xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/1x16-minmax-fp32-neonv8-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/neon-mlal-lane.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/igemm.h>
16 #include <xnnpack/intrinsics-polyfill.h>
17 
18 
xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const uint8_t** restrict a,
25     const void* restrict w,
26     uint8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const uint8_t* zero,
31     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 1);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(uint8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   uint8_t* c0 = c;
45 
46   const uint8x8_t vb_zero_point = vld1_dup_u8(&params->fp32_neonv8.kernel_zero_point[0]);
47   do {
48     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
52 
53     size_t p = ks;
54     do {
55       const uint8_t* restrict a0 = a[0];
56       if XNN_UNPREDICTABLE(a0 != zero) {
57         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
58       }
59       a += 1;
60 
61       size_t k = kc;
62       while (k >= 8 * sizeof(uint8_t)) {
63         const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
64         const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
65 
66         const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
67         const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
68 
69         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
70         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
71         const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
72         const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
73 
74         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
75         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
76         const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
77         const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
78 
79         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
80         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
81         const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
82         const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
83 
84         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
85         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
86         const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
87         const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
88 
89         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
90         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
91         const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
92         const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
93 
94         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
95         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
96         const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
97         const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
98 
99         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
100         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
101         const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
102         const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
103 
104         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
105         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
106 
107 
108         const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
109         const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
110 
111         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
112         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
113         const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
114         const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
115 
116         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
117         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
118         const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
119         const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
120 
121         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
122         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
123         const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
124         const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
125 
126         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
127         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
128         const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
129         const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
130 
131         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
132         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
133         const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
134         const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
135 
136         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
137         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
138         const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
139         const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
140 
141         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
142         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
143         const uint8x8_t vb89ABCDEFc7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
144         const int16x8_t vxb89ABCDEFc7 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc7, vb_zero_point));
145 
146         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
147         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
148 
149         k -= 8 * sizeof(uint8_t);
150       }
151       if XNN_UNLIKELY(k != 0) {
152         const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
153         const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
154 
155         const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
156         const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
157         const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
158         const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
159 
160         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
161         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
162         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
163         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
164 
165         if (k >= 2 * sizeof(uint8_t)) {
166           const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
167           const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
168           const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
169           const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
170 
171           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
172           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
173           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
174           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
175 
176           if (k > 2 * sizeof(uint8_t)) {
177             const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
178             const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
179             const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
180             const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
181 
182             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
183             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
184             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
185             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
186 
187             if (k >= 4 * sizeof(uint8_t)) {
188               const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
189               const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
190               const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
191               const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
192 
193               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
194               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
195               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
196               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
197 
198               if (k > 4 * sizeof(uint8_t)) {
199                 const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
200                 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
201                 const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
202                 const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
203 
204                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
205                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
206                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
207                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
208 
209                 if (k >= 6 * sizeof(uint8_t)) {
210                   const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
211                   const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
212                   const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
213                   const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
214 
215                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
216                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
217                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
218                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
219 
220                   if (k > 6 * sizeof(uint8_t)) {
221                     const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
222                     const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
223                     const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
224                     const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
225 
226                     vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
227                     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
228                     vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
229                     vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
230                   }
231                 }
232               }
233             }
234           }
235         }
236       }
237       p -= 1 * sizeof(void*);
238     } while (p != 0);
239 
240     // Post-accumulation work
241     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
242     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
243     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
244     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
245 
246     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
247     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
248     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
249     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
250     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
251 
252     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
253     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
254     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
255     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
256 
257     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
258 #if XNN_ARCH_ARM64
259     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
260     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
261 
262     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
263     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
264 
265     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
266 #else
267     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
268     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
269 
270     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
271     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
272 
273     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
274 #endif
275 
276     const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neonv8.output_min);
277     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
278 
279     const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neonv8.output_max);
280     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
281 
282     if (nc >= 16) {
283       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
284 
285       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
286 
287       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
288 
289       nc -= 16;
290     } else {
291       uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
292       if (nc & 8) {
293         vst1_u8(c0, vout0x01234567); c0 += 8;
294         vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
295       }
296       if (nc & 4) {
297         vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
298         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
299       }
300       if (nc & 2) {
301         vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
302         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
303       }
304       if (nc & 1) {
305         vst1_lane_u8(c0, vout0x01234567, 0);
306       }
307 
308       nc = 0;
309     }
310   } while (nc != 0);
311 }
312