xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x25-minmax-fp32-neonv8-mul8-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul8.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/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul8_ld128(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul8_ld128(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
34   const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     const int8_t* i3 = input[3];
53     assert(i3 != NULL);
54     if XNN_UNPREDICTABLE(i3 != zero) {
55       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56     }
57     const int8_t* i4 = input[4];
58     assert(i4 != NULL);
59     if XNN_UNPREDICTABLE(i4 != zero) {
60       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61     }
62     const int8_t* i5 = input[5];
63     assert(i5 != NULL);
64     if XNN_UNPREDICTABLE(i5 != zero) {
65       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66     }
67     const int8_t* i6 = input[6];
68     assert(i6 != NULL);
69     if XNN_UNPREDICTABLE(i6 != zero) {
70       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71     }
72     const int8_t* i7 = input[7];
73     assert(i7 != NULL);
74     if XNN_UNPREDICTABLE(i7 != zero) {
75       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76     }
77     const int8_t* i8 = input[8];
78     assert(i8 != NULL);
79     if XNN_UNPREDICTABLE(i8 != zero) {
80       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81     }
82     const int8_t* i9 = input[9];
83     assert(i9 != NULL);
84     if XNN_UNPREDICTABLE(i9 != zero) {
85       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86     }
87     const int8_t* i10 = input[10];
88     assert(i10 != NULL);
89     if XNN_UNPREDICTABLE(i10 != zero) {
90       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91     }
92     const int8_t* i11 = input[11];
93     assert(i11 != NULL);
94     if XNN_UNPREDICTABLE(i11 != zero) {
95       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96     }
97     const int8_t* i12 = input[12];
98     assert(i12 != NULL);
99     if XNN_UNPREDICTABLE(i12 != zero) {
100       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101     }
102     const int8_t* i13 = input[13];
103     assert(i13 != NULL);
104     if XNN_UNPREDICTABLE(i13 != zero) {
105       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106     }
107     const int8_t* i14 = input[14];
108     assert(i14 != NULL);
109     if XNN_UNPREDICTABLE(i14 != zero) {
110       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111     }
112     const int8_t* i15 = input[15];
113     assert(i15 != NULL);
114     if XNN_UNPREDICTABLE(i15 != zero) {
115       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116     }
117     const int8_t* i16 = input[16];
118     assert(i16 != NULL);
119     if XNN_UNPREDICTABLE(i16 != zero) {
120       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121     }
122     const int8_t* i17 = input[17];
123     assert(i17 != NULL);
124     if XNN_UNPREDICTABLE(i17 != zero) {
125       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126     }
127     const int8_t* i18 = input[18];
128     assert(i18 != NULL);
129     if XNN_UNPREDICTABLE(i18 != zero) {
130       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131     }
132     const int8_t* i19 = input[19];
133     assert(i19 != NULL);
134     if XNN_UNPREDICTABLE(i19 != zero) {
135       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136     }
137     const int8_t* i20 = input[20];
138     assert(i20 != NULL);
139     if XNN_UNPREDICTABLE(i20 != zero) {
140       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141     }
142     const int8_t* i21 = input[21];
143     assert(i21 != NULL);
144     if XNN_UNPREDICTABLE(i21 != zero) {
145       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146     }
147     const int8_t* i22 = input[22];
148     assert(i22 != NULL);
149     if XNN_UNPREDICTABLE(i22 != zero) {
150       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151     }
152     const int8_t* i23 = input[23];
153     assert(i23 != NULL);
154     if XNN_UNPREDICTABLE(i23 != zero) {
155       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156     }
157     const int8_t* i24 = input[24];
158     assert(i24 != NULL);
159     if XNN_UNPREDICTABLE(i24 != zero) {
160       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161     }
162     input = (const int8_t**) ((uintptr_t) input + input_stride);
163 
164     size_t c = channels;
165     const void* w = weights;
166     for (; c >= 16; c -= 16) {
167       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
168       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171 
172       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
173       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
174 
175       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
176       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
177 
178       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
179       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
180       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
181       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
182       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
183       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
184 
185       vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
186       vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
187 
188       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
189       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
190       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
191       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
192       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
193       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
194 
195       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
196       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
197 
198       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
199       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
200       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
201       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
202       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
203       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
204 
205       vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
206       vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
207 
208       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
209       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
210       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
211       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
212       const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
213       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
214 
215       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
216       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
217 
218       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
219       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
220       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
221       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
222       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
223       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
224 
225       vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
226       vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
227 
228       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
229       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
230       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
231       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
232       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
233       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
234 
235       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
236       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
237 
238       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
239       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
240       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
241       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
242       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
243       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
244 
245       vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
246       vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
247 
248       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
249       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
250       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
251       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
252       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
253       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
254 
255       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
256       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
257 
258       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
259       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
260       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
261       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
262       const int8x16_t vi9x0123456789ABCDEF = vld1q_s8(i9); i9 += 16;
263       const int8x16_t vk9x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
264 
265       vprod01234567 = vmull_s8(vget_low_s8(vi9x0123456789ABCDEF), vget_low_s8(vk9x0123456789ABCDEF));
266       vprod89ABCDEF = vmull_s8(vget_high_s8(vi9x0123456789ABCDEF), vget_high_s8(vk9x0123456789ABCDEF));
267 
268       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
269       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
270       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
271       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
272       const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16;
273       const int8x16_t vk10x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
274 
275       vprod01234567 = vmull_s8(vget_low_s8(vi10x0123456789ABCDEF), vget_low_s8(vk10x0123456789ABCDEF));
276       vprod89ABCDEF = vmull_s8(vget_high_s8(vi10x0123456789ABCDEF), vget_high_s8(vk10x0123456789ABCDEF));
277 
278       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
279       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
280       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
281       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
282       const int8x16_t vi11x0123456789ABCDEF = vld1q_s8(i11); i11 += 16;
283       const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
284 
285       vprod01234567 = vmull_s8(vget_low_s8(vi11x0123456789ABCDEF), vget_low_s8(vk11x0123456789ABCDEF));
286       vprod89ABCDEF = vmull_s8(vget_high_s8(vi11x0123456789ABCDEF), vget_high_s8(vk11x0123456789ABCDEF));
287 
288       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
289       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
290       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
291       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
292       const int8x16_t vi12x0123456789ABCDEF = vld1q_s8(i12); i12 += 16;
293       const int8x16_t vk12x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
294 
295       vprod01234567 = vmull_s8(vget_low_s8(vi12x0123456789ABCDEF), vget_low_s8(vk12x0123456789ABCDEF));
296       vprod89ABCDEF = vmull_s8(vget_high_s8(vi12x0123456789ABCDEF), vget_high_s8(vk12x0123456789ABCDEF));
297 
298       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
299       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
300       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
301       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
302       const int8x16_t vi13x0123456789ABCDEF = vld1q_s8(i13); i13 += 16;
303       const int8x16_t vk13x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
304 
305       vprod01234567 = vmull_s8(vget_low_s8(vi13x0123456789ABCDEF), vget_low_s8(vk13x0123456789ABCDEF));
306       vprod89ABCDEF = vmull_s8(vget_high_s8(vi13x0123456789ABCDEF), vget_high_s8(vk13x0123456789ABCDEF));
307 
308       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
309       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
310       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
311       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
312       const int8x16_t vi14x0123456789ABCDEF = vld1q_s8(i14); i14 += 16;
313       const int8x16_t vk14x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
314 
315       vprod01234567 = vmull_s8(vget_low_s8(vi14x0123456789ABCDEF), vget_low_s8(vk14x0123456789ABCDEF));
316       vprod89ABCDEF = vmull_s8(vget_high_s8(vi14x0123456789ABCDEF), vget_high_s8(vk14x0123456789ABCDEF));
317 
318       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
319       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
320       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
321       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
322       const int8x16_t vi15x0123456789ABCDEF = vld1q_s8(i15); i15 += 16;
323       const int8x16_t vk15x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
324 
325       vprod01234567 = vmull_s8(vget_low_s8(vi15x0123456789ABCDEF), vget_low_s8(vk15x0123456789ABCDEF));
326       vprod89ABCDEF = vmull_s8(vget_high_s8(vi15x0123456789ABCDEF), vget_high_s8(vk15x0123456789ABCDEF));
327 
328       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
329       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
330       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
331       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
332       const int8x16_t vi16x0123456789ABCDEF = vld1q_s8(i16); i16 += 16;
333       const int8x16_t vk16x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
334 
335       vprod01234567 = vmull_s8(vget_low_s8(vi16x0123456789ABCDEF), vget_low_s8(vk16x0123456789ABCDEF));
336       vprod89ABCDEF = vmull_s8(vget_high_s8(vi16x0123456789ABCDEF), vget_high_s8(vk16x0123456789ABCDEF));
337 
338       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
339       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
340       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
341       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
342       const int8x16_t vi17x0123456789ABCDEF = vld1q_s8(i17); i17 += 16;
343       const int8x16_t vk17x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
344 
345       vprod01234567 = vmull_s8(vget_low_s8(vi17x0123456789ABCDEF), vget_low_s8(vk17x0123456789ABCDEF));
346       vprod89ABCDEF = vmull_s8(vget_high_s8(vi17x0123456789ABCDEF), vget_high_s8(vk17x0123456789ABCDEF));
347 
348       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
349       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
350       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
351       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
352       const int8x16_t vi18x0123456789ABCDEF = vld1q_s8(i18); i18 += 16;
353       const int8x16_t vk18x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
354 
355       vprod01234567 = vmull_s8(vget_low_s8(vi18x0123456789ABCDEF), vget_low_s8(vk18x0123456789ABCDEF));
356       vprod89ABCDEF = vmull_s8(vget_high_s8(vi18x0123456789ABCDEF), vget_high_s8(vk18x0123456789ABCDEF));
357 
358       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
359       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
360       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
361       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
362       const int8x16_t vi19x0123456789ABCDEF = vld1q_s8(i19); i19 += 16;
363       const int8x16_t vk19x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
364 
365       vprod01234567 = vmull_s8(vget_low_s8(vi19x0123456789ABCDEF), vget_low_s8(vk19x0123456789ABCDEF));
366       vprod89ABCDEF = vmull_s8(vget_high_s8(vi19x0123456789ABCDEF), vget_high_s8(vk19x0123456789ABCDEF));
367 
368       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
369       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
370       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
371       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
372       const int8x16_t vi20x0123456789ABCDEF = vld1q_s8(i20); i20 += 16;
373       const int8x16_t vk20x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
374 
375       vprod01234567 = vmull_s8(vget_low_s8(vi20x0123456789ABCDEF), vget_low_s8(vk20x0123456789ABCDEF));
376       vprod89ABCDEF = vmull_s8(vget_high_s8(vi20x0123456789ABCDEF), vget_high_s8(vk20x0123456789ABCDEF));
377 
378       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
379       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
380       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
381       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
382       const int8x16_t vi21x0123456789ABCDEF = vld1q_s8(i21); i21 += 16;
383       const int8x16_t vk21x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
384 
385       vprod01234567 = vmull_s8(vget_low_s8(vi21x0123456789ABCDEF), vget_low_s8(vk21x0123456789ABCDEF));
386       vprod89ABCDEF = vmull_s8(vget_high_s8(vi21x0123456789ABCDEF), vget_high_s8(vk21x0123456789ABCDEF));
387 
388       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
389       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
390       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
391       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
392       const int8x16_t vi22x0123456789ABCDEF = vld1q_s8(i22); i22 += 16;
393       const int8x16_t vk22x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
394 
395       vprod01234567 = vmull_s8(vget_low_s8(vi22x0123456789ABCDEF), vget_low_s8(vk22x0123456789ABCDEF));
396       vprod89ABCDEF = vmull_s8(vget_high_s8(vi22x0123456789ABCDEF), vget_high_s8(vk22x0123456789ABCDEF));
397 
398       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
399       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
400       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
401       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
402       const int8x16_t vi23x0123456789ABCDEF = vld1q_s8(i23); i23 += 16;
403       const int8x16_t vk23x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
404 
405       vprod01234567 = vmull_s8(vget_low_s8(vi23x0123456789ABCDEF), vget_low_s8(vk23x0123456789ABCDEF));
406       vprod89ABCDEF = vmull_s8(vget_high_s8(vi23x0123456789ABCDEF), vget_high_s8(vk23x0123456789ABCDEF));
407 
408       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
409       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
410       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
411       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
412       const int8x16_t vi24x0123456789ABCDEF = vld1q_s8(i24); i24 += 16;
413       const int8x16_t vk24x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
414 
415       vprod01234567 = vmull_s8(vget_low_s8(vi24x0123456789ABCDEF), vget_low_s8(vk24x0123456789ABCDEF));
416       vprod89ABCDEF = vmull_s8(vget_high_s8(vi24x0123456789ABCDEF), vget_high_s8(vk24x0123456789ABCDEF));
417 
418       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
419       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
420       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
421       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
422 
423       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
424       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
425       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
426       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
427 
428       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
429       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
430       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
431       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
432 
433       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
434       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
435       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
436       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
437 
438       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
439       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
440       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
441       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
442 
443 #if XNN_ARCH_ARM64
444       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
445       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
446 
447       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
448       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
449 
450       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
451 #else  // !XNN_ARCH_ARM64
452       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
453       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
454 
455       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
456       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
457 
458       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
459 #endif  // !XNN_ARCH_ARM64
460 
461       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
462 
463       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
464 
465       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
466     }
467     if XNN_UNLIKELY(c != 0) {
468       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
469       do {
470         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
471         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
472 
473         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
474         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
475 
476         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
477 
478         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
479         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
480         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
481         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
482 
483         vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
484 
485         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
486         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
487         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
488         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
489 
490         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
491 
492         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
493         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
494         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
495         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
496 
497         vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
498 
499         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
500         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
501         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
502         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
503 
504         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
505 
506         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
507         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
508         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
509         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
510 
511         vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
512 
513         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
514         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
515         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
516         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
517 
518         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
519 
520         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
521         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
522         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
523         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
524 
525         vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
526 
527         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
528         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
529         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
530         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
531 
532         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
533 
534         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
535         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
536         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
537         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
538 
539         vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
540 
541         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
542         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
543         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
544         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
545 
546         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
547 
548         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
549         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
550         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
551         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
552 
553         vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
554 
555         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
556         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
557         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
558         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
559 
560         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
561 
562         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
563         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
564         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
565         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
566 
567         vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
568 
569         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
570         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
571         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
572         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
573 
574         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
575 
576         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
577         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
578         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
579         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
580 
581         vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
582 
583         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
584         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
585         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
586         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
587 
588         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
589 
590         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
591         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
592         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
593         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
594 
595         vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
596 
597         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
598         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
599         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
600         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
601 
602         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
603 
604         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
605         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
606         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
607         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
608 
609         vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
610 
611         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
612         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
613         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
614         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
615 
616         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
617 
618         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
619         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
620         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
621         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
622 
623         vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
624 
625         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
626         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
627         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
628         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
629 
630         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
631 
632         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
633         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
634         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
635         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
636 
637         vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
638 
639         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
640         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
641         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
642         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
643 
644         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
645 
646         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
647         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
648 
649         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
650         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
651 
652         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t)));
653         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
654         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
655         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
656 
657         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
658         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
659 
660 #if XNN_ARCH_ARM64
661         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
662 #else
663         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
664 #endif
665         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
666 
667         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
668         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
669         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
670 
671         if XNN_LIKELY(c >= 8) {
672           vst1_s8(output, vout01234567); output += 8;
673           c -= 8;
674         } else {
675           if (c & 4) {
676             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
677             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
678           }
679           if (c & 2) {
680             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
681             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
682           }
683           if (c & 1) {
684             vst1_lane_s8(output, vout01234567, 0); output += 1;
685           }
686           c = 0;
687         }
688       } while (c != 0);
689     }
690 
691     output = (int8_t*) ((uintptr_t) output + output_increment);
692   } while (--output_width != 0);
693 }
694