xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x25-minmax-fp32-neonv8-mla8-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul8.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mla8_ld128(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mla8_ld128(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
34   const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     const int8_t* i3 = input[3];
53     assert(i3 != NULL);
54     if XNN_UNPREDICTABLE(i3 != zero) {
55       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56     }
57     const int8_t* i4 = input[4];
58     assert(i4 != NULL);
59     if XNN_UNPREDICTABLE(i4 != zero) {
60       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61     }
62     const int8_t* i5 = input[5];
63     assert(i5 != NULL);
64     if XNN_UNPREDICTABLE(i5 != zero) {
65       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66     }
67     const int8_t* i6 = input[6];
68     assert(i6 != NULL);
69     if XNN_UNPREDICTABLE(i6 != zero) {
70       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71     }
72     const int8_t* i7 = input[7];
73     assert(i7 != NULL);
74     if XNN_UNPREDICTABLE(i7 != zero) {
75       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76     }
77     const int8_t* i8 = input[8];
78     assert(i8 != NULL);
79     if XNN_UNPREDICTABLE(i8 != zero) {
80       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81     }
82     const int8_t* i9 = input[9];
83     assert(i9 != NULL);
84     if XNN_UNPREDICTABLE(i9 != zero) {
85       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86     }
87     const int8_t* i10 = input[10];
88     assert(i10 != NULL);
89     if XNN_UNPREDICTABLE(i10 != zero) {
90       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91     }
92     const int8_t* i11 = input[11];
93     assert(i11 != NULL);
94     if XNN_UNPREDICTABLE(i11 != zero) {
95       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96     }
97     const int8_t* i12 = input[12];
98     assert(i12 != NULL);
99     if XNN_UNPREDICTABLE(i12 != zero) {
100       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101     }
102     const int8_t* i13 = input[13];
103     assert(i13 != NULL);
104     if XNN_UNPREDICTABLE(i13 != zero) {
105       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106     }
107     const int8_t* i14 = input[14];
108     assert(i14 != NULL);
109     if XNN_UNPREDICTABLE(i14 != zero) {
110       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111     }
112     const int8_t* i15 = input[15];
113     assert(i15 != NULL);
114     if XNN_UNPREDICTABLE(i15 != zero) {
115       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116     }
117     const int8_t* i16 = input[16];
118     assert(i16 != NULL);
119     if XNN_UNPREDICTABLE(i16 != zero) {
120       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121     }
122     const int8_t* i17 = input[17];
123     assert(i17 != NULL);
124     if XNN_UNPREDICTABLE(i17 != zero) {
125       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126     }
127     const int8_t* i18 = input[18];
128     assert(i18 != NULL);
129     if XNN_UNPREDICTABLE(i18 != zero) {
130       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131     }
132     const int8_t* i19 = input[19];
133     assert(i19 != NULL);
134     if XNN_UNPREDICTABLE(i19 != zero) {
135       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136     }
137     const int8_t* i20 = input[20];
138     assert(i20 != NULL);
139     if XNN_UNPREDICTABLE(i20 != zero) {
140       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141     }
142     const int8_t* i21 = input[21];
143     assert(i21 != NULL);
144     if XNN_UNPREDICTABLE(i21 != zero) {
145       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146     }
147     const int8_t* i22 = input[22];
148     assert(i22 != NULL);
149     if XNN_UNPREDICTABLE(i22 != zero) {
150       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151     }
152     const int8_t* i23 = input[23];
153     assert(i23 != NULL);
154     if XNN_UNPREDICTABLE(i23 != zero) {
155       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156     }
157     const int8_t* i24 = input[24];
158     assert(i24 != NULL);
159     if XNN_UNPREDICTABLE(i24 != zero) {
160       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161     }
162     input = (const int8_t**) ((uintptr_t) input + input_stride);
163 
164     size_t c = channels;
165     const void* w = weights;
166     for (; c >= 16; c -= 16) {
167       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
168       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171 
172       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
173       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
174 
175       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
176       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
177 
178       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
179       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
180 
181       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
182       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
183 
184       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
185       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
186       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
187       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
188       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
189       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
190 
191       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
192       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
193 
194       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
195       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
196 
197       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
198       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
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 vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
205       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
206 
207       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
208       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
209 
210       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
211       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
212 
213       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
214       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
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 int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
221       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
222 
223       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
224       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
225 
226       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
227       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
228 
229       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
230       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
231 
232       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
233       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
234       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
235       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
236       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
237       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
238 
239       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
240       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
241 
242       const int8x16_t vi9x0123456789ABCDEF = vld1q_s8(i9); i9 += 16;
243       const int8x16_t vk9x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
244 
245       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi9x0123456789ABCDEF), vget_low_s8(vk9x0123456789ABCDEF));
246       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi9x0123456789ABCDEF), vget_high_s8(vk9x0123456789ABCDEF));
247 
248       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
249       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
250       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
251       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
252       const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16;
253       const int8x16_t vk10x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
254 
255       vprod01234567 = vmull_s8(vget_low_s8(vi10x0123456789ABCDEF), vget_low_s8(vk10x0123456789ABCDEF));
256       vprod89ABCDEF = vmull_s8(vget_high_s8(vi10x0123456789ABCDEF), vget_high_s8(vk10x0123456789ABCDEF));
257 
258       const int8x16_t vi11x0123456789ABCDEF = vld1q_s8(i11); i11 += 16;
259       const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
260 
261       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi11x0123456789ABCDEF), vget_low_s8(vk11x0123456789ABCDEF));
262       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi11x0123456789ABCDEF), vget_high_s8(vk11x0123456789ABCDEF));
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 int8x16_t vi12x0123456789ABCDEF = vld1q_s8(i12); i12 += 16;
269       const int8x16_t vk12x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
270 
271       vprod01234567 = vmull_s8(vget_low_s8(vi12x0123456789ABCDEF), vget_low_s8(vk12x0123456789ABCDEF));
272       vprod89ABCDEF = vmull_s8(vget_high_s8(vi12x0123456789ABCDEF), vget_high_s8(vk12x0123456789ABCDEF));
273 
274       const int8x16_t vi13x0123456789ABCDEF = vld1q_s8(i13); i13 += 16;
275       const int8x16_t vk13x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
276 
277       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi13x0123456789ABCDEF), vget_low_s8(vk13x0123456789ABCDEF));
278       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi13x0123456789ABCDEF), vget_high_s8(vk13x0123456789ABCDEF));
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 vi14x0123456789ABCDEF = vld1q_s8(i14); i14 += 16;
285       const int8x16_t vk14x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
286 
287       vprod01234567 = vmull_s8(vget_low_s8(vi14x0123456789ABCDEF), vget_low_s8(vk14x0123456789ABCDEF));
288       vprod89ABCDEF = vmull_s8(vget_high_s8(vi14x0123456789ABCDEF), vget_high_s8(vk14x0123456789ABCDEF));
289 
290       const int8x16_t vi15x0123456789ABCDEF = vld1q_s8(i15); i15 += 16;
291       const int8x16_t vk15x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
292 
293       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi15x0123456789ABCDEF), vget_low_s8(vk15x0123456789ABCDEF));
294       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi15x0123456789ABCDEF), vget_high_s8(vk15x0123456789ABCDEF));
295 
296       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
297       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
298       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
299       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
300       const int8x16_t vi16x0123456789ABCDEF = vld1q_s8(i16); i16 += 16;
301       const int8x16_t vk16x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
302 
303       vprod01234567 = vmull_s8(vget_low_s8(vi16x0123456789ABCDEF), vget_low_s8(vk16x0123456789ABCDEF));
304       vprod89ABCDEF = vmull_s8(vget_high_s8(vi16x0123456789ABCDEF), vget_high_s8(vk16x0123456789ABCDEF));
305 
306       const int8x16_t vi17x0123456789ABCDEF = vld1q_s8(i17); i17 += 16;
307       const int8x16_t vk17x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
308 
309       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi17x0123456789ABCDEF), vget_low_s8(vk17x0123456789ABCDEF));
310       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi17x0123456789ABCDEF), vget_high_s8(vk17x0123456789ABCDEF));
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 int8x16_t vi18x0123456789ABCDEF = vld1q_s8(i18); i18 += 16;
317       const int8x16_t vk18x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
318 
319       vprod01234567 = vmull_s8(vget_low_s8(vi18x0123456789ABCDEF), vget_low_s8(vk18x0123456789ABCDEF));
320       vprod89ABCDEF = vmull_s8(vget_high_s8(vi18x0123456789ABCDEF), vget_high_s8(vk18x0123456789ABCDEF));
321 
322       const int8x16_t vi19x0123456789ABCDEF = vld1q_s8(i19); i19 += 16;
323       const int8x16_t vk19x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
324 
325       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi19x0123456789ABCDEF), vget_low_s8(vk19x0123456789ABCDEF));
326       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi19x0123456789ABCDEF), vget_high_s8(vk19x0123456789ABCDEF));
327 
328       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
329       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
330       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
331       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
332       const int8x16_t vi20x0123456789ABCDEF = vld1q_s8(i20); i20 += 16;
333       const int8x16_t vk20x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
334 
335       vprod01234567 = vmull_s8(vget_low_s8(vi20x0123456789ABCDEF), vget_low_s8(vk20x0123456789ABCDEF));
336       vprod89ABCDEF = vmull_s8(vget_high_s8(vi20x0123456789ABCDEF), vget_high_s8(vk20x0123456789ABCDEF));
337 
338       const int8x16_t vi21x0123456789ABCDEF = vld1q_s8(i21); i21 += 16;
339       const int8x16_t vk21x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
340 
341       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi21x0123456789ABCDEF), vget_low_s8(vk21x0123456789ABCDEF));
342       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi21x0123456789ABCDEF), vget_high_s8(vk21x0123456789ABCDEF));
343 
344       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
345       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
346       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
347       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
348       const int8x16_t vi22x0123456789ABCDEF = vld1q_s8(i22); i22 += 16;
349       const int8x16_t vk22x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
350 
351       vprod01234567 = vmull_s8(vget_low_s8(vi22x0123456789ABCDEF), vget_low_s8(vk22x0123456789ABCDEF));
352       vprod89ABCDEF = vmull_s8(vget_high_s8(vi22x0123456789ABCDEF), vget_high_s8(vk22x0123456789ABCDEF));
353 
354       const int8x16_t vi23x0123456789ABCDEF = vld1q_s8(i23); i23 += 16;
355       const int8x16_t vk23x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
356 
357       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi23x0123456789ABCDEF), vget_low_s8(vk23x0123456789ABCDEF));
358       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi23x0123456789ABCDEF), vget_high_s8(vk23x0123456789ABCDEF));
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 vi24x0123456789ABCDEF = vld1q_s8(i24); i24 += 16;
365       const int8x16_t vk24x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
366 
367       vprod01234567 = vmull_s8(vget_low_s8(vi24x0123456789ABCDEF), vget_low_s8(vk24x0123456789ABCDEF));
368       vprod89ABCDEF = vmull_s8(vget_high_s8(vi24x0123456789ABCDEF), vget_high_s8(vk24x0123456789ABCDEF));
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 
375       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
376       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
377       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
378       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
379 
380       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
381       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
382       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
383       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
384 
385       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
386       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
387       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
388       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
389 
390       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
391       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
392       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
393       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
394 
395 #if XNN_ARCH_ARM64
396       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
397       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
398 
399       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
400       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
401 
402       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
403 #else  // !XNN_ARCH_ARM64
404       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
405       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
406 
407       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
408       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
409 
410       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
411 #endif  // !XNN_ARCH_ARM64
412 
413       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
414 
415       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
416 
417       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
418     }
419     if XNN_UNLIKELY(c != 0) {
420       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
421       do {
422         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
423         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
424 
425         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
426         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
427 
428         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
429 
430         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
431         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
432 
433         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
434 
435         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
436         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
437         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
438         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
439 
440         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
441 
442         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
443         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
444 
445         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
446 
447         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
448         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
449         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
450         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
451 
452         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
453 
454         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
455         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
456 
457         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
458 
459         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
460         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
461         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
462         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
463 
464         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
465 
466         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
467         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
468 
469         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
470 
471         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
472         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
473         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
474         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
475 
476         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
477 
478         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
479         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
480 
481         vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
482 
483         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
484         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
485         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
486         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
487 
488         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
489 
490         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
491         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
492 
493         vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
494 
495         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
496         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
497         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
498         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
499 
500         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
501 
502         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
503         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
504 
505         vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
506 
507         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
508         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
509         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
510         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
511 
512         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
513 
514         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
515         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
516 
517         vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
518 
519         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
520         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
521         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
522         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
523 
524         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
525 
526         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
527         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
528 
529         vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
530 
531         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
532         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
533         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
534         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
535 
536         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
537 
538         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
539         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
540 
541         vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
542 
543         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
544         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
545         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
546         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
547 
548         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
549 
550         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
551         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
552 
553         vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
554 
555         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
556         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
557         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
558         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
559 
560         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
561 
562         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
563         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
564 
565         vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
566 
567         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
568         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
569         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
570         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
571 
572         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
573 
574         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
575         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
576 
577         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
578         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
579 
580         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t)));
581         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
582         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
583         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
584 
585         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
586         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
587 
588 #if XNN_ARCH_ARM64
589         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
590 #else
591         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
592 #endif
593         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
594 
595         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
596         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
597         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
598 
599         if XNN_LIKELY(c >= 8) {
600           vst1_s8(output, vout01234567); output += 8;
601           c -= 8;
602         } else {
603           if (c & 4) {
604             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
605             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
606           }
607           if (c & 2) {
608             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
609             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
610           }
611           if (c & 1) {
612             vst1_lane_s8(output, vout01234567, 0); output += 1;
613           }
614           c = 0;
615         }
616       } while (c != 0);
617     }
618 
619     output = (int8_t*) ((uintptr_t) output + output_increment);
620   } while (--output_width != 0);
621 }
622