xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r(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_3x8c4__neon_mlal_ld2r(
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 <= 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, 4 * 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 
54   do {
55     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59     int32x4_t vacc1x01 = vacc0x01;
60     int32x4_t vacc1x23 = vacc0x23;
61     int32x4_t vacc1x45 = vacc0x45;
62     int32x4_t vacc1x67 = vacc0x67;
63     int32x4_t vacc2x01 = vacc0x01;
64     int32x4_t vacc2x23 = vacc0x23;
65     int32x4_t vacc2x45 = vacc0x45;
66     int32x4_t vacc2x67 = vacc0x67;
67 
68     size_t p = ks;
69     do {
70       const int8_t* restrict a0 = a[0];
71       if XNN_UNPREDICTABLE(a0 != zero) {
72         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73       }
74       const int8_t* restrict a1 = a[1];
75       if XNN_UNPREDICTABLE(a1 != zero) {
76         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77       }
78       const int8_t* restrict a2 = a[2];
79       if XNN_UNPREDICTABLE(a2 != zero) {
80         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81       }
82       a += 3;
83 
84       size_t k = kc;
85 
86       while (k >= 16 * sizeof(int8_t)) {
87         const int32x2x2_t va0x0 = vld2_dup_s32((const void*)a0); a0 += 8;
88         const int32x2x2_t va0x1 = vld2_dup_s32((const void*)a0); a0 += 8;
89         const int32x2x2_t va1x0 = vld2_dup_s32((const void*)a1); a1 += 8;
90         const int32x2x2_t va1x1 = vld2_dup_s32((const void*)a1); a1 += 8;
91         const int32x2x2_t va2x0 = vld2_dup_s32((const void*)a2); a2 += 8;
92         const int32x2x2_t va2x1 = vld2_dup_s32((const void*)a2); a2 += 8;
93 
94         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 
103         const int8x8_t va0c0x0 = vreinterpret_s8_s32(va0x0.val[0]);
104         const int8x8_t va0c0x1 = vreinterpret_s8_s32(va0x1.val[0]);
105         const int8x8_t va1c0x0 = vreinterpret_s8_s32(va1x0.val[0]);
106         const int8x8_t va1c0x1 = vreinterpret_s8_s32(va1x1.val[0]);
107         const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]);
108         const int8x8_t va2c0x1 = vreinterpret_s8_s32(va2x1.val[0]);
109 
110         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
111         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
112         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
113         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
115         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
116         vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
117         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
118         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
119         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
120         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
121         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
122         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
123         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
125         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
126         vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
127         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
128         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
129         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
130         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
131         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
132         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
133         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
135         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
136         vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
137         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
138         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
139         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
140         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
141         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
142         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
143         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
145         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
146         vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
147         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
148         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
149         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
150         const int8x8_t va0c1x0 = vreinterpret_s8_s32(va0x0.val[1]);
151         const int8x8_t va0c1x1 = vreinterpret_s8_s32(va0x1.val[1]);
152         const int8x8_t va1c1x0 = vreinterpret_s8_s32(va1x0.val[1]);
153         const int8x8_t va1c1x1 = vreinterpret_s8_s32(va1x1.val[1]);
154         const int8x8_t va2c1x0 = vreinterpret_s8_s32(va2x0.val[1]);
155         const int8x8_t va2c1x1 = vreinterpret_s8_s32(va2x1.val[1]);
156 
157         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
158         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
159         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
160         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
162         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
163         vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
164         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
165         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
166         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
167         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
168         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
169         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
170         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
172         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
173         vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
174         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
175         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
176         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
177         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
178         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
179         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
180         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
182         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
183         vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
184         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
185         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
186         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
187         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
188         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
189         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
190         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
191         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
192         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
193         vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
194         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
195         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
196         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
197 
198         k -= 16 * sizeof(int8_t);
199       }
200 
201       if (k >= 8 * sizeof(int8_t)) {
202         const int32x2x2_t va0 = vld2_dup_s32((const void*)a0); a0 += 8;
203         const int32x2x2_t va1 = vld2_dup_s32((const void*)a1); a1 += 8;
204         const int32x2x2_t va2 = vld2_dup_s32((const void*)a2); a2 += 8;
205 
206         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
211         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214 
215         const int8x8_t va0c0 = vreinterpret_s8_s32(va0.val[0]);
216         const int8x8_t va1c0 = vreinterpret_s8_s32(va1.val[0]);
217         const int8x8_t va2c0 = vreinterpret_s8_s32(va2.val[0]);
218 
219         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
220         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
221         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
222         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
223         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
224         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
225         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
226         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
227         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
228         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
229         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
230         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
231         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
232         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
233         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
234         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
235         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
236         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
237         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
238         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
239         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
240         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
241         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
242         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
243         const int8x8_t va0c1 = vreinterpret_s8_s32(va0.val[1]);
244         const int8x8_t va1c1 = vreinterpret_s8_s32(va1.val[1]);
245         const int8x8_t va2c1 = vreinterpret_s8_s32(va2.val[1]);
246 
247         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
248         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
249         const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
250         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
251         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
252         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
253         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
254         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
255         const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
256         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
257         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
258         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
259         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
260         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
261         const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
262         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
263         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
264         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
265         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
266         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
267         const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
268         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
269         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
270         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
271 
272         k -= 8 * sizeof(int8_t);
273       }
274 
275       if XNN_UNLIKELY(k != 0) {
276         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
277         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
278         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
279 
280         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
282         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
283         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
284 
285         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
286         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
287         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
288         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
289         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
290         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
291         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
292         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
293         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
294         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
295         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
296         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
297         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
298         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
299         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
300         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
301         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
302         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
303         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
304         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
305         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
306         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
307         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
308         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
309         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
310         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
311         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
312       }
313       p -= 3 * sizeof(void*);
314     } while (p != 0);
315 
316 #if XNN_ARCH_ARM64
317     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
318     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
319     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
320     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
321     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
322     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
323 #else
324     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
325     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
326     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
327     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
328     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
329     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
330     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
331     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
332     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
333     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
334     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
335     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
336     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
337     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
338     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
339     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
340     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
341     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
342 #endif
343 
344     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
345     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
346     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
347 
348     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
349     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
350     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
351     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
352     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
353     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
354 
355     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
356     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
357     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
358     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
359     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
360     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
361 
362     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
363     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
364     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
365     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
366     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
367     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
368 
369     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
370 #if XNN_ARCH_ARM64
371     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
372     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
373     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
374 
375     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
376     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
377     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
378 
379     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
380     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
381 #else
382     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
383     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
384     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
385 
386     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
387     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
388     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
389 
390     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
391     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
392 #endif
393 
394     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
395     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
396     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
397 
398     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
399     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
400     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
401 
402     if (nc >= 8) {
403       vst1_s8(c2 + 0, vout2x01234567);
404       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
405       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
406 
407       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
408       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
409       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
410 
411       a = (const int8_t**restrict) ((uintptr_t) a - ks);
412 
413       nc -= 8;
414     } else {
415       if (nc & 4) {
416         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
417         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
418         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
419         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
420         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
421       }
422       if (nc & 2) {
423         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
424         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
425         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
426         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
427         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
428       }
429       if (nc & 1) {
430         vst1_lane_s8(c2, vout2x01234567, 0);
431         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
432         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
433       }
434 
435       nc = 0;
436     }
437   } while (nc != 0);
438 }
439