xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.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/math.h>
16 
17 
xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r(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)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 2);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr != 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48 
49   do {
50     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc1x01 = vacc0x01;
55     int32x4_t vacc1x23 = vacc0x23;
56     int32x4_t vacc1x45 = vacc0x45;
57     int32x4_t vacc1x67 = vacc0x67;
58 
59     size_t k = kc;
60 
61     while (k >= 16 * sizeof(int8_t)) {
62       const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
63       const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
64       const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
65       const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
66       const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
67       const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
68       const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
69       const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
70 
71       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 
80       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
81       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
82       const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
83       const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
84 
85       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
86       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
87       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
89       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
90       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
91       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
92       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
93       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
94       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
96       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
97       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
98       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
99       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
100       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
101       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
103       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
104       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
105       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
106       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
107       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
108       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
110       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
111       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
112       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
113       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
114       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
115       const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
116       const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
117 
118       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
119       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
120       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
122       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
123       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
124       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
125       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
126       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
127       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
129       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
130       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
131       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
132       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
133       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
134       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
136       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
137       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
138       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
139       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
140       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
141       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
143       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
144       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
145       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
146 
147       k -= 16 * sizeof(int8_t);
148     }
149 
150     if (k >= 8 * sizeof(int8_t)) {
151       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
152       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
153       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
154       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
155 
156       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 
165       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
166       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
167 
168       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
169       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
170       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
171       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
172       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
173       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
174       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
175       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
176       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
177       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
178       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
179       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
180       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
181       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
182       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
183       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
184       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
185       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
186 
187       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
188       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
189       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
190       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
191       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
192       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
193       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
194       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
195       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
196       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
197       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
198       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
199       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
200       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
201       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
202       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
203 
204       k -= 8 * sizeof(int8_t);
205     }
206 
207     if XNN_UNLIKELY(k != 0) {
208       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
209       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
210 
211       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215 
216       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
217       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
218       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
219       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
220       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
221       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
222       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
223       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
224       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
225       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
226       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
227       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
228       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
229       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
230       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
231       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
232       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
233       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
234     }
235 
236 #if XNN_ARCH_ARM64
237     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
238     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
239     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
240     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
241 #else
242     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
243     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
244     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
245     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
246     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
247     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
248     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
249     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
250     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
251     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
252     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
253     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
254 #endif
255 
256     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
257     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
258     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
259     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
260 
261     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
262     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
263     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
264     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
265     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
266     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
267 
268     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
269     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
270     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
271     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
272     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
273 
274     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
275     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
276     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
277     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
278     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
279 
280 #if XNN_ARCH_ARM64
281     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
282     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
283 
284 
285     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
286 #else
287     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
288     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
289 
290 
291     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
292 #endif
293 
294     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
295     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
296 
297     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
298     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
299 
300     if (nc >= 8) {
301       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
302       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
303 
304       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
305       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
306 
307       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
308       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
309 
310       nc -= 8;
311     } else {
312       // Final case where not all of the 8 columns fit in the destination.
313       if (nc & 4) {
314         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
315         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
316         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
317       }
318       if (nc & 2) {
319         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
320         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
321         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
322       }
323       if (nc & 1) {
324         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
325         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
326       }
327 
328       nc = 0;
329     }
330   } while (nc != 0);
331 }
332