xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x25-minmax-fp32-neonv8-mul8-ld64.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_ld64(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_ld64(
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 int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
173       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
174       const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
175       const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176 
177       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
178       int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
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 int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
185       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
186       const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
187       const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
188 
189       vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
190       vprod89ABCDEF = vmull_s8(vi1x89ABCDEF, vk1x89ABCDEF);
191 
192       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
193       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
194       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
195       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
196       const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
197       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
198       const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
199       const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
200 
201       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
202       vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
203 
204       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
205       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
206       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
207       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
208       const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
209       const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
210       const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
211       const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
212 
213       vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
214       vprod89ABCDEF = vmull_s8(vi3x89ABCDEF, vk3x89ABCDEF);
215 
216       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
217       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
218       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
219       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
220       const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
221       const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
222       const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
223       const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
224 
225       vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
226       vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
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 int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
233       const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
234       const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
235       const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
236 
237       vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
238       vprod89ABCDEF = vmull_s8(vi5x89ABCDEF, vk5x89ABCDEF);
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 int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
245       const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
246       const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
247       const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
248 
249       vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
250       vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
251 
252       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
253       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
254       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
255       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
256       const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
257       const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
258       const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
259       const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
260 
261       vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
262       vprod89ABCDEF = vmull_s8(vi7x89ABCDEF, vk7x89ABCDEF);
263 
264       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
265       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
266       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
267       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
268       const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
269       const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
270       const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
271       const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
272 
273       vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
274       vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
275 
276       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
277       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
278       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
279       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
280       const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
281       const int8x8_t vk9x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
282       const int8x8_t vi9x89ABCDEF = vld1_s8(i9); i9 += 8;
283       const int8x8_t vk9x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
284 
285       vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
286       vprod89ABCDEF = vmull_s8(vi9x89ABCDEF, vk9x89ABCDEF);
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 int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
293       const int8x8_t vk10x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
294       const int8x8_t vi10x89ABCDEF = vld1_s8(i10); i10 += 8;
295       const int8x8_t vk10x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
296 
297       vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
298       vprod89ABCDEF = vmull_s8(vi10x89ABCDEF, vk10x89ABCDEF);
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 int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
305       const int8x8_t vk11x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
306       const int8x8_t vi11x89ABCDEF = vld1_s8(i11); i11 += 8;
307       const int8x8_t vk11x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
308 
309       vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
310       vprod89ABCDEF = vmull_s8(vi11x89ABCDEF, vk11x89ABCDEF);
311 
312       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
313       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
314       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
315       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
316       const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
317       const int8x8_t vk12x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
318       const int8x8_t vi12x89ABCDEF = vld1_s8(i12); i12 += 8;
319       const int8x8_t vk12x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
320 
321       vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
322       vprod89ABCDEF = vmull_s8(vi12x89ABCDEF, vk12x89ABCDEF);
323 
324       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
325       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
326       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
327       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
328       const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
329       const int8x8_t vk13x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
330       const int8x8_t vi13x89ABCDEF = vld1_s8(i13); i13 += 8;
331       const int8x8_t vk13x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
332 
333       vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
334       vprod89ABCDEF = vmull_s8(vi13x89ABCDEF, vk13x89ABCDEF);
335 
336       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
337       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
338       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
339       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
340       const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
341       const int8x8_t vk14x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
342       const int8x8_t vi14x89ABCDEF = vld1_s8(i14); i14 += 8;
343       const int8x8_t vk14x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
344 
345       vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
346       vprod89ABCDEF = vmull_s8(vi14x89ABCDEF, vk14x89ABCDEF);
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 int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
353       const int8x8_t vk15x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
354       const int8x8_t vi15x89ABCDEF = vld1_s8(i15); i15 += 8;
355       const int8x8_t vk15x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
356 
357       vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
358       vprod89ABCDEF = vmull_s8(vi15x89ABCDEF, vk15x89ABCDEF);
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 int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
365       const int8x8_t vk16x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
366       const int8x8_t vi16x89ABCDEF = vld1_s8(i16); i16 += 8;
367       const int8x8_t vk16x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
368 
369       vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
370       vprod89ABCDEF = vmull_s8(vi16x89ABCDEF, vk16x89ABCDEF);
371 
372       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
373       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
374       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
375       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
376       const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
377       const int8x8_t vk17x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
378       const int8x8_t vi17x89ABCDEF = vld1_s8(i17); i17 += 8;
379       const int8x8_t vk17x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
380 
381       vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
382       vprod89ABCDEF = vmull_s8(vi17x89ABCDEF, vk17x89ABCDEF);
383 
384       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
385       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
386       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
387       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
388       const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
389       const int8x8_t vk18x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
390       const int8x8_t vi18x89ABCDEF = vld1_s8(i18); i18 += 8;
391       const int8x8_t vk18x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
392 
393       vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
394       vprod89ABCDEF = vmull_s8(vi18x89ABCDEF, vk18x89ABCDEF);
395 
396       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
397       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
398       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
399       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
400       const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
401       const int8x8_t vk19x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
402       const int8x8_t vi19x89ABCDEF = vld1_s8(i19); i19 += 8;
403       const int8x8_t vk19x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
404 
405       vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
406       vprod89ABCDEF = vmull_s8(vi19x89ABCDEF, vk19x89ABCDEF);
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 int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
413       const int8x8_t vk20x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
414       const int8x8_t vi20x89ABCDEF = vld1_s8(i20); i20 += 8;
415       const int8x8_t vk20x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
416 
417       vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
418       vprod89ABCDEF = vmull_s8(vi20x89ABCDEF, vk20x89ABCDEF);
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       const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
425       const int8x8_t vk21x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
426       const int8x8_t vi21x89ABCDEF = vld1_s8(i21); i21 += 8;
427       const int8x8_t vk21x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
428 
429       vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
430       vprod89ABCDEF = vmull_s8(vi21x89ABCDEF, vk21x89ABCDEF);
431 
432       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
433       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
434       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
435       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
436       const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
437       const int8x8_t vk22x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
438       const int8x8_t vi22x89ABCDEF = vld1_s8(i22); i22 += 8;
439       const int8x8_t vk22x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
440 
441       vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
442       vprod89ABCDEF = vmull_s8(vi22x89ABCDEF, vk22x89ABCDEF);
443 
444       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
445       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
446       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
447       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
448       const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
449       const int8x8_t vk23x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
450       const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8;
451       const int8x8_t vk23x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
452 
453       vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
454       vprod89ABCDEF = vmull_s8(vi23x89ABCDEF, vk23x89ABCDEF);
455 
456       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
457       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
458       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
459       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
460       const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
461       const int8x8_t vk24x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
462       const int8x8_t vi24x89ABCDEF = vld1_s8(i24); i24 += 8;
463       const int8x8_t vk24x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
464 
465       vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
466       vprod89ABCDEF = vmull_s8(vi24x89ABCDEF, vk24x89ABCDEF);
467 
468       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
469       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
470       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
471       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
472 
473       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
474       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
475       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
476       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
477 
478       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
479       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
480       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
481       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
482 
483       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
484       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
485       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
486       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
487 
488       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
489       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
490       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
491       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
492 
493 #if XNN_ARCH_ARM64
494       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
495       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
496 
497       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
498       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
499 
500       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
501 #else  // !XNN_ARCH_ARM64
502       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
503       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
504 
505       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
506       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
507 
508       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
509 #endif  // !XNN_ARCH_ARM64
510 
511       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
512 
513       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
514 
515       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
516     }
517     if XNN_UNLIKELY(c != 0) {
518       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
519       do {
520         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
521         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
522 
523         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
524         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
525 
526         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
527 
528         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
529         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
530         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
531         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
532 
533         vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
534 
535         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
536         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
537         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
538         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
539 
540         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
541 
542         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
543         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
544         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
545         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
546 
547         vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
548 
549         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
550         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
551         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
552         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
553 
554         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
555 
556         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
557         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
558         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
559         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
560 
561         vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
562 
563         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
564         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
565         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
566         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
567 
568         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
569 
570         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
571         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
572         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
573         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
574 
575         vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
576 
577         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
578         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
579         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
580         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
581 
582         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
583 
584         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
585         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
586         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
587         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
588 
589         vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
590 
591         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
592         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
593         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
594         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
595 
596         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
597 
598         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
599         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
600         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
601         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
602 
603         vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
604 
605         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
606         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
607         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
608         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
609 
610         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
611 
612         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
613         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
614         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
615         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
616 
617         vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
618 
619         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
620         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
621         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
622         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
623 
624         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
625 
626         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
627         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
628         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
629         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
630 
631         vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
632 
633         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
634         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
635         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
636         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
637 
638         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
639 
640         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
641         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
642         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
643         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
644 
645         vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
646 
647         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
648         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
649         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
650         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
651 
652         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
653 
654         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
655         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
656         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
657         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
658 
659         vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
660 
661         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
662         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
663         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
664         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
665 
666         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
667 
668         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
669         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
670         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
671         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
672 
673         vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
674 
675         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
676         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
677         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
678         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
679 
680         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
681 
682         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
683         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
684         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
685         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
686 
687         vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
688 
689         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
690         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
691         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
692         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
693 
694         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
695 
696         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
697         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
698 
699         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
700         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
701 
702         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t)));
703         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
704         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
705         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
706 
707         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
708         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
709 
710 #if XNN_ARCH_ARM64
711         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
712 #else
713         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
714 #endif
715         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
716 
717         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
718         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
719         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
720 
721         if XNN_LIKELY(c >= 8) {
722           vst1_s8(output, vout01234567); output += 8;
723           c -= 8;
724         } else {
725           if (c & 4) {
726             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
727             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
728           }
729           if (c & 2) {
730             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
731             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
732           }
733           if (c & 1) {
734             vst1_lane_s8(output, vout01234567, 0); output += 1;
735           }
736           c = 0;
737         }
738       } while (c != 0);
739     }
740 
741     output = (int8_t*) ((uintptr_t) output + output_increment);
742   } while (--output_width != 0);
743 }
744