xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/2x8c2-minmax-fp32-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_ld2r(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_ld2r(
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 int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
58       const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
59       const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
60       const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
61       const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
62       const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
63       const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
64       const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
65       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 
74       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
75       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
76       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
77       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
78 
79       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
80       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
81       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
83       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
84       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
85       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
86       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
87       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
88       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
90       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
91       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
92       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
93       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
94       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
95       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
96       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
97 
98       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
99       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
100       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
102       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
103       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
104       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
105       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
106       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
107       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
109       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
110       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
111       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
112       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
113       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
114       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
115       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
116 
117       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
118       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
119       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
121       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
122       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
123       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
124       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
125       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
126       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
128       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
129       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
130       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
131       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
132       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
133       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
134       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
135 
136       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
137       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
138       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
140       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
141       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
142       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
143       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
144       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
145       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
147       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
148       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
149       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
150 
151       k -= 16 * sizeof(int8_t);
152     }
153 
154     if (k >= 8 * sizeof(int8_t)) {
155       const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
156       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
157       const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
158       const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
159 
160       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 
169       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
170       const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
171 
172       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
173       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
174       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
175       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
176       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
177       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
178       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
179       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
180       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
181       const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
182 
183       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
184       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
185       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
186       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
187       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
188       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
189       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
190       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
191       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
192       const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
193 
194       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
195       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
196       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
197       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
198       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
199       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
200       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
201       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
202       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
203       const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
204 
205       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
206       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
207       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
208       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
209       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
210       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
211       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
212       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
213 
214       k -= 8 * sizeof(int8_t);
215     }
216 
217     if XNN_UNLIKELY(k != 0) {
218       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
219       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
220 
221       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223 
224       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
225       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
226       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
227       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
228       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
229       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
230       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
231       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
232       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
233       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
234 
235       if (k > 2 * sizeof(int8_t)) {
236         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 
239         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
240         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
241         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
242         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
243         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
244         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
245         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
246         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
247         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
248         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
249 
250         if (k > 4 * sizeof(int8_t)) {
251           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 
254           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
255           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
256           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
257           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
258           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
259           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
260           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
261           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
262           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
263           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
264         }
265       }
266     }
267 
268     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
269     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
270     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
271     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
272 
273     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
274     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
275     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
276     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
277     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
278     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
279 
280     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
281     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
282     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
283     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
284     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
285 
286     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
287     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
288     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
289     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
290     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
291 
292 #if XNN_ARCH_ARM64
293     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
294     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
295 
296 
297     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
298 #else
299     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
300     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
301 
302 
303     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
304 #endif
305 
306     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
307     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
308 
309     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
310     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
311 
312     if (nc >= 8) {
313       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
314       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
315 
316       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
317       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
318 
319       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
320       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
321 
322       nc -= 8;
323     } else {
324       // Final case where not all of the 8 columns fit in the destination.
325       if (nc & 4) {
326         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
327         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
328         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
329       }
330       if (nc & 2) {
331         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
332         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
333         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
334       }
335       if (nc & 1) {
336         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
337         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
338       }
339 
340       nc = 0;
341     }
342   } while (nc != 0);
343 }
344