xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x16c2-minmax-rndnu-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_ld2r(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_ld2r(
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 int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
52       const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
53       const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
54       const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
55       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 
72       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
73       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
74 
75       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
76       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
78       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
79       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
80       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
82       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
83       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
84       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
86       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
87       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
88       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
90       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
91       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
92       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
93 
94       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
95       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
97       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
98       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
99       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
101       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
102       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
103       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
105       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
106       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
107       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
109       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
110       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
111       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
112 
113       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
114       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
116       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
117       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
118       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
120       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
121       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
122       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
124       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
125       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
126       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
128       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
129       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
130       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
131 
132       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
133       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
135       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
136       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
137       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
139       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
140       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
141       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
143       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
144       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
145       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
147       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
148 
149       k -= 16 * sizeof(int8_t);
150     }
151 
152     if (k >= 8 * sizeof(int8_t)) {
153       const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
154       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
155 
156       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 
173       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
174 
175       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
176       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
177       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
178       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
179       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
180       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
181       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
182       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
183       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
184 
185       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
186       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
187       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
188       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
189       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
190       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
191       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
192       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
193       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
194 
195       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
196       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
197       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
198       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
199       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
200       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
201       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
202       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
203       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
204 
205       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
206       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
207       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
208       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
209       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
210       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
211       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
212       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
213 
214       k -= 8 * sizeof(int8_t);
215     }
216 
217     if XNN_UNLIKELY(k != 0) {
218       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
219 
220       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224 
225       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
226       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
227       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
228       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
229       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
230       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
231       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
232       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
233       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
234 
235       if (k > 2 * sizeof(int8_t)) {
236         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240 
241         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
242         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
243         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
244         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
245         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
246         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
247         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
248         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
249         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
250 
251         if (k > 4 * sizeof(int8_t)) {
252           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 
257           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
258           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
259           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
260           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
261           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
262           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
263           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
264           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
265           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
266         }
267       }
268     }
269 
270     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
271     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
272     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
273 
274     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
275     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
276     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
277     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
278 
279     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
280     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
281     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
282     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
283 
284     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
285     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
286     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
287     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
288 
289     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
290 #if XNN_ARCH_ARM64
291     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
292     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
293 
294     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
295     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
296 
297     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
298 #else
299     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
300     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
301 
302     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
303     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
304 
305     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
306 #endif
307 
308     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
309     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
310 
311     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
312     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
313 
314     if (nc >= 16) {
315       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
316 
317       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
318 
319       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
320 
321       nc -= 16;
322     } else {
323       // Final case where not all of the 16 columns fit in the destination.
324       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
325       if (nc & 8) {
326         vst1_s8(c0, vout0x01234567); c0 += 8;
327         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
328       }
329       if (nc & 4) {
330         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
331         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
332       }
333       if (nc & 2) {
334         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
335         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
336       }
337       if (nc & 1) {
338         vst1_lane_s8(c0, vout0x01234567, 0);
339       }
340 
341       nc = 0;
342     }
343   } while (nc != 0);
344 }
345