• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_ld4r(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_ld4r(
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 int16x4x4_t va0x0 = vld4_dup_s16((const void*)a0); a0 += 8;
72         const int16x4x4_t va0x1 = vld4_dup_s16((const void*)a0); a0 += 8;
73         const int16x4x4_t va1x0 = vld4_dup_s16((const void*)a1); a1 += 8;
74         const int16x4x4_t va1x1 = vld4_dup_s16((const void*)a1); a1 += 8;
75 
76         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 
85         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va0x0.val[0]);
86         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va0x1.val[0]);
87         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va1x0.val[0]);
88         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va1x1.val[0]);
89 
90         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
91         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
92         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
94         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
95         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
96         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
97         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
98         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
99         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
101         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
102         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
103         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
104         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va0x0.val[1]);
105         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va0x1.val[1]);
106         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va1x0.val[1]);
107         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va1x1.val[1]);
108 
109         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
110         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
111         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
113         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
114         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
115         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
116         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
117         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
118         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
120         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
121         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
122         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
123         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va0x0.val[2]);
124         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va0x1.val[2]);
125         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va1x0.val[2]);
126         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va1x1.val[2]);
127 
128         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
129         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
130         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
132         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
133         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
134         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
135         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
136         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
137         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
139         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
140         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
141         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
142         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va0x0.val[3]);
143         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]);
144         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va1x0.val[3]);
145         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]);
146 
147         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
148         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
149         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
151         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
152         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
153         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
154         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
155         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
156         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
158         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
159         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
160         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
161 
162         k -= 16 * sizeof(int8_t);
163       }
164 
165       if (k >= 8 * sizeof(int8_t)) {
166         const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
167         const int16x4x4_t va1 = vld4_dup_s16((const void*)a1); a1 += 8;
168 
169         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 
178         const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
179         const int8x8_t va1c0 = vreinterpret_s8_s16(va1.val[0]);
180 
181         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
182         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
183         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
184         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
185         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
186         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
187         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
188         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
189         const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
190         const int8x8_t va1c1 = vreinterpret_s8_s16(va1.val[1]);
191 
192         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
193         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
194         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
195         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
196         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
197         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
198         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
199         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
200         const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
201         const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]);
202 
203         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
204         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
205         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
206         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
207         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
208         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
209         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
210         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
211         const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
212         const int8x8_t va1c3 = vreinterpret_s8_s16(va1.val[3]);
213 
214         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
215         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
216         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
217         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
218         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
219         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
220         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
221         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
222 
223         k -= 8 * sizeof(int8_t);
224       }
225 
226       if XNN_UNLIKELY(k != 0) {
227         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
228         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
229 
230         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 
233         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
234         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
235         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
236         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
237         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
238         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
239         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
240         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
241         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
242         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
243 
244         if (k > 2 * sizeof(int8_t)) {
245           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247 
248           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
249           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
250           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
251           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
252           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
253           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
254           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
255           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
256           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
257           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
258 
259           if (k > 4 * sizeof(int8_t)) {
260             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
261             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
262 
263             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
264             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
265             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
266             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
267             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
268             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
269             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
270             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
271             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
272             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
273           }
274         }
275       }
276       p -= 2 * sizeof(void*);
277     } while (p != 0);
278 
279     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
280     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
281     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
282     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
283 
284     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
285     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
286     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
287     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
288     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
289     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
290 
291     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
292     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
293     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
294     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
295 
296     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
297 #if XNN_ARCH_ARM64
298     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
299     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
300 
301     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
302     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
303 
304     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
305 #else
306     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
307     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
308 
309     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
310     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
311 
312     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
313 #endif
314 
315     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
316     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
317 
318     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
319     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
320 
321     if (nc >= 8) {
322       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
323       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
324 
325       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
326       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
327 
328       a = (const int8_t**restrict) ((uintptr_t) a - ks);
329 
330       nc -= 8;
331     } else {
332       if (nc & 4) {
333         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
334         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
335         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
336       }
337       if (nc & 2) {
338         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
339         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
340         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
341       }
342       if (nc & 1) {
343         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
344         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
345       }
346 
347       nc = 0;
348     }
349   } while (nc != 0);
350 }
351