xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c2-minmax-rndnu-neon-mull-ld2r.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_3x16c2__neon_mull_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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r(
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 vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc1x0123 = vacc0x0123;
59     int32x4_t vacc1x4567 = vacc0x4567;
60     int32x4_t vacc1x89AB = vacc0x89AB;
61     int32x4_t vacc1xCDEF = vacc0xCDEF;
62     int32x4_t vacc2x0123 = vacc0x0123;
63     int32x4_t vacc2x4567 = vacc0x4567;
64     int32x4_t vacc2x89AB = vacc0x89AB;
65     int32x4_t vacc2xCDEF = vacc0xCDEF;
66 
67     size_t p = ks;
68     do {
69       const int8_t* restrict a0 = a[0];
70       if XNN_UNPREDICTABLE(a0 != zero) {
71         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72       }
73       const int8_t* restrict a1 = a[1];
74       if XNN_UNPREDICTABLE(a1 != zero) {
75         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76       }
77       const int8_t* restrict a2 = a[2];
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80       }
81       a += 3;
82 
83       size_t k = kc;
84 
85 
86       while (k >= 8 * sizeof(int8_t)) {
87         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
88         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
89         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
90         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
91         const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
92         const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
93 
94         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 
111         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
112         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
113         const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
114 
115         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
116         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
117         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
118         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
119         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
120         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
121         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
122         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
123         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
124         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
125         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
126         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
127         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
128         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
129         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
130         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
131         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
132         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
133         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
134         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
135         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
136         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
137         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
138         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
139         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
140         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
141         const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
142 
143         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
144         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
145         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
146         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
147         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
148         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
149         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
150         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
151         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
152         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
153         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
154         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
155         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
156         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
157         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
158         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
159         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
160         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
161         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
162         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
163         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
164         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
165         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
166         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
167         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
168         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
169         const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
170 
171         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
172         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
173         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
174         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
175         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
176         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
177         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
178         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
179         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
180         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
181         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
182         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
183         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
184         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
185         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
186         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
187         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
188         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
189         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
190         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
191         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
192         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
193         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
194         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
195         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
196         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
197         const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
198 
199         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
200         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
201         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
202         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
203         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
204         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
205         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
206         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
207         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
208         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
209         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
210         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
211         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
212         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
213         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
214         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
215         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
216         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
217         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
218         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
219         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
220         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
221         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
222         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
223 
224         k -= 8 * sizeof(int8_t);
225       }
226 
227       if XNN_UNLIKELY(k != 0) {
228         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
229         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
230         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
231 
232         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236 
237         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
238         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
239         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
240         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
241         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
242         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
243         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
244         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
245         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
246         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
247         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
248         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
249         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
250         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
251         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
252         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
253         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
254         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
255         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
256         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
257         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
258         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
259         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
260         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
261         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
262         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
263         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
264 
265         if (k > 2 * sizeof(int8_t)) {
266           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
269           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
270 
271           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
272           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
273           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
274           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
275           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
276           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
277           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
278           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
279           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
280           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
281           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
282           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
283           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
284           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
285           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
286           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
287           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
288           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
289           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
290           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
291           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
292           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
293           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
294           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
295           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
296           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
297           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
298 
299           if (k > 4 * sizeof(int8_t)) {
300             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
301             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
302             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
303             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
304 
305             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
306             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
307             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
308             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
309             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
310             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
311             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
312             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
313             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
314             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
315             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
316             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
317             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
318             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
319             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
320             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
321             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
322             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
323             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
324             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
325             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
326             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
327             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
328             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
329             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
330             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
331             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
332           }
333         }
334       }
335       p -= 3 * sizeof(void*);
336     } while (p != 0);
337 
338     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
339     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
340     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
341 
342     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
343     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
344     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
345     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
346     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
347     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
348     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
349     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
350     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
351     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
352     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
353     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
354 
355     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
356     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
357     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
358     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
359     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
360     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
361     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
362     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
363     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
364     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
365     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
366     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
367 
368     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
369     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
370     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
371     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
372     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
373     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
374     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
375     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
376     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
377     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
378     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
379     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
380 
381     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
382 #if XNN_ARCH_ARM64
383     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
384     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
385     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
386     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
387     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
388     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
389 
390     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
391     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
392     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
393     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
394     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
395     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
396 
397     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
398     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
399     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
400 #else
401     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
402     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
403     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
404     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
405     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
406     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
407 
408     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
409     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
410     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
411     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
412     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
413     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
414 
415     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
416     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
417     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
418 #endif
419 
420     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
421     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
422     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
423     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
424 
425     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
426     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
427     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
428     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
429 
430     if (nc >= 16) {
431       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
432       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
433       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
434 
435       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
436       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
437       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
438 
439       a = (const int8_t**restrict) ((uintptr_t) a - ks);
440 
441       nc -= 16;
442     } else {
443       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
444       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
445       if (nc & 8) {
446         vst1_s8(c2, vout2x01234567); c2 += 8;
447         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
448         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
449         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
450         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
451       }
452       if (nc & 4) {
453         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
454         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
455         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
456         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
457         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
458       }
459       if (nc & 2) {
460         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
461         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
462         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
463         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
464         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
465       }
466       if (nc & 1) {
467         vst1_lane_s8(c2, vout2x01234567, 0);
468         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
469         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
470       }
471 
472       nc = 0;
473     }
474   } while (nc != 0);
475 }
476