xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16c2-minmax-rndnu-neon-mlal-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_1x16c2__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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_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 <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (1 * 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 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 
51     size_t p = ks;
52     do {
53       const int8_t* restrict a0 = a[0];
54       if XNN_UNPREDICTABLE(a0 != zero) {
55         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56       }
57       a += 1;
58 
59       size_t k = kc;
60 
61       while (k >= 16 * sizeof(int8_t)) {
62         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
63         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
64         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
65         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
66 
67         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78         const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 
84         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
85         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
86 
87         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
88         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
90         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
91         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
92         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
94         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
95         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
96         const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
98         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
99         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
100         const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
102         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
103         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
104         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
105 
106         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
107         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
109         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
110         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
111         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
113         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
114         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
115         const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
117         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
118         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
119         const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
121         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
122         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
123         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
124 
125         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
126         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
128         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
129         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
130         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
132         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
133         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
134         const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
136         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
137         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
138         const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139         vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
140         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
141         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
142         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
143 
144         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
145         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
147         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
148         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
149         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
151         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
152         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
153         const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154         vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
155         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
156         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
157         const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158         vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
159         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
160 
161         k -= 16 * sizeof(int8_t);
162       }
163 
164       if (k >= 8 * sizeof(int8_t)) {
165         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
166         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
167 
168         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 
185         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
186 
187         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
188         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
189         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
190         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
191         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
192         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
193         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
194         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
195         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
196 
197         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
198         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
199         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
200         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
201         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
202         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
203         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
204         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
205         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
206 
207         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
208         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
209         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
210         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
211         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
212         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
213         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
214         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
215         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
216 
217         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
218         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
219         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
220         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
221         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
222         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
223         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
224         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
225 
226         k -= 8 * sizeof(int8_t);
227       }
228 
229       if XNN_UNLIKELY(k != 0) {
230         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + 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 
247         if (k > 2 * sizeof(int8_t)) {
248           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252 
253           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
254           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
255           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
256           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
257           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
258           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
259           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
260           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
261           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
262 
263           if (k > 4 * sizeof(int8_t)) {
264             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
265             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
266             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268 
269             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
270             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
271             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
272             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
273             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
274             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
275             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
276             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
277             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
278           }
279         }
280       }
281       p -= 1 * sizeof(void*);
282     } while (p != 0);
283 
284     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
285     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
286     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
287 
288     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
289     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
290     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
291     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
292 
293     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
294     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
295     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
296     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
297 
298     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
299     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
300     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
301     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
302 
303     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
304 #if XNN_ARCH_ARM64
305     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
306     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
307 
308     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
309     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
310 
311     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
312 #else
313     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
314     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
315 
316     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
317     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
318 
319     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
320 #endif
321 
322     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
323     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
324 
325     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
326     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
327 
328     if (nc >= 16) {
329       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
330 
331       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
332 
333       a = (const int8_t**restrict) ((uintptr_t) a - ks);
334 
335       nc -= 16;
336     } else {
337       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
338       if (nc & 8) {
339         vst1_s8(c0, vout0x01234567); c0 += 8;
340         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
341       }
342       if (nc & 4) {
343         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
344         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
345       }
346       if (nc & 2) {
347         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
348         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
349       }
350       if (nc & 1) {
351         vst1_lane_s8(c0, vout0x01234567, 0);
352       }
353 
354       nc = 0;
355     }
356   } while (nc != 0);
357 }
358