xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42 
43   do {
44     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 
53     size_t k = kc;
54 
55     while (k >= 16 * sizeof(int8_t)) {
56       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
57       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
58 
59       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 
76       const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
77       const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
78 
79       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
80       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
82       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
83       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
84       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
86       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
87       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
88       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
90       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
91       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
92       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
94       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
95       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
96       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
98       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
99       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
100       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
102       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
103       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
104       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
106       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
107       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
108       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
110       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
111       const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
112       const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
113 
114       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
115       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
117       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
118       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
119       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
121       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
122       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
123       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
125       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
126       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
127       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
129       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
130       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
131       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
133       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
134       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
135       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
137       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
138       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
139       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
141       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
142       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
143       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
145       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
146 
147       k -= 16 * sizeof(int8_t);
148     }
149 
150     if (k >= 8 * sizeof(int8_t)) {
151       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
152 
153       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 
170       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
171 
172       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
173       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
174       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
175       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
176       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
177       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
178       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
179       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
180       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
181       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
182       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
183       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
184       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
185       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
186       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
187       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
188       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
189 
190       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
191       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
192       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
193       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
194       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
195       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
196       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
197       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
198       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
199       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
200       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
201       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
202       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
203       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
204       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
205       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
206 
207       k -= 8 * sizeof(int8_t);
208     }
209 
210     if XNN_UNLIKELY(k != 0) {
211       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
212 
213       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221 
222       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
223       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
224       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
225       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
226       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
227       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
228       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
229       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
230       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
231       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
232       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
233       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
234       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
235       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
236       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
237       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
238       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
239     }
240 
241 #if XNN_ARCH_ARM64
242     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
243     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
244     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
245     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
246 #else
247     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
248     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
249     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
250     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
251     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
252     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
253     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
254     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
255     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
256     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
257     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
258     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
259 #endif
260 
261     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
262     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
263     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
264 
265     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
266     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
267     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
268     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
269 
270     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
271     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
272     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
273     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
274 
275     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
276     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
277     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
278     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
279 
280     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
281 #if XNN_ARCH_ARM64
282     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
283     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
284 
285     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
286     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
287 
288     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
289 #else
290     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
291     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
292 
293     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
294     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
295 
296     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
297 #endif
298 
299     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
300     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
301 
302     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
303     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
304 
305     if (nc >= 16) {
306       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
307 
308       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
309 
310       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
311 
312       nc -= 16;
313     } else {
314       // Final case where not all of the 16 columns fit in the destination.
315       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
316       if (nc & 8) {
317         vst1_s8(c0, vout0x01234567); c0 += 8;
318         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
319       }
320       if (nc & 4) {
321         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
322         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
323       }
324       if (nc & 2) {
325         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
326         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
327       }
328       if (nc & 1) {
329         vst1_lane_s8(c0, vout0x01234567, 0);
330       }
331 
332       nc = 0;
333     }
334   } while (nc != 0);
335 }
336