xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-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/math.h>
16 
17 
xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_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)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup(
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 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
64       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
65       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
66 
67       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 
76       const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
77       const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
78       const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
79       const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
80 
81       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
82       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
83       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
85       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
86       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
87       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
88       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
89       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
90       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
92       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
93       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
94       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
95       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
96       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
97       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
99       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
100       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
101       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
102       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
103       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
104       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
106       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
107       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
108       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
109       const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
110       const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
111       const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
112       const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
113 
114       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
115       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
116       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
118       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
119       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
120       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
121       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
122       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
123       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
125       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
126       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
127       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
128       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
129       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
130       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
132       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
133       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
134       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
135       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
136       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
137       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
139       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
140       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
141       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
142 
143       k -= 16 * sizeof(int8_t);
144     }
145 
146     if (k >= 8 * sizeof(int8_t)) {
147       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
148       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
149 
150       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 
159       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
160       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
161 
162       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
163       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
164       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
165       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
166       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
167       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
168       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
169       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
170       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
171       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
172       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
173       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
174       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
175       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
176       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
177       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
178       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
179       const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
180 
181       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
182       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
183       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
184       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
185       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
186       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
187       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
188       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
189       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
190       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
191       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
192       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
193       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
194       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
195       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
196       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
197 
198       k -= 8 * sizeof(int8_t);
199     }
200 
201     if XNN_UNLIKELY(k != 0) {
202       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
203       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
204 
205       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209 
210       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
211       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
212       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
213       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
214       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
215       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
216       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
217       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
218       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
219       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
220       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
221       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
222       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
223       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
224       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
225       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
226       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
227       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
228     }
229 
230 #if XNN_ARCH_ARM64
231     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
232     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
233     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
234     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
235 #else
236     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
237     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
238     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
239     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
240     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
241     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
242     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
243     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
244     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
245     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
246     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
247     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
248 #endif
249 
250     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
251     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
252     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
253     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
254 
255     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
256     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
257     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
258     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
259     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
260     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
261 
262     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
263     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
264     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
265     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
266     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
267 
268     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
269     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
270     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
271     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
272     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
273 
274 #if XNN_ARCH_ARM64
275     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
276     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
277 
278 
279     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
280 #else
281     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
282     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
283 
284 
285     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
286 #endif
287 
288     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
289     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
290 
291     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
292     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
293 
294     if (nc >= 8) {
295       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
296       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
297 
298       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
299       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
300 
301       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
302       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
303 
304       nc -= 8;
305     } else {
306       // Final case where not all of the 8 columns fit in the destination.
307       if (nc & 4) {
308         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
309         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
310         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
311       }
312       if (nc & 2) {
313         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
314         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
315         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
316       }
317       if (nc & 1) {
318         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
319         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
320       }
321 
322       nc = 0;
323     }
324   } while (nc != 0);
325 }
326