xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r(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_2x8c4__neon_mlal_ld1r(
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, 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 int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
76         const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
77         const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
78         const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
79         const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
80         const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
81         const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
82         const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
83 
84         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 
93         const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
94         const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
95         const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
96         const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
97 
98         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
99         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
100         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
102         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
103         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
104         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
105         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
106         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
107         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
109         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
110         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
111         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
112         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
113         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
114         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
116         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
117         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
118         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
119         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
120         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
121         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
123         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
124         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
125         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
126         const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
127         const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
128         const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
129         const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
130 
131         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
132         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
133         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
135         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
136         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
137         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
138         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
139         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
140         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
142         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
143         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
144         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
145         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
146         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
147         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
149         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
150         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
151         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
152         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
153         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
154         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
156         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
157         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
158         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
159 
160         k -= 16 * sizeof(int8_t);
161       }
162 
163       if (k >= 8 * sizeof(int8_t)) {
164         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
165         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
166         const int32x2_t va10 = vld1_dup_s32((const void*)a1);
167         const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
168 
169         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 
178         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
179         const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
180 
181         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
182         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
183         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
184         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
185         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
186         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
187         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
188         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
189         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
190         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
191         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
192         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
193         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
194         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
195         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
196         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
197         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
198         const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
199 
200         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
201         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
202         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
203         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
204         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
205         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
206         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
207         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
208         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
209         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
210         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
211         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
212         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
213         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
214         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
215         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
216 
217         k -= 8 * sizeof(int8_t);
218       }
219 
220       if XNN_UNLIKELY(k != 0) {
221         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
222         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
223 
224         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228 
229         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
230         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
231         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
232         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
233         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
234         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
235         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
236         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
237         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
238         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
239         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
240         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
241         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
242         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
243         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
244         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
245         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
246         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
247       }
248       p -= 2 * sizeof(void*);
249     } while (p != 0);
250 
251 #if XNN_ARCH_ARM64
252     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
253     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
254     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
255     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
256 #else
257     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
258     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
259     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
260     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
261     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
262     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
263     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
264     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
265     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
266     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
267     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
268     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
269 #endif
270 
271     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
272     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
273     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
274     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
275 
276     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
277     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
278     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
279     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
280     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
281     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
282 
283     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
284     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
285     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
286     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
287     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
288 
289     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
290     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
291     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
292     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
293     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
294 
295 #if XNN_ARCH_ARM64
296     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
297     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
298 
299 
300     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
301 #else
302     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
303     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
304 
305 
306     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
307 #endif
308 
309     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
310     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
311 
312     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
313     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
314 
315     if (nc >= 8) {
316       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
317       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
318 
319       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
320       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
321 
322       a = (const int8_t**restrict) ((uintptr_t) a - ks);
323 
324       nc -= 8;
325     } else {
326       if (nc & 4) {
327         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
328         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
329         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
330       }
331       if (nc & 2) {
332         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
333         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
334         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
335       }
336       if (nc & 1) {
337         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
338         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
339       }
340 
341       nc = 0;
342     }
343   } while (nc != 0);
344 }
345