• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r(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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 1);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41 
42   do {
43     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 
48     size_t k = kc;
49 
50     while (k >= 16 * sizeof(int8_t)) {
51       const int16x4x4_t va0x0 = vld4_dup_s16((const void*)a0); a0 += 8;
52       const int16x4x4_t va0x1 = vld4_dup_s16((const void*)a0); a0 += 8;
53       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
54       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 
70       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va0x0.val[0]);
71       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va0x1.val[0]);
72 
73       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
74       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
76       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
77       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
78       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
80       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
81       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
82       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
84       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
85       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
86       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
88       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
89       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va0x0.val[1]);
90       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va0x1.val[1]);
91 
92       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
93       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
95       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
96       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
97       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
99       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
100       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
101       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
103       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
104       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
105       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
107       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
108       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va0x0.val[2]);
109       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va0x1.val[2]);
110 
111       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
112       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
114       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
115       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
116       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
118       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
119       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
120       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
122       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
123       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
124       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
126       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
127       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va0x0.val[3]);
128       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]);
129 
130       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
131       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
133       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
134       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
135       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
137       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
138       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
139       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
141       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
142       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
143       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
145       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
146 
147       k -= 16 * sizeof(int8_t);
148     }
149 
150     if (k >= 8 * sizeof(int8_t)) {
151       const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
152 
153       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 
170       const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
171 
172       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
173       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
174       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
175       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
176       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
177       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
178       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
179       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
180       const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
181 
182       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
183       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
184       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
185       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
186       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
187       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
188       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
189       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
190       const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
191 
192       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
193       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
194       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
195       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
196       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
197       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
198       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
199       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
200       const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
201 
202       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
203       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
204       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
205       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
206       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
207       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
208       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
209       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
210 
211       k -= 8 * sizeof(int8_t);
212     }
213 
214     if XNN_UNLIKELY(k != 0) {
215       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
216 
217       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221 
222       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
223       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
224       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
225       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
226       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
227       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
228       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
229       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
230       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
231 
232       if (k > 2 * sizeof(int8_t)) {
233         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237 
238         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
239         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
240         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
241         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
242         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
243         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
244         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
245         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
246         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
247 
248         if (k > 4 * sizeof(int8_t)) {
249           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 
254           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
255           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
256           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
257           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
258           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
259           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
260           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
261           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
262           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
263         }
264       }
265     }
266 
267     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
268     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
269     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
270 
271     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
272     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
273     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
274     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
275 
276     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
277     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
278     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
279     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
280 
281     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
282     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
283     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
284     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
285 
286     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
287 #if XNN_ARCH_ARM64
288     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
289     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
290 
291     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
292     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
293 
294     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
295 #else
296     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
297     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
298 
299     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
300     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
301 
302     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
303 #endif
304 
305     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
306     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
307 
308     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
309     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
310 
311     if (nc >= 16) {
312       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
313 
314       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
315 
316       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
317 
318       nc -= 16;
319     } else {
320       // Final case where not all of the 16 columns fit in the destination.
321       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
322       if (nc & 8) {
323         vst1_s8(c0, vout0x01234567); c0 += 8;
324         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
325       }
326       if (nc & 4) {
327         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
328         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
329       }
330       if (nc & 2) {
331         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
332         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
333       }
334       if (nc & 1) {
335         vst1_lane_s8(c0, vout0x01234567, 0);
336       }
337 
338       nc = 0;
339     }
340   } while (nc != 0);
341 }
342