xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x25-minmax-rndnu-neon-mla8-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 
16 
xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld64(
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 int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
175       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176       const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
177       const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178 
179       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
180       int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
181 
182       const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
183       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
184       const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
185       const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
186 
187       vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
188       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi1x89ABCDEF, vk1x89ABCDEF);
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 int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
195       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
196       const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
197       const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
198 
199       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
200       vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
201 
202       const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
203       const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
204       const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
205       const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
206 
207       vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
208       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi3x89ABCDEF, vk3x89ABCDEF);
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 int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
215       const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
216       const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
217       const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
218 
219       vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
220       vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
221 
222       const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
223       const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
224       const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
225       const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
226 
227       vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
228       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi5x89ABCDEF, vk5x89ABCDEF);
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 int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
235       const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
236       const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
237       const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
238 
239       vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
240       vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
241 
242       const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
243       const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
244       const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
245       const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
246 
247       vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
248       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi7x89ABCDEF, vk7x89ABCDEF);
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 int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
255       const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
256       const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
257       const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
258 
259       vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
260       vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
261 
262       const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
263       const int8x8_t vk9x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
264       const int8x8_t vi9x89ABCDEF = vld1_s8(i9); i9 += 8;
265       const int8x8_t vk9x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
266 
267       vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
268       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi9x89ABCDEF, vk9x89ABCDEF);
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 int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
275       const int8x8_t vk10x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
276       const int8x8_t vi10x89ABCDEF = vld1_s8(i10); i10 += 8;
277       const int8x8_t vk10x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
278 
279       vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
280       vprod89ABCDEF = vmull_s8(vi10x89ABCDEF, vk10x89ABCDEF);
281 
282       const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
283       const int8x8_t vk11x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
284       const int8x8_t vi11x89ABCDEF = vld1_s8(i11); i11 += 8;
285       const int8x8_t vk11x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
286 
287       vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
288       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi11x89ABCDEF, vk11x89ABCDEF);
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 int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
295       const int8x8_t vk12x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
296       const int8x8_t vi12x89ABCDEF = vld1_s8(i12); i12 += 8;
297       const int8x8_t vk12x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
298 
299       vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
300       vprod89ABCDEF = vmull_s8(vi12x89ABCDEF, vk12x89ABCDEF);
301 
302       const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
303       const int8x8_t vk13x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
304       const int8x8_t vi13x89ABCDEF = vld1_s8(i13); i13 += 8;
305       const int8x8_t vk13x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
306 
307       vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
308       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi13x89ABCDEF, vk13x89ABCDEF);
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 int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
315       const int8x8_t vk14x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
316       const int8x8_t vi14x89ABCDEF = vld1_s8(i14); i14 += 8;
317       const int8x8_t vk14x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
318 
319       vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
320       vprod89ABCDEF = vmull_s8(vi14x89ABCDEF, vk14x89ABCDEF);
321 
322       const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
323       const int8x8_t vk15x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
324       const int8x8_t vi15x89ABCDEF = vld1_s8(i15); i15 += 8;
325       const int8x8_t vk15x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
326 
327       vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
328       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi15x89ABCDEF, vk15x89ABCDEF);
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 int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
335       const int8x8_t vk16x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
336       const int8x8_t vi16x89ABCDEF = vld1_s8(i16); i16 += 8;
337       const int8x8_t vk16x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
338 
339       vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
340       vprod89ABCDEF = vmull_s8(vi16x89ABCDEF, vk16x89ABCDEF);
341 
342       const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
343       const int8x8_t vk17x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
344       const int8x8_t vi17x89ABCDEF = vld1_s8(i17); i17 += 8;
345       const int8x8_t vk17x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
346 
347       vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
348       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi17x89ABCDEF, vk17x89ABCDEF);
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 int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
355       const int8x8_t vk18x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
356       const int8x8_t vi18x89ABCDEF = vld1_s8(i18); i18 += 8;
357       const int8x8_t vk18x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
358 
359       vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
360       vprod89ABCDEF = vmull_s8(vi18x89ABCDEF, vk18x89ABCDEF);
361 
362       const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
363       const int8x8_t vk19x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
364       const int8x8_t vi19x89ABCDEF = vld1_s8(i19); i19 += 8;
365       const int8x8_t vk19x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
366 
367       vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
368       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi19x89ABCDEF, vk19x89ABCDEF);
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 int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
375       const int8x8_t vk20x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
376       const int8x8_t vi20x89ABCDEF = vld1_s8(i20); i20 += 8;
377       const int8x8_t vk20x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
378 
379       vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
380       vprod89ABCDEF = vmull_s8(vi20x89ABCDEF, vk20x89ABCDEF);
381 
382       const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
383       const int8x8_t vk21x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
384       const int8x8_t vi21x89ABCDEF = vld1_s8(i21); i21 += 8;
385       const int8x8_t vk21x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
386 
387       vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
388       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi21x89ABCDEF, vk21x89ABCDEF);
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 int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
395       const int8x8_t vk22x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
396       const int8x8_t vi22x89ABCDEF = vld1_s8(i22); i22 += 8;
397       const int8x8_t vk22x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
398 
399       vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
400       vprod89ABCDEF = vmull_s8(vi22x89ABCDEF, vk22x89ABCDEF);
401 
402       const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
403       const int8x8_t vk23x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
404       const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8;
405       const int8x8_t vk23x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
406 
407       vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
408       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi23x89ABCDEF, vk23x89ABCDEF);
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 int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
415       const int8x8_t vk24x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
416       const int8x8_t vi24x89ABCDEF = vld1_s8(i24); i24 += 8;
417       const int8x8_t vk24x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
418 
419       vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
420       vprod89ABCDEF = vmull_s8(vi24x89ABCDEF, vk24x89ABCDEF);
421 
422       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
423       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
424       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
425       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
426 
427       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
428       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
429       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
430       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
431 
432       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
433       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
434       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
435       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
436 
437       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
438       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
439       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
440       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
441 
442 #if XNN_ARCH_ARM64
443       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
444       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
445 
446       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
447       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
448 
449       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
450 #else  // !XNN_ARCH_ARM64
451       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
452       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
453 
454       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
455       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
456 
457       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
458 #endif  // !XNN_ARCH_ARM64
459 
460       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
461 
462       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
463 
464       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
465     }
466     if XNN_UNLIKELY(c != 0) {
467       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
468       do {
469         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
470         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
471 
472         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
473         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
474 
475         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
476 
477         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
478         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
479 
480         vprod01234567 = vmlal_s8(vprod01234567, 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         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
490         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
491 
492         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
493 
494         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
495         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
496         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
497         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
498 
499         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
500 
501         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
502         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
503 
504         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
505 
506         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
507         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
508         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
509         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
510 
511         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
512 
513         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
514         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
515 
516         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
517 
518         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
519         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
520         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
521         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
522 
523         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
524 
525         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
526         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
527 
528         vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
529 
530         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
531         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
532         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
533         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
534 
535         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
536 
537         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
538         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
539 
540         vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
541 
542         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
543         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
544         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
545         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
546 
547         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
548 
549         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
550         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
551 
552         vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
553 
554         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
555         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
556         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
557         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
558 
559         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
560 
561         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
562         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
563 
564         vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
565 
566         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
567         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
568         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
569         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
570 
571         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
572 
573         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
574         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
575 
576         vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
577 
578         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
579         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
580         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
581         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
582 
583         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
584 
585         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
586         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
587 
588         vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
589 
590         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
591         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
592         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
593         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
594 
595         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
596 
597         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
598         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
599 
600         vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
601 
602         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
603         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
604         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
605         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
606 
607         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
608 
609         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
610         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
611 
612         vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
613 
614         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
615         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
616         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
617         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
618 
619         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
620 
621         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
622         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
623 
624         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
625         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
626 
627         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
628         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
629 
630         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
631         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
632 
633 #if XNN_ARCH_ARM64
634         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
635 #else
636         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
637 #endif
638         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
639 
640         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
641         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
642         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
643 
644         if XNN_LIKELY(c >= 8) {
645           vst1_s8(output, vout01234567); output += 8;
646           c -= 8;
647         } else {
648           if (c & 4) {
649             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
650             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
651           }
652           if (c & 2) {
653             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
654             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
655           }
656           if (c & 1) {
657             vst1_lane_s8(output, vout01234567, 0); output += 1;
658           }
659           c = 0;
660         }
661       } while (c != 0);
662     }
663 
664     output = (int8_t*) ((uintptr_t) output + output_increment);
665   } while (--output_width != 0);
666 }
667