xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.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_1x16c4__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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r(
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 <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * 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 
46   do {
47     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55 
56     size_t p = ks;
57     do {
58       const int8_t* restrict a0 = a[0];
59       if XNN_UNPREDICTABLE(a0 != zero) {
60         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61       }
62       a += 1;
63 
64       size_t k = kc;
65 
66       while (k >= 16 * sizeof(int8_t)) {
67         const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
68         const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
69         const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
70         const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
71 
72         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 
89         const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
90         const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
91 
92         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
93         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
95         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
96         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
97         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
99         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
100         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
101         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
103         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
104         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
105         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
107         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
108         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
109         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
111         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
112         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
113         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
115         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
116         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
117         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
119         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
120         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
121         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
123         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
124         const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
125         const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
126 
127         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
128         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
130         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
131         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
132         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
134         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
135         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
136         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
138         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
139         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
140         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
142         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
143         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
144         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
146         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
147         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
148         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
150         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
151         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
152         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
154         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
155         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
156         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
158         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
159 
160         k -= 16 * sizeof(int8_t);
161       }
162 
163       if (k >= 8 * sizeof(int8_t)) {
164         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
165         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
166 
167         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 
184         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
185 
186         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
187         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
188         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
189         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
190         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
191         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
192         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
193         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
194         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
195         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
196         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
197         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
198         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
199         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
200         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
201         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
202         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
203 
204         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
205         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
206         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
207         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
208         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
209         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
210         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
211         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
212         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
213         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
214         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
215         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
216         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
217         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
218         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
219         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
220 
221         k -= 8 * sizeof(int8_t);
222       }
223 
224       if XNN_UNLIKELY(k != 0) {
225         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
226 
227         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235 
236         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
237         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
238         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
239         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
240         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
241         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
242         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
243         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
244         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
245         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
246         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
247         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
248         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
249         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
250         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
251         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
252         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
253       }
254       p -= 1 * sizeof(void*);
255     } while (p != 0);
256 
257 #if XNN_ARCH_ARM64
258     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
259     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
260     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
261     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
262 #else
263     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
264     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
265     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
266     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
267     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
268     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
269     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
270     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
271     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
272     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
273     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
274     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
275 #endif
276 
277     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
278     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
279     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
280 
281     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
282     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
283     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
284     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
285 
286     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
287     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
288     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
289     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
290 
291     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
292     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
293     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
294     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
295 
296     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
297 #if XNN_ARCH_ARM64
298     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
299     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
300 
301     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
302     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
303 
304     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
305 #else
306     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
307     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
308 
309     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
310     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
311 
312     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
313 #endif
314 
315     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
316     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
317 
318     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
319     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
320 
321     if (nc >= 16) {
322       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
323 
324       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
325 
326       a = (const int8_t**restrict) ((uintptr_t) a - ks);
327 
328       nc -= 16;
329     } else {
330       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
331       if (nc & 8) {
332         vst1_s8(c0, vout0x01234567); c0 += 8;
333         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
334       }
335       if (nc & 4) {
336         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
337         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
338       }
339       if (nc & 2) {
340         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
341         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
342       }
343       if (nc & 1) {
344         vst1_lane_s8(c0, vout0x01234567, 0);
345       }
346 
347       nc = 0;
348     }
349   } while (nc != 0);
350 }
351