xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x8c2-minmax-rndnu-neon-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-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 
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (3 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   do {
54     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc1x0123 = vacc0x0123;
57     int32x4_t vacc1x4567 = vacc0x4567;
58     int32x4_t vacc2x0123 = vacc0x0123;
59     int32x4_t vacc2x4567 = vacc0x4567;
60 
61     size_t p = ks;
62     do {
63       const int8_t* restrict a0 = a[0];
64       if XNN_UNPREDICTABLE(a0 != zero) {
65         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66       }
67       const int8_t* restrict a1 = a[1];
68       if XNN_UNPREDICTABLE(a1 != zero) {
69         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70       }
71       const int8_t* restrict a2 = a[2];
72       if XNN_UNPREDICTABLE(a2 != zero) {
73         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
74       }
75       a += 3;
76 
77       size_t k = kc;
78 
79       while (k >= 16 * sizeof(int8_t)) {
80         const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
81         const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
82         const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
83         const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
84         const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
85         const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
86         const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
87         const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
88         const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
89         const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
90         const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
91         const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
92         const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
93         const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
94         const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
95         const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
96         const int16x4_t va20x0 = vld1_dup_s16((const void*)a2);
97         const int16x4_t va21x0 = vld1_dup_s16((const void*)(a2 + 2));
98         const int16x4_t va22x0 = vld1_dup_s16((const void*)(a2 + 4));
99         const int16x4_t va23x0 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
100         const int16x4_t va20x1 = vld1_dup_s16((const void*)a2);
101         const int16x4_t va21x1 = vld1_dup_s16((const void*)(a2 + 2));
102         const int16x4_t va22x1 = vld1_dup_s16((const void*)(a2 + 4));
103         const int16x4_t va23x1 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
104 
105         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 
114         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
115         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
116         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
117         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
118         const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0);
119         const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1);
120 
121         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
122         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
123         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
124         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
126         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
127         vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
128         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
129         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
130         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
131         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
132         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
133         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
134         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
136         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
137         vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
138         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
139         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
140         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
141         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
142         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
143         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
144         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
145         const int8x8_t va2c1x0 = vreinterpret_s8_s16(va21x0);
146         const int8x8_t va2c1x1 = vreinterpret_s8_s16(va21x1);
147 
148         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
149         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
150         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
151         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
153         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
154         vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
155         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
156         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
157         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
158         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
159         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
160         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
161         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
163         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
164         vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
165         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
166         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
167         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
168         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
169         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
170         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
171         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
172         const int8x8_t va2c2x0 = vreinterpret_s8_s16(va22x0);
173         const int8x8_t va2c2x1 = vreinterpret_s8_s16(va22x1);
174 
175         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
176         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
177         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
178         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
180         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
181         vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
182         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
183         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
184         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
185         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
186         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
187         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
188         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
190         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
191         vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
192         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
193         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
194         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
195         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
196         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
197         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
198         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
199         const int8x8_t va2c3x0 = vreinterpret_s8_s16(va23x0);
200         const int8x8_t va2c3x1 = vreinterpret_s8_s16(va23x1);
201 
202         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
203         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
204         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
205         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
207         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
208         vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
209         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
210         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
211         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
212         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
213         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
214         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
215         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
217         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
218         vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
219         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
220         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
221         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
222 
223         k -= 16 * sizeof(int8_t);
224       }
225 
226       if (k >= 8 * sizeof(int8_t)) {
227         const int16x4_t va00 = vld1_dup_s16((const void*)a0);
228         const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
229         const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
230         const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
231         const int16x4_t va10 = vld1_dup_s16((const void*)a1);
232         const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
233         const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
234         const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
235         const int16x4_t va20 = vld1_dup_s16((const void*)a2);
236         const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
237         const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
238         const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
239 
240         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248 
249         const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
250         const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
251         const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
252 
253         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
254         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
255         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
256         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
257         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
258         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
259         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
260         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
261         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
262         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
263         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
264         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
265         const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
266         const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
267         const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
268 
269         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
270         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
271         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
272         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
273         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
274         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
275         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
276         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
277         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
278         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
279         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
280         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
281         const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
282         const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
283         const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
284 
285         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
286         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
287         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
288         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
289         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
290         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
291         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
292         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
293         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
294         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
295         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
296         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
297         const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
298         const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
299         const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
300 
301         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
302         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
303         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
304         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
305         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
306         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
307         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
308         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
309         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
310         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
311         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
312         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
313 
314         k -= 8 * sizeof(int8_t);
315       }
316 
317       if XNN_UNLIKELY(k != 0) {
318         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
319         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
320         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
321 
322         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324 
325         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
326         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
327         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
328         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
329         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
330         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
331         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
332         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
333         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
334         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
335         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
336         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
337         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
338         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
339         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
340 
341         if (k > 2 * sizeof(int8_t)) {
342           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
343           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
344 
345           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
346           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
347           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
348           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
349           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
350           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
351           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
352           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
353           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
354           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
355           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
356           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
357           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
358           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
359           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
360 
361           if (k > 4 * sizeof(int8_t)) {
362             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
363             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
364 
365             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
366             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
367             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
368             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
369             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
370             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
371             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
372             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
373             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
374             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
375             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
376             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
377             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
378             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
379             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
380           }
381         }
382       }
383       p -= 3 * sizeof(void*);
384     } while (p != 0);
385 
386     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
387     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
388     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
389 
390     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
391     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
392     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
393     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
394     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
395     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
396 
397     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
398     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
399     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
400     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
401     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
402     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
403 
404     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
405     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
406     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
407     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
408     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
409     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
410 
411     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
412 #if XNN_ARCH_ARM64
413     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
414     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
415     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
416 
417     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
418     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
419     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
420 
421     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
422     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
423 #else
424     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
425     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
426     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
427 
428     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
429     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
430     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
431 
432     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
433     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
434 #endif
435 
436     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
437     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
438     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
439 
440     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
441     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
442     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
443 
444     if (nc >= 8) {
445       vst1_s8(c2 + 0, vout2x01234567);
446       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
447       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
448 
449       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
450       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
451       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
452 
453       a = (const int8_t**restrict) ((uintptr_t) a - ks);
454 
455       nc -= 8;
456     } else {
457       if (nc & 4) {
458         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
459         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
460         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
461         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
462         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
463       }
464       if (nc & 2) {
465         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
466         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
467         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
468         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
469         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
470       }
471       if (nc & 1) {
472         vst1_lane_s8(c2, vout2x01234567, 0);
473         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
474         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
475       }
476 
477       nc = 0;
478     }
479   } while (nc != 0);
480 }
481