xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x25-minmax-rndnu-neon-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 
16 
xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8_ld128(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
33   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
34   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
35   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
36   const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
37   const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
38   do {
39     const int8_t* i0 = input[0];
40     assert(i0 != NULL);
41     if XNN_UNPREDICTABLE(i0 != zero) {
42       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43     }
44     const int8_t* i1 = input[1];
45     assert(i1 != NULL);
46     if XNN_UNPREDICTABLE(i1 != zero) {
47       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48     }
49     const int8_t* i2 = input[2];
50     assert(i2 != NULL);
51     if XNN_UNPREDICTABLE(i2 != zero) {
52       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53     }
54     const int8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const int8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const int8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const int8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const int8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const int8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     const int8_t* i9 = input[9];
85     assert(i9 != NULL);
86     if XNN_UNPREDICTABLE(i9 != zero) {
87       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
88     }
89     const int8_t* i10 = input[10];
90     assert(i10 != NULL);
91     if XNN_UNPREDICTABLE(i10 != zero) {
92       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
93     }
94     const int8_t* i11 = input[11];
95     assert(i11 != NULL);
96     if XNN_UNPREDICTABLE(i11 != zero) {
97       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
98     }
99     const int8_t* i12 = input[12];
100     assert(i12 != NULL);
101     if XNN_UNPREDICTABLE(i12 != zero) {
102       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
103     }
104     const int8_t* i13 = input[13];
105     assert(i13 != NULL);
106     if XNN_UNPREDICTABLE(i13 != zero) {
107       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
108     }
109     const int8_t* i14 = input[14];
110     assert(i14 != NULL);
111     if XNN_UNPREDICTABLE(i14 != zero) {
112       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
113     }
114     const int8_t* i15 = input[15];
115     assert(i15 != NULL);
116     if XNN_UNPREDICTABLE(i15 != zero) {
117       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
118     }
119     const int8_t* i16 = input[16];
120     assert(i16 != NULL);
121     if XNN_UNPREDICTABLE(i16 != zero) {
122       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
123     }
124     const int8_t* i17 = input[17];
125     assert(i17 != NULL);
126     if XNN_UNPREDICTABLE(i17 != zero) {
127       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
128     }
129     const int8_t* i18 = input[18];
130     assert(i18 != NULL);
131     if XNN_UNPREDICTABLE(i18 != zero) {
132       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
133     }
134     const int8_t* i19 = input[19];
135     assert(i19 != NULL);
136     if XNN_UNPREDICTABLE(i19 != zero) {
137       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
138     }
139     const int8_t* i20 = input[20];
140     assert(i20 != NULL);
141     if XNN_UNPREDICTABLE(i20 != zero) {
142       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
143     }
144     const int8_t* i21 = input[21];
145     assert(i21 != NULL);
146     if XNN_UNPREDICTABLE(i21 != zero) {
147       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
148     }
149     const int8_t* i22 = input[22];
150     assert(i22 != NULL);
151     if XNN_UNPREDICTABLE(i22 != zero) {
152       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
153     }
154     const int8_t* i23 = input[23];
155     assert(i23 != NULL);
156     if XNN_UNPREDICTABLE(i23 != zero) {
157       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
158     }
159     const int8_t* i24 = input[24];
160     assert(i24 != NULL);
161     if XNN_UNPREDICTABLE(i24 != zero) {
162       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
163     }
164     input = (const int8_t**) ((uintptr_t) input + input_stride);
165 
166     size_t c = channels;
167     const void* w = weights;
168     for (; c >= 16; c -= 16) {
169       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
172       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
173 
174       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
175       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
176 
177       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
178       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
179 
180       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
181       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
182       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
183       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
184       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
185       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
186 
187       vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
188       vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
189 
190       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
191       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
192       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
193       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
194       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
195       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
196 
197       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
198       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
199 
200       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
201       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
202       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
203       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
204       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
205       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
206 
207       vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
208       vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
209 
210       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
211       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
212       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
213       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
214       const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
215       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
216 
217       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
218       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
219 
220       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
221       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
222       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
223       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
224       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
225       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
226 
227       vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
228       vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
229 
230       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
231       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
232       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
233       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
234       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
235       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
236 
237       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
238       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
239 
240       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
241       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
242       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
243       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
244       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
245       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
246 
247       vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
248       vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
249 
250       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
251       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
252       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
253       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
254       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
255       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
256 
257       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
258       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
259 
260       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
261       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
262       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
263       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
264       const int8x16_t vi9x0123456789ABCDEF = vld1q_s8(i9); i9 += 16;
265       const int8x16_t vk9x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
266 
267       vprod01234567 = vmull_s8(vget_low_s8(vi9x0123456789ABCDEF), vget_low_s8(vk9x0123456789ABCDEF));
268       vprod89ABCDEF = vmull_s8(vget_high_s8(vi9x0123456789ABCDEF), vget_high_s8(vk9x0123456789ABCDEF));
269 
270       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
271       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
272       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
273       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
274       const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16;
275       const int8x16_t vk10x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
276 
277       vprod01234567 = vmull_s8(vget_low_s8(vi10x0123456789ABCDEF), vget_low_s8(vk10x0123456789ABCDEF));
278       vprod89ABCDEF = vmull_s8(vget_high_s8(vi10x0123456789ABCDEF), vget_high_s8(vk10x0123456789ABCDEF));
279 
280       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
281       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
282       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
283       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
284       const int8x16_t vi11x0123456789ABCDEF = vld1q_s8(i11); i11 += 16;
285       const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
286 
287       vprod01234567 = vmull_s8(vget_low_s8(vi11x0123456789ABCDEF), vget_low_s8(vk11x0123456789ABCDEF));
288       vprod89ABCDEF = vmull_s8(vget_high_s8(vi11x0123456789ABCDEF), vget_high_s8(vk11x0123456789ABCDEF));
289 
290       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
291       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
292       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
293       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
294       const int8x16_t vi12x0123456789ABCDEF = vld1q_s8(i12); i12 += 16;
295       const int8x16_t vk12x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
296 
297       vprod01234567 = vmull_s8(vget_low_s8(vi12x0123456789ABCDEF), vget_low_s8(vk12x0123456789ABCDEF));
298       vprod89ABCDEF = vmull_s8(vget_high_s8(vi12x0123456789ABCDEF), vget_high_s8(vk12x0123456789ABCDEF));
299 
300       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
301       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
302       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
303       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
304       const int8x16_t vi13x0123456789ABCDEF = vld1q_s8(i13); i13 += 16;
305       const int8x16_t vk13x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
306 
307       vprod01234567 = vmull_s8(vget_low_s8(vi13x0123456789ABCDEF), vget_low_s8(vk13x0123456789ABCDEF));
308       vprod89ABCDEF = vmull_s8(vget_high_s8(vi13x0123456789ABCDEF), vget_high_s8(vk13x0123456789ABCDEF));
309 
310       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
311       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
312       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
313       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
314       const int8x16_t vi14x0123456789ABCDEF = vld1q_s8(i14); i14 += 16;
315       const int8x16_t vk14x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
316 
317       vprod01234567 = vmull_s8(vget_low_s8(vi14x0123456789ABCDEF), vget_low_s8(vk14x0123456789ABCDEF));
318       vprod89ABCDEF = vmull_s8(vget_high_s8(vi14x0123456789ABCDEF), vget_high_s8(vk14x0123456789ABCDEF));
319 
320       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
321       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
322       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
323       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
324       const int8x16_t vi15x0123456789ABCDEF = vld1q_s8(i15); i15 += 16;
325       const int8x16_t vk15x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
326 
327       vprod01234567 = vmull_s8(vget_low_s8(vi15x0123456789ABCDEF), vget_low_s8(vk15x0123456789ABCDEF));
328       vprod89ABCDEF = vmull_s8(vget_high_s8(vi15x0123456789ABCDEF), vget_high_s8(vk15x0123456789ABCDEF));
329 
330       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
331       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
332       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
333       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
334       const int8x16_t vi16x0123456789ABCDEF = vld1q_s8(i16); i16 += 16;
335       const int8x16_t vk16x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
336 
337       vprod01234567 = vmull_s8(vget_low_s8(vi16x0123456789ABCDEF), vget_low_s8(vk16x0123456789ABCDEF));
338       vprod89ABCDEF = vmull_s8(vget_high_s8(vi16x0123456789ABCDEF), vget_high_s8(vk16x0123456789ABCDEF));
339 
340       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
341       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
342       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
343       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
344       const int8x16_t vi17x0123456789ABCDEF = vld1q_s8(i17); i17 += 16;
345       const int8x16_t vk17x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
346 
347       vprod01234567 = vmull_s8(vget_low_s8(vi17x0123456789ABCDEF), vget_low_s8(vk17x0123456789ABCDEF));
348       vprod89ABCDEF = vmull_s8(vget_high_s8(vi17x0123456789ABCDEF), vget_high_s8(vk17x0123456789ABCDEF));
349 
350       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
351       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
352       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
353       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
354       const int8x16_t vi18x0123456789ABCDEF = vld1q_s8(i18); i18 += 16;
355       const int8x16_t vk18x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
356 
357       vprod01234567 = vmull_s8(vget_low_s8(vi18x0123456789ABCDEF), vget_low_s8(vk18x0123456789ABCDEF));
358       vprod89ABCDEF = vmull_s8(vget_high_s8(vi18x0123456789ABCDEF), vget_high_s8(vk18x0123456789ABCDEF));
359 
360       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
361       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
362       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
363       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
364       const int8x16_t vi19x0123456789ABCDEF = vld1q_s8(i19); i19 += 16;
365       const int8x16_t vk19x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
366 
367       vprod01234567 = vmull_s8(vget_low_s8(vi19x0123456789ABCDEF), vget_low_s8(vk19x0123456789ABCDEF));
368       vprod89ABCDEF = vmull_s8(vget_high_s8(vi19x0123456789ABCDEF), vget_high_s8(vk19x0123456789ABCDEF));
369 
370       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
371       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
372       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
373       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
374       const int8x16_t vi20x0123456789ABCDEF = vld1q_s8(i20); i20 += 16;
375       const int8x16_t vk20x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
376 
377       vprod01234567 = vmull_s8(vget_low_s8(vi20x0123456789ABCDEF), vget_low_s8(vk20x0123456789ABCDEF));
378       vprod89ABCDEF = vmull_s8(vget_high_s8(vi20x0123456789ABCDEF), vget_high_s8(vk20x0123456789ABCDEF));
379 
380       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
381       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
382       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
383       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
384       const int8x16_t vi21x0123456789ABCDEF = vld1q_s8(i21); i21 += 16;
385       const int8x16_t vk21x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
386 
387       vprod01234567 = vmull_s8(vget_low_s8(vi21x0123456789ABCDEF), vget_low_s8(vk21x0123456789ABCDEF));
388       vprod89ABCDEF = vmull_s8(vget_high_s8(vi21x0123456789ABCDEF), vget_high_s8(vk21x0123456789ABCDEF));
389 
390       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
391       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
392       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
393       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
394       const int8x16_t vi22x0123456789ABCDEF = vld1q_s8(i22); i22 += 16;
395       const int8x16_t vk22x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
396 
397       vprod01234567 = vmull_s8(vget_low_s8(vi22x0123456789ABCDEF), vget_low_s8(vk22x0123456789ABCDEF));
398       vprod89ABCDEF = vmull_s8(vget_high_s8(vi22x0123456789ABCDEF), vget_high_s8(vk22x0123456789ABCDEF));
399 
400       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
401       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
402       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
403       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
404       const int8x16_t vi23x0123456789ABCDEF = vld1q_s8(i23); i23 += 16;
405       const int8x16_t vk23x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
406 
407       vprod01234567 = vmull_s8(vget_low_s8(vi23x0123456789ABCDEF), vget_low_s8(vk23x0123456789ABCDEF));
408       vprod89ABCDEF = vmull_s8(vget_high_s8(vi23x0123456789ABCDEF), vget_high_s8(vk23x0123456789ABCDEF));
409 
410       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
411       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
412       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
413       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
414       const int8x16_t vi24x0123456789ABCDEF = vld1q_s8(i24); i24 += 16;
415       const int8x16_t vk24x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
416 
417       vprod01234567 = vmull_s8(vget_low_s8(vi24x0123456789ABCDEF), vget_low_s8(vk24x0123456789ABCDEF));
418       vprod89ABCDEF = vmull_s8(vget_high_s8(vi24x0123456789ABCDEF), vget_high_s8(vk24x0123456789ABCDEF));
419 
420       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
421       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
422       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
423       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
424 
425       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
426       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
427       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
428       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
429 
430       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
431       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
432       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
433       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
434 
435       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
436       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
437       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
438       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
439 
440 #if XNN_ARCH_ARM64
441       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
442       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
443 
444       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
445       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
446 
447       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
448 #else  // !XNN_ARCH_ARM64
449       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
450       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
451 
452       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
453       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
454 
455       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
456 #endif  // !XNN_ARCH_ARM64
457 
458       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
459 
460       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
461 
462       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
463     }
464     if XNN_UNLIKELY(c != 0) {
465       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
466       do {
467         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
468         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
469 
470         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
471         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
472 
473         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
474 
475         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
476         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
477         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
478         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
479 
480         vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
481 
482         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
483         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
484         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
485         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
486 
487         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
488 
489         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
490         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
491         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
492         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
493 
494         vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
495 
496         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
497         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
498         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
499         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
500 
501         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
502 
503         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
504         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
505         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
506         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
507 
508         vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
509 
510         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
511         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
512         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
513         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
514 
515         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
516 
517         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
518         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
519         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
520         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
521 
522         vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
523 
524         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
525         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
526         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
527         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
528 
529         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
530 
531         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
532         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
533         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
534         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
535 
536         vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
537 
538         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
539         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
540         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
541         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
542 
543         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
544 
545         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
546         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
547         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
548         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
549 
550         vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
551 
552         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
553         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
554         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
555         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
556 
557         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
558 
559         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
560         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
561         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
562         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
563 
564         vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
565 
566         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
567         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
568         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
569         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
570 
571         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
572 
573         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
574         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
575         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
576         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
577 
578         vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
579 
580         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
581         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
582         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
583         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
584 
585         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
586 
587         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
588         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
589         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
590         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
591 
592         vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
593 
594         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
595         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
596         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
597         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
598 
599         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
600 
601         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
602         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
603         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
604         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
605 
606         vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
607 
608         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
609         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
610         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
611         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
612 
613         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
614 
615         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
616         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
617         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
618         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
619 
620         vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
621 
622         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
623         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
624         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
625         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
626 
627         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
628 
629         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
630         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
631         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
632         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
633 
634         vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
635 
636         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
637         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
638         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
639         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
640 
641         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
642 
643         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
644         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
645 
646         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
647         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
648 
649         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
650         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
651 
652         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
653         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
654 
655 #if XNN_ARCH_ARM64
656         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
657 #else
658         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
659 #endif
660         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
661 
662         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
663         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
664         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
665 
666         if XNN_LIKELY(c >= 8) {
667           vst1_s8(output, vout01234567); output += 8;
668           c -= 8;
669         } else {
670           if (c & 4) {
671             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
672             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
673           }
674           if (c & 2) {
675             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
676             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
677           }
678           if (c & 1) {
679             vst1_lane_s8(output, vout01234567, 0); output += 1;
680           }
681           c = 0;
682         }
683       } while (c != 0);
684     }
685 
686     output = (int8_t*) ((uintptr_t) output + output_increment);
687   } while (--output_width != 0);
688 }
689