• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup(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)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 2);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (2 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 4 * sizeof(int8_t));
45   int8_t* c0 = c;
46   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr != 2) {
48     c1 = c0;
49   }
50 
51   do {
52     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56     int32x4_t vacc1x01 = vacc0x01;
57     int32x4_t vacc1x23 = vacc0x23;
58     int32x4_t vacc1x45 = vacc0x45;
59     int32x4_t vacc1x67 = vacc0x67;
60 
61     size_t p = ks;
62     do {
63       const int8_t* restrict a0 = a[0];
64       if XNN_UNPREDICTABLE(a0 != zero) {
65         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66       }
67       const int8_t* restrict a1 = a[1];
68       if XNN_UNPREDICTABLE(a1 != zero) {
69         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70       }
71       a += 2;
72 
73       size_t k = kc;
74 
75       while (k >= 16 * sizeof(int8_t)) {
76         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
77         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
78         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
79         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
80 
81         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 
90         const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
91         const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
92         const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
93         const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
94 
95         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
96         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
97         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
99         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
100         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
101         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
102         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
103         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
104         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
106         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
107         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
108         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
109         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
110         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
111         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
113         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
114         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
115         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
116         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
117         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
118         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
120         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
121         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
122         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
123         const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
124         const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
125         const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
126         const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
127 
128         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
129         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
130         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
132         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
133         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
134         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
135         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
136         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
137         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
139         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
140         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
141         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
142         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
143         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
144         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
146         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
147         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
148         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
149         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
150         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
151         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
153         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
154         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
155         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
156 
157         k -= 16 * sizeof(int8_t);
158       }
159 
160       if (k >= 8 * sizeof(int8_t)) {
161         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
162         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
163 
164         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 
173         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
174         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
175 
176         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
177         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
178         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
179         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
180         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
181         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
182         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
183         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
184         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
185         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
186         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
187         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
188         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
189         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
190         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
191         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
192         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
193         const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
194 
195         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
196         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
197         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
198         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
199         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
200         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
201         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
202         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
203         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
204         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
205         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
206         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
207         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
208         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
209         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
210         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
211 
212         k -= 8 * sizeof(int8_t);
213       }
214 
215       if XNN_UNLIKELY(k != 0) {
216         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
217         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
218 
219         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223 
224         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
225         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
226         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
227         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
228         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
229         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
230         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
231         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
232         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
233         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
234         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
235         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
236         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
237         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
238         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
239         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
240         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
241         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
242       }
243       p -= 2 * sizeof(void*);
244     } while (p != 0);
245 
246 #if XNN_ARCH_ARM64
247     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
248     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
249     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
250     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
251 #else
252     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
253     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
254     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
255     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
256     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
257     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
258     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
259     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
260     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
261     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
262     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
263     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
264 #endif
265 
266     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
267     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
268     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
269     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
270 
271     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
272     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
273     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
274     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
275     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
276     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
277 
278     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
279     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
280     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
281     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
282 
283     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
284 #if XNN_ARCH_ARM64
285     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
286     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
287 
288     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
289     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
290 
291     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
292 #else
293     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
294     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
295 
296     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
297     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
298 
299     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
300 #endif
301 
302     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
303     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
304 
305     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
306     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
307 
308     if (nc >= 8) {
309       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
310       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
311 
312       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
313       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
314 
315       a = (const int8_t**restrict) ((uintptr_t) a - ks);
316 
317       nc -= 8;
318     } else {
319       if (nc & 4) {
320         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
321         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
322         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
323       }
324       if (nc & 2) {
325         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
326         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
327         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
328       }
329       if (nc & 1) {
330         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
331         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
332       }
333 
334       nc = 0;
335     }
336   } while (nc != 0);
337 }
338