• 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/math.h>
16 
17 
xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup(
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_qs8_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, 4 * 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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55     int32x4_t vacc1x01 = vacc0x01;
56     int32x4_t vacc1x23 = vacc0x23;
57     int32x4_t vacc1x45 = vacc0x45;
58     int32x4_t vacc1x67 = vacc0x67;
59 
60     size_t p = ks;
61     do {
62       const int8_t* restrict a0 = a[0];
63       if XNN_UNPREDICTABLE(a0 != zero) {
64         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65       }
66       const int8_t* restrict a1 = a[1];
67       if XNN_UNPREDICTABLE(a1 != zero) {
68         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69       }
70       a += 2;
71 
72       size_t k = kc;
73 
74       while (k >= 16 * sizeof(int8_t)) {
75         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
76         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
77         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
78         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
79 
80         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 
89         const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
90         const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
91         const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
92         const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
93 
94         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
95         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
96         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
98         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
99         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
100         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
101         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
102         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
103         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
105         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
106         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
107         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
108         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
109         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
110         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
112         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
113         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
114         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
115         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
116         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
117         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
119         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
120         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
121         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
122         const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
123         const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
124         const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
125         const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
126 
127         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
128         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
129         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
131         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
132         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
133         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
134         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
135         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
136         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
138         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
139         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
140         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
141         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
142         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
143         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
145         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
146         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
147         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
148         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
149         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
150         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
152         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
153         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
154         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
155 
156         k -= 16 * sizeof(int8_t);
157       }
158 
159       if (k >= 8 * sizeof(int8_t)) {
160         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
161         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
162 
163         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 
172         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
173         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
174 
175         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
176         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
177         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
178         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
179         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
180         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
181         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
182         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
183         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
184         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
185         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
186         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
187         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
188         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
189         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
190         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
191         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
192         const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
193 
194         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
195         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
196         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
197         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
198         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
199         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
200         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
201         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
202         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
203         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
204         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
205         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
206         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
207         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
208         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
209         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
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         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
217 
218         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222 
223         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
224         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
225         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
226         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
227         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
228         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
229         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
230         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
231         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
232         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
233         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
234         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
235         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
236         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
237         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
238         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
239         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
240         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
241       }
242       p -= 2 * sizeof(void*);
243     } while (p != 0);
244 
245 #if XNN_ARCH_ARM64
246     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
247     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
248     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
249     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
250 #else
251     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
252     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
253     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
254     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
255     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
256     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
257     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
258     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
259     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
260     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
261     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
262     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
263 #endif
264 
265     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
266     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
267     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
268     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
269 
270     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
271     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
272     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
273     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
274     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
275 
276     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
277     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
278     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
279     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
280     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
281 
282     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
283     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
284     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
285     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
286     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
287 
288 #if XNN_ARCH_ARM64
289     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
290     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
291 
292 
293     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
294 #else
295     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
296     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
297 
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_neon.output_min);
303     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
304 
305     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.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