xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x8c16-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_4x8c16__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_4x8c16__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 <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (4 * 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   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr != 4) {
55     c3 = c2;
56   }
57 
58   do {
59     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x7 = 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 vacc2x0 = vacc0x0;
76     int32x4_t vacc2x1 = vacc0x1;
77     int32x4_t vacc2x2 = vacc0x2;
78     int32x4_t vacc2x3 = vacc0x3;
79     int32x4_t vacc2x4 = vacc0x4;
80     int32x4_t vacc2x5 = vacc0x5;
81     int32x4_t vacc2x6 = vacc0x6;
82     int32x4_t vacc2x7 = vacc0x7;
83     int32x4_t vacc3x0 = vacc0x0;
84     int32x4_t vacc3x1 = vacc0x1;
85     int32x4_t vacc3x2 = vacc0x2;
86     int32x4_t vacc3x3 = vacc0x3;
87     int32x4_t vacc3x4 = vacc0x4;
88     int32x4_t vacc3x5 = vacc0x5;
89     int32x4_t vacc3x6 = vacc0x6;
90     int32x4_t vacc3x7 = vacc0x7;
91 
92     size_t p = ks;
93     do {
94       const int8_t* restrict a0 = a[0];
95       if XNN_UNPREDICTABLE(a0 != zero) {
96         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97       }
98       const int8_t* restrict a1 = a[1];
99       if XNN_UNPREDICTABLE(a1 != zero) {
100         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101       }
102       const int8_t* restrict a2 = a[2];
103       if XNN_UNPREDICTABLE(a2 != zero) {
104         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105       }
106       const int8_t* restrict a3 = a[3];
107       if XNN_UNPREDICTABLE(a3 != zero) {
108         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109       }
110       a += 4;
111 
112       // KC loop of 16 with up to 15 remainder
113       size_t k = kc;
114       while (k != 0) {
115         const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
116         const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
117         const int8x16_t va2 = vld1q_s8(a2); a2 += 16;
118         const int8x16_t va3 = vld1q_s8(a3); a3 += 16;
119 
120         const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
121         const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
122         const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
123         const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
124         const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
125         const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
126         const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
127         const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
128 
129         int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
130         int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
131         int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2));
132         int16x8_t vprod3x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va3));
133         vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
134         vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
135         vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2));
136         vprod3x0 = vmlal_s8(vprod3x0, vget_high_s8(vb0), vget_high_s8(va3));
137         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
138         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
139         vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
140         vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
141         int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
142         int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
143         int16x8_t vprod2x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va2));
144         int16x8_t vprod3x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va3));
145         vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
146         vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
147         vprod2x1 = vmlal_s8(vprod2x1, vget_high_s8(vb1), vget_high_s8(va2));
148         vprod3x1 = vmlal_s8(vprod3x1, vget_high_s8(vb1), vget_high_s8(va3));
149         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
150         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
151         vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
152         vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
153         int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
154         int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
155         int16x8_t vprod2x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va2));
156         int16x8_t vprod3x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va3));
157         vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
158         vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
159         vprod2x2 = vmlal_s8(vprod2x2, vget_high_s8(vb2), vget_high_s8(va2));
160         vprod3x2 = vmlal_s8(vprod3x2, vget_high_s8(vb2), vget_high_s8(va3));
161         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
162         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
163         vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
164         vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
165         int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
166         int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
167         int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2));
168         int16x8_t vprod3x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va3));
169         vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
170         vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
171         vprod2x3 = vmlal_s8(vprod2x3, vget_high_s8(vb3), vget_high_s8(va2));
172         vprod3x3 = vmlal_s8(vprod3x3, vget_high_s8(vb3), vget_high_s8(va3));
173         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
174         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
175         vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
176         vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
177         int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
178         int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
179         int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2));
180         int16x8_t vprod3x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va3));
181         vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
182         vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
183         vprod2x4 = vmlal_s8(vprod2x4, vget_high_s8(vb4), vget_high_s8(va2));
184         vprod3x4 = vmlal_s8(vprod3x4, vget_high_s8(vb4), vget_high_s8(va3));
185         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
186         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
187         vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
188         vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
189         int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
190         int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
191         int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2));
192         int16x8_t vprod3x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va3));
193         vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
194         vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
195         vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2));
196         vprod3x5 = vmlal_s8(vprod3x5, vget_high_s8(vb5), vget_high_s8(va3));
197         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
198         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
199         vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
200         vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
201         int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
202         int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
203         int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2));
204         int16x8_t vprod3x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va3));
205         vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
206         vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
207         vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2));
208         vprod3x6 = vmlal_s8(vprod3x6, vget_high_s8(vb6), vget_high_s8(va3));
209         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
210         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
211         vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
212         vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
213         int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
214         int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
215         int16x8_t vprod2x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va2));
216         int16x8_t vprod3x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va3));
217         vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
218         vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
219         vprod2x7 = vmlal_s8(vprod2x7, vget_high_s8(vb7), vget_high_s8(va2));
220         vprod3x7 = vmlal_s8(vprod3x7, vget_high_s8(vb7), vget_high_s8(va3));
221         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
222         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
223         vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
224         vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
225 
226         k -= 16 * sizeof(int8_t);
227       }
228 
229       p -= 4 * sizeof(void*);
230     } while (p != 0);
231 
232 #if XNN_ARCH_ARM64
233     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
234     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
235     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
236     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
237     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
238     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
239     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
240     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
241     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
242     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
243     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
244     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
245     const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1);
246     const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3);
247     const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5);
248     const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7);
249     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
250     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
251     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
252     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
253     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
254     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
255     int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23);
256     int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67);
257 #else
258     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
259     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
260     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
261     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
262     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
263     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
264     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
265     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
266     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
267     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
268     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
269     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
270     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
271     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
272     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
273     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
274     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
275     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
276     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
277     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
278     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
279     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
280     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
281     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
282     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
283     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
284     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
285     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
286     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
287     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
288     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
289     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
290     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
291     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
292     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
293     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
294     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
295     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
296     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
297     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
298     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
299     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
300     const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));
301     const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));
302     const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));
303     const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));
304     const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);
305     const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);
306     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23 );
307     const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));
308     const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));
309     const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));
310     const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));
311     const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);
312     const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);
313     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67 );
314 #endif
315 
316     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
317     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
318     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
319 
320     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
321     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
322     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
323     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
324     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
325     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
326     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
327     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
328 
329     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
330     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
331     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
332     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
333     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
334     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
335     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
336     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
337 
338     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
339     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
340     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
341     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
342     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
343     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
344     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
345     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
346 
347     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
348 #if XNN_ARCH_ARM64
349     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
350     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
351     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
352     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
353     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
354     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
355 #else
356     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
357     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
358     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
359     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
360 
361     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
362     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
363 #endif
364     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
365     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
366 
367     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
368     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
369 
370     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
371     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
372 
373     if (nc >= 8) {
374       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
375       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
376       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
377       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
378 
379       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
380       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
381       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
382       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
383 
384       a = (const int8_t**restrict) ((uintptr_t) a - ks);
385 
386       nc -= 8;
387     } else {
388       if (nc & 4) {
389         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
390         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
391         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
392         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
393         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
394         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
395       }
396       if (nc & 2) {
397         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
398         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
399         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
400         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
401         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
402         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
403       }
404       if (nc & 1) {
405         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
406         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
407         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
408         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
409       }
410 
411       nc = 0;
412     }
413   } while (nc != 0);
414 }
415