_high_s8(vb10), vget_high_s8(va1));
183         vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
184         vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
185         int16x8_t vprod0x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va0));
186         int16x8_t vprod1x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va1));
187         vprod0x11 = vmlal_s8(vprod0x11, vget_high_s8(vb11), vget_high_s8(va0));
188         vprod1x11 = vmlal_s8(vprod1x11, vget_high_s8(vb11), vget_high_s8(va1));
189         vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
190         vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
191         int16x8_t vprod0x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va0));
192         int16x8_t vprod1x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va1));
193         vprod0x12 = vmlal_s8(vprod0x12, vget_high_s8(vb12), vget_high_s8(va0));
194         vprod1x12 = vmlal_s8(vprod1x12, vget_high_s8(vb12), vget_high_s8(va1));
195         vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
196         vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
197         int16x8_t vprod0x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va0));
198         int16x8_t vprod1x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va1));
199         vprod0x13 = vmlal_s8(vprod0x13, vget_high_s8(vb13), vget_high_s8(va0));
200         vprod1x13 = vmlal_s8(vprod1x13, vget_high_s8(vb13), vget_high_s8(va1));
201         vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
202         vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
203         int16x8_t vprod0x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va0));
204         int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1));
205         vprod0x14 = vmlal_s8(vprod0x14, vget_high_s8(vb14), vget_high_s8(va0));
206         vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1));
207         vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
208         vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
209         int16x8_t vprod0x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va0));
210         int16x8_t vprod1x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va1));
211         vprod0x15 = vmlal_s8(vprod0x15, vget_high_s8(vb15), vget_high_s8(va0));
212         vprod1x15 = vmlal_s8(vprod1x15, vget_high_s8(vb15), vget_high_s8(va1));
213         vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
214         vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
215 
216         k -= 16 * sizeof(int8_t);
217       }
218 
219       p -= 2 * sizeof(void*);
220     } while (p != 0);
221 
222 #if XNN_ARCH_ARM64
223     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
224     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
225     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
226     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
227     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
228     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
229     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
230     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
231     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
232     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
233     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
234     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
235     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
236     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
237     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
238     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
239     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
240     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
241     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
242     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
243     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
244     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
245     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
246     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
247 #else
248     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
249     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
250     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
251     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
252     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
253     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
254     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
255     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
256     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
257     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
258     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
259     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
260     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
261     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
262     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
263     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
264     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
265     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
266     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
267     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
268     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
269     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
270     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
271     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
272     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
273     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
274     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
275     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
276     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
277     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
278     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
279     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
280     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
281     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
282     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
283     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
284     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
285     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
286     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
287     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
288     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
289     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
290     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
291     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
292     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
293     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
294     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
295     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
296     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
297     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
298     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
299     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
300     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
301     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
302     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
303     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
304 #endif
305 
306     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
307     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
308     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
309 
310     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
311     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
312     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
313     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
314     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
315     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
316     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
317     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
318 
319     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
320     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
321     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
322     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
323     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
324     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
325     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
326     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
327 
328     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
329     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
330     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
331     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
332     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
333     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
334     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
335     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
336 
337     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
338 #if XNN_ARCH_ARM64
339     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
340     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
341     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
342     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
343     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
344     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
345 #else
346     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
347     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
348     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
349     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
350 
351     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
352     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
353 #endif
354     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
355     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
356 
357     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
358     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
359 
360     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
361     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
362 
363     if (nc >= 16) {
364       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
365       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
366 
367       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
368       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
369 
370       a = (const int8_t**restrict) ((uintptr_t) a - ks);
371 
372       nc -= 16;
373     } else {
374       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
375       if (nc & 8) {
376         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
377         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
378         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
379       }
380       if (nc & 4) {
381         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
382         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
383         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
384       }
385       if (nc & 2) {
386         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
387         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
388         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
389       }
390       if (nc & 1) {
391         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
392         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
393       }
394 
395       nc = 0;
396     }
397   } while (nc != 0);
398 }
399