• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-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 
xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r(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)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 2);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr != 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47 
48   do {
49     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc1x0123 = vacc0x0123;
52     int32x4_t vacc1x4567 = vacc0x4567;
53 
54     size_t k = kc;
55 
56     while (k >= 16 * sizeof(int8_t)) {
57       const int16x4x4_t va0x0 = vld4_dup_s16((const void*)a0); a0 += 8;
58       const int16x4x4_t va0x1 = vld4_dup_s16((const void*)a0); a0 += 8;
59       const int16x4x4_t va1x0 = vld4_dup_s16((const void*)a1); a1 += 8;
60       const int16x4x4_t va1x1 = vld4_dup_s16((const void*)a1); a1 += 8;
61       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 
70       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va0x0.val[0]);
71       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va0x1.val[0]);
72       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va1x0.val[0]);
73       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va1x1.val[0]);
74 
75       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
76       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
77       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
79       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
80       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
81       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
82       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
83       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
84       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
86       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
87       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
88       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
89       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va0x0.val[1]);
90       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va0x1.val[1]);
91       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va1x0.val[1]);
92       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va1x1.val[1]);
93 
94       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
95       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
96       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
98       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
99       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
100       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
101       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
102       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
103       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
105       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
106       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
107       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
108       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va0x0.val[2]);
109       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va0x1.val[2]);
110       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va1x0.val[2]);
111       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va1x1.val[2]);
112 
113       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
114       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
115       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
117       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
118       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
119       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
120       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
121       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
122       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
124       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
125       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
126       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
127       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va0x0.val[3]);
128       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]);
129       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va1x0.val[3]);
130       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va1x1.val[3]);
131 
132       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
133       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
134       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
136       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
137       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
138       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
139       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
140       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
141       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
143       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
144       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
145       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
146 
147       k -= 16 * sizeof(int8_t);
148     }
149 
150     if (k >= 8 * sizeof(int8_t)) {
151       const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
152       const int16x4x4_t va1 = vld4_dup_s16((const void*)a1); a1 += 8;
153 
154       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 
163       const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
164       const int8x8_t va1c0 = vreinterpret_s8_s16(va1.val[0]);
165 
166       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
167       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
168       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
169       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
170       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
171       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
172       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
173       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
174       const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
175       const int8x8_t va1c1 = vreinterpret_s8_s16(va1.val[1]);
176 
177       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
178       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
179       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
180       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
181       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
182       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
183       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
184       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
185       const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
186       const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]);
187 
188       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
189       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
190       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
191       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
192       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
193       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
194       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
195       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
196       const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
197       const int8x8_t va1c3 = vreinterpret_s8_s16(va1.val[3]);
198 
199       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
200       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
201       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
202       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
203       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
204       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
205       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
206       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
207 
208       k -= 8 * sizeof(int8_t);
209     }
210 
211     if XNN_UNLIKELY(k != 0) {
212       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
213       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
214 
215       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217 
218       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
219       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
220       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
221       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
222       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
223       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
224       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
225       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
226       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
227       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
228 
229       if (k > 2 * sizeof(int8_t)) {
230         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 
233         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
234         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
235         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
236         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
237         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
238         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
239         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
240         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
241         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
242         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
243 
244         if (k > 4 * sizeof(int8_t)) {
245           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247 
248           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
249           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
250           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
251           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
252           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
253           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
254           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
255           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
256           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
257           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
258         }
259       }
260     }
261 
262     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
263     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
264     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
265     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
266 
267     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
268     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
269     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
270     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
271     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
272     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
273 
274     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
275     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
276     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
277     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
278     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
279 
280     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
281     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
282     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
283     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
284     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
285 
286 #if XNN_ARCH_ARM64
287     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
288     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
289 
290 
291     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
292 #else
293     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
294     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
295 
296 
297     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
298 #endif
299 
300     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
301     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
302 
303     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
304     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
305 
306     if (nc >= 8) {
307       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
308       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
309 
310       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
311       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
312 
313       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
314       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
315 
316       nc -= 8;
317     } else {
318       // Final case where not all of the 8 columns fit in the destination.
319       if (nc & 4) {
320         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
321         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
322         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
323       }
324       if (nc & 2) {
325         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
326         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
327         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
328       }
329       if (nc & 1) {
330         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
331         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
332       }
333 
334       nc = 0;
335     }
336   } while (nc != 0);
337 }
338