xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 2);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 4 * sizeof(int8_t));
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr != 2) {
46     a1 = a0;
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 k = kc;
61 
62     while (k >= 16 * sizeof(int8_t)) {
63       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
64       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
65       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
66       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
67 
68       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 
77       const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
78       const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
79       const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
80       const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
81 
82       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
83       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
84       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
86       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
87       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
88       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
89       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
90       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
91       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
93       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
94       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
95       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
96       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
97       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
98       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
100       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
101       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
102       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
103       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
104       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
105       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
107       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
108       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
109       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
110       const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
111       const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
112       const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
113       const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
114 
115       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
116       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
117       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
119       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
120       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
121       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
122       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
123       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
124       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
126       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
127       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
128       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
129       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
130       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
131       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
133       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
134       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
135       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
136       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
137       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
138       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
140       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
141       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
142       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
143 
144       k -= 16 * sizeof(int8_t);
145     }
146 
147     if (k >= 8 * sizeof(int8_t)) {
148       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
149       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
150 
151       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159 
160       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
161       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
162 
163       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
164       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
165       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
166       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
167       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
168       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
169       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
170       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
171       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
172       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
173       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
174       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
175       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
176       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
177       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
178       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
179       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
180       const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
181 
182       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
183       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
184       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
185       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
186       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
187       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
188       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
189       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
190       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
191       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
192       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
193       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
194       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
195       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
196       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
197       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
198 
199       k -= 8 * sizeof(int8_t);
200     }
201 
202     if XNN_UNLIKELY(k != 0) {
203       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
204       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
205 
206       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210 
211       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
212       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
213       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
214       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
215       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
216       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
217       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
218       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
219       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
220       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
221       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
222       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
223       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
224       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
225       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
226       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
227       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
228       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
229     }
230 
231 #if XNN_ARCH_ARM64
232     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
233     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
234     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
235     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
236 #else
237     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
238     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
239     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
240     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
241     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
242     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
243     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
244     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
245     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
246     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
247     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
248     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
249 #endif
250 
251     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
252     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
253     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
254     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
255 
256     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
257     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
258     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
259     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
260     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
261     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
262 
263     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
264     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
265     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
266     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
267 
268     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
269 #if XNN_ARCH_ARM64
270     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
271     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
272 
273     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
274     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
275 
276     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
277 #else
278     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
279     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
280 
281     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
282     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
283 
284     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
285 #endif
286 
287     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
288     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
289 
290     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
291     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
292 
293     if (nc >= 8) {
294       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
295       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
296 
297       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
298       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
299 
300       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
301       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
302 
303       nc -= 8;
304     } else {
305       // Final case where not all of the 8 columns fit in the destination.
306       if (nc & 4) {
307         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
308         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
309         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
310       }
311       if (nc & 2) {
312         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
313         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
314         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
315       }
316       if (nc & 1) {
317         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
318         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
319       }
320 
321       nc = 0;
322     }
323   } while (nc != 0);
324 }
325