• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
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(ks != 0);
36   assert(ks % (2 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr != 2) {
46     c1 = c0;
47   }
48 
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc1x0123 = vacc0x0123;
53     int32x4_t vacc1x4567 = vacc0x4567;
54 
55     size_t p = ks;
56     do {
57       const int8_t* restrict a0 = a[0];
58       if XNN_UNPREDICTABLE(a0 != zero) {
59         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60       }
61       const int8_t* restrict a1 = a[1];
62       if XNN_UNPREDICTABLE(a1 != zero) {
63         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64       }
65       a += 2;
66 
67       size_t k = kc;
68 
69       while (k >= 16 * sizeof(int8_t)) {
70         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
71         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
72         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
73         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
74         const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
75         const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
76         const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
77         const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
78 
79         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 
88         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
89         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
90         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
91         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
92 
93         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
94         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
95         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
97         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
98         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
99         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
100         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
101         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
102         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
104         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
105         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
106         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
107         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
108         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
109         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
110         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
111 
112         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
113         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
114         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
116         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
117         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
118         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
119         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
120         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
121         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
123         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
124         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
125         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
126         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
127         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
128         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
129         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
130 
131         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
132         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
133         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
135         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
136         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
137         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
138         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
139         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
140         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
142         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
143         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
144         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
145         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
146         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
147         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
148         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
149 
150         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
151         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
152         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
154         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
155         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
156         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
157         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
158         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
159         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
161         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
162         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
163         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
164 
165         k -= 16 * sizeof(int8_t);
166       }
167 
168       if (k >= 8 * sizeof(int8_t)) {
169         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
170         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
171         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
172         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
173 
174         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182 
183         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
184         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
185 
186         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
187         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
188         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
189         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
190         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
191         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
192         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
193         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
194         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
195         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
196 
197         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
198         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
199         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
200         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
201         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
202         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
203         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
204         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
205         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
206         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
207 
208         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
209         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
210         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
211         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
212         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
213         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
214         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
215         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
216         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
217         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
218 
219         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
220         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
221         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
222         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
223         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
224         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
225         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
226         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
227 
228         k -= 8 * sizeof(int8_t);
229       }
230 
231       if XNN_UNLIKELY(k != 0) {
232         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
233         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
234 
235         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237 
238         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
239         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
240         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
241         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
242         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
243         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
244         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
245         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
246         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
247         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
248 
249         if (k > 2 * sizeof(int8_t)) {
250           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252 
253           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
254           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
255           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
256           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
257           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
258           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
259           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
260           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
261           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
262           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
263 
264           if (k > 4 * sizeof(int8_t)) {
265             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
266             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267 
268             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
269             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
270             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
271             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
272             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
273             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
274             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
275             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
276             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
277             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
278           }
279         }
280       }
281       p -= 2 * sizeof(void*);
282     } while (p != 0);
283 
284     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
285     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
286     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
287     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
288 
289     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
290     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
291     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
292     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
293     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
294     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
295 
296     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
297     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
298     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
299     vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
300     vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
301 
302     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
303     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
304     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
305     vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
306     vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
307 
308 #if XNN_ARCH_ARM64
309     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
310     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
311 
312 
313     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
314 #else
315     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
316     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
317 
318 
319     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
320 #endif
321 
322     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
323     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
324 
325     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
326     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
327 
328     if (nc >= 8) {
329       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
330       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
331 
332       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
333       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
334 
335       a = (const int8_t**restrict) ((uintptr_t) a - ks);
336 
337       nc -= 8;
338     } else {
339       if (nc & 4) {
340         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
341         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
342         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
343       }
344       if (nc & 2) {
345         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
346         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
347         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
348       }
349       if (nc & 1) {
350         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
351         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
352       }
353 
354       nc = 0;
355     }
356   } while (nc != 0);
357 }
358