xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/2x8c2-minmax-fp32-neonv8-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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r(
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * 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, 2 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     c1 = c0;
48   }
49 
50   do {
51     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53     int32x4_t vacc1x0123 = vacc0x0123;
54     int32x4_t vacc1x4567 = vacc0x4567;
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       const int8_t* restrict a1 = a[1];
63       if XNN_UNPREDICTABLE(a1 != zero) {
64         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65       }
66       a += 2;
67 
68       size_t k = kc;
69 
70       while (k >= 16 * sizeof(int8_t)) {
71         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
72         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
73         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
74         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
75         const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
76         const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
77         const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
78         const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
79 
80         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 
89         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
90         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
91         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
92         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
93 
94         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
95         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
96         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
98         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
99         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
100         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
101         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
102         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
103         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
105         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
106         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
107         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
108         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
109         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
110         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
111         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
112 
113         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
114         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
115         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
117         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
118         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
119         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
120         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
121         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
122         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
124         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
125         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
126         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
127         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
128         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
129         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
130         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
131 
132         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
133         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
134         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
136         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
137         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
138         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
139         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
140         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
141         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
143         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
144         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
145         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
146         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
147         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
148         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
149         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
150 
151         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
152         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
153         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
155         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
156         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
157         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
158         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
159         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
160         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
162         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
163         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
164         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
165 
166         k -= 16 * sizeof(int8_t);
167       }
168 
169       if (k >= 8 * sizeof(int8_t)) {
170         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
171         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
172         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
173         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
174 
175         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 
184         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
185         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
186 
187         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
188         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
189         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
190         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
191         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
192         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
193         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
194         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
195         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
196         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
197 
198         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
199         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
200         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
201         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
202         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
203         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
204         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
205         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
206         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
207         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
208 
209         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
210         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
211         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
212         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
213         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
214         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
215         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
216         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
217         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
218         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
219 
220         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
221         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
222         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
223         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
224         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
225         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
226         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
227         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
228 
229         k -= 8 * sizeof(int8_t);
230       }
231 
232       if XNN_UNLIKELY(k != 0) {
233         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
234         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
235 
236         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 
239         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
240         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
241         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
242         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
243         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
244         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
245         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
246         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
247         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
248         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
249 
250         if (k > 2 * sizeof(int8_t)) {
251           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 
254           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
255           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
256           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
257           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
258           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
259           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
260           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
261           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
262           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
263           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
264 
265           if (k > 4 * sizeof(int8_t)) {
266             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267             const int8x8_t vb4567c2 = 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 int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
275             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
276             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
277             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
278             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
279           }
280         }
281       }
282       p -= 2 * sizeof(void*);
283     } while (p != 0);
284 
285     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
286     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
287     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
288     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
289 
290     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
291     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
292     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
293     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
294     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
295     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
296 
297     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
298     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
299     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
300     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
301 
302     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
303 #if XNN_ARCH_ARM64
304     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
305     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
306 
307     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
308     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
309 
310     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
311 #else
312     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
313     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
314 
315     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
316     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
317 
318     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
319 #endif
320 
321     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
322     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
323 
324     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
325     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
326 
327     if (nc >= 8) {
328       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
329       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
330 
331       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
332       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
333 
334       a = (const int8_t**restrict) ((uintptr_t) a - ks);
335 
336       nc -= 8;
337     } else {
338       if (nc & 4) {
339         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
340         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
341         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
342       }
343       if (nc & 2) {
344         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
345         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
346         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
347       }
348       if (nc & 1) {
349         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
350         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
351       }
352 
353       nc = 0;
354     }
355   } while (nc != 0);
356 }
357