xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x25-minmax-rndnu-neon-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 
16 
xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld128(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
33   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
34   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
35   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
36   const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
37   const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
38   do {
39     const int8_t* i0 = input[0];
40     assert(i0 != NULL);
41     if XNN_UNPREDICTABLE(i0 != zero) {
42       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43     }
44     const int8_t* i1 = input[1];
45     assert(i1 != NULL);
46     if XNN_UNPREDICTABLE(i1 != zero) {
47       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48     }
49     const int8_t* i2 = input[2];
50     assert(i2 != NULL);
51     if XNN_UNPREDICTABLE(i2 != zero) {
52       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53     }
54     const int8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const int8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const int8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const int8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const int8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const int8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     const int8_t* i9 = input[9];
85     assert(i9 != NULL);
86     if XNN_UNPREDICTABLE(i9 != zero) {
87       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
88     }
89     const int8_t* i10 = input[10];
90     assert(i10 != NULL);
91     if XNN_UNPREDICTABLE(i10 != zero) {
92       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
93     }
94     const int8_t* i11 = input[11];
95     assert(i11 != NULL);
96     if XNN_UNPREDICTABLE(i11 != zero) {
97       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
98     }
99     const int8_t* i12 = input[12];
100     assert(i12 != NULL);
101     if XNN_UNPREDICTABLE(i12 != zero) {
102       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
103     }
104     const int8_t* i13 = input[13];
105     assert(i13 != NULL);
106     if XNN_UNPREDICTABLE(i13 != zero) {
107       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
108     }
109     const int8_t* i14 = input[14];
110     assert(i14 != NULL);
111     if XNN_UNPREDICTABLE(i14 != zero) {
112       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
113     }
114     const int8_t* i15 = input[15];
115     assert(i15 != NULL);
116     if XNN_UNPREDICTABLE(i15 != zero) {
117       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
118     }
119     const int8_t* i16 = input[16];
120     assert(i16 != NULL);
121     if XNN_UNPREDICTABLE(i16 != zero) {
122       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
123     }
124     const int8_t* i17 = input[17];
125     assert(i17 != NULL);
126     if XNN_UNPREDICTABLE(i17 != zero) {
127       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
128     }
129     const int8_t* i18 = input[18];
130     assert(i18 != NULL);
131     if XNN_UNPREDICTABLE(i18 != zero) {
132       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
133     }
134     const int8_t* i19 = input[19];
135     assert(i19 != NULL);
136     if XNN_UNPREDICTABLE(i19 != zero) {
137       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
138     }
139     const int8_t* i20 = input[20];
140     assert(i20 != NULL);
141     if XNN_UNPREDICTABLE(i20 != zero) {
142       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
143     }
144     const int8_t* i21 = input[21];
145     assert(i21 != NULL);
146     if XNN_UNPREDICTABLE(i21 != zero) {
147       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
148     }
149     const int8_t* i22 = input[22];
150     assert(i22 != NULL);
151     if XNN_UNPREDICTABLE(i22 != zero) {
152       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
153     }
154     const int8_t* i23 = input[23];
155     assert(i23 != NULL);
156     if XNN_UNPREDICTABLE(i23 != zero) {
157       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
158     }
159     const int8_t* i24 = input[24];
160     assert(i24 != NULL);
161     if XNN_UNPREDICTABLE(i24 != zero) {
162       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
163     }
164     input = (const int8_t**) ((uintptr_t) input + input_stride);
165 
166     size_t c = channels;
167     const void* w = weights;
168     for (; c >= 16; c -= 16) {
169       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
172       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
173 
174       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
175       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
176 
177       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
178       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
179 
180       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
181       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
182 
183       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
184       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
185 
186       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
187       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
188       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
189       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
190       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
191       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
192 
193       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
194       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
195 
196       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
197       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
198 
199       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
200       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
201 
202       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
203       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
204       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
205       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
206       const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
207       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
208 
209       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
210       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
211 
212       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
213       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
214 
215       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
216       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
217 
218       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
219       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
220       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
221       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
222       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
223       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
224 
225       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
226       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
227 
228       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
229       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
230 
231       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
232       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
233 
234       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
235       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
236       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
237       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
238       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
239       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
240 
241       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
242       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
243 
244       const int8x16_t vi9x0123456789ABCDEF = vld1q_s8(i9); i9 += 16;
245       const int8x16_t vk9x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
246 
247       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi9x0123456789ABCDEF), vget_low_s8(vk9x0123456789ABCDEF));
248       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi9x0123456789ABCDEF), vget_high_s8(vk9x0123456789ABCDEF));
249 
250       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
251       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
252       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
253       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
254       const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16;
255       const int8x16_t vk10x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
256 
257       vprod01234567 = vmull_s8(vget_low_s8(vi10x0123456789ABCDEF), vget_low_s8(vk10x0123456789ABCDEF));
258       vprod89ABCDEF = vmull_s8(vget_high_s8(vi10x0123456789ABCDEF), vget_high_s8(vk10x0123456789ABCDEF));
259 
260       const int8x16_t vi11x0123456789ABCDEF = vld1q_s8(i11); i11 += 16;
261       const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
262 
263       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi11x0123456789ABCDEF), vget_low_s8(vk11x0123456789ABCDEF));
264       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi11x0123456789ABCDEF), vget_high_s8(vk11x0123456789ABCDEF));
265 
266       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
267       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
268       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
269       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
270       const int8x16_t vi12x0123456789ABCDEF = vld1q_s8(i12); i12 += 16;
271       const int8x16_t vk12x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
272 
273       vprod01234567 = vmull_s8(vget_low_s8(vi12x0123456789ABCDEF), vget_low_s8(vk12x0123456789ABCDEF));
274       vprod89ABCDEF = vmull_s8(vget_high_s8(vi12x0123456789ABCDEF), vget_high_s8(vk12x0123456789ABCDEF));
275 
276       const int8x16_t vi13x0123456789ABCDEF = vld1q_s8(i13); i13 += 16;
277       const int8x16_t vk13x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
278 
279       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi13x0123456789ABCDEF), vget_low_s8(vk13x0123456789ABCDEF));
280       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi13x0123456789ABCDEF), vget_high_s8(vk13x0123456789ABCDEF));
281 
282       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
283       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
284       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
285       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
286       const int8x16_t vi14x0123456789ABCDEF = vld1q_s8(i14); i14 += 16;
287       const int8x16_t vk14x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
288 
289       vprod01234567 = vmull_s8(vget_low_s8(vi14x0123456789ABCDEF), vget_low_s8(vk14x0123456789ABCDEF));
290       vprod89ABCDEF = vmull_s8(vget_high_s8(vi14x0123456789ABCDEF), vget_high_s8(vk14x0123456789ABCDEF));
291 
292       const int8x16_t vi15x0123456789ABCDEF = vld1q_s8(i15); i15 += 16;
293       const int8x16_t vk15x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
294 
295       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi15x0123456789ABCDEF), vget_low_s8(vk15x0123456789ABCDEF));
296       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi15x0123456789ABCDEF), vget_high_s8(vk15x0123456789ABCDEF));
297 
298       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
299       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
300       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
301       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
302       const int8x16_t vi16x0123456789ABCDEF = vld1q_s8(i16); i16 += 16;
303       const int8x16_t vk16x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
304 
305       vprod01234567 = vmull_s8(vget_low_s8(vi16x0123456789ABCDEF), vget_low_s8(vk16x0123456789ABCDEF));
306       vprod89ABCDEF = vmull_s8(vget_high_s8(vi16x0123456789ABCDEF), vget_high_s8(vk16x0123456789ABCDEF));
307 
308       const int8x16_t vi17x0123456789ABCDEF = vld1q_s8(i17); i17 += 16;
309       const int8x16_t vk17x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
310 
311       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi17x0123456789ABCDEF), vget_low_s8(vk17x0123456789ABCDEF));
312       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi17x0123456789ABCDEF), vget_high_s8(vk17x0123456789ABCDEF));
313 
314       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
315       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
316       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
317       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
318       const int8x16_t vi18x0123456789ABCDEF = vld1q_s8(i18); i18 += 16;
319       const int8x16_t vk18x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
320 
321       vprod01234567 = vmull_s8(vget_low_s8(vi18x0123456789ABCDEF), vget_low_s8(vk18x0123456789ABCDEF));
322       vprod89ABCDEF = vmull_s8(vget_high_s8(vi18x0123456789ABCDEF), vget_high_s8(vk18x0123456789ABCDEF));
323 
324       const int8x16_t vi19x0123456789ABCDEF = vld1q_s8(i19); i19 += 16;
325       const int8x16_t vk19x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
326 
327       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi19x0123456789ABCDEF), vget_low_s8(vk19x0123456789ABCDEF));
328       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi19x0123456789ABCDEF), vget_high_s8(vk19x0123456789ABCDEF));
329 
330       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
331       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
332       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
333       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
334       const int8x16_t vi20x0123456789ABCDEF = vld1q_s8(i20); i20 += 16;
335       const int8x16_t vk20x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
336 
337       vprod01234567 = vmull_s8(vget_low_s8(vi20x0123456789ABCDEF), vget_low_s8(vk20x0123456789ABCDEF));
338       vprod89ABCDEF = vmull_s8(vget_high_s8(vi20x0123456789ABCDEF), vget_high_s8(vk20x0123456789ABCDEF));
339 
340       const int8x16_t vi21x0123456789ABCDEF = vld1q_s8(i21); i21 += 16;
341       const int8x16_t vk21x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
342 
343       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi21x0123456789ABCDEF), vget_low_s8(vk21x0123456789ABCDEF));
344       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi21x0123456789ABCDEF), vget_high_s8(vk21x0123456789ABCDEF));
345 
346       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
347       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
348       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
349       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
350       const int8x16_t vi22x0123456789ABCDEF = vld1q_s8(i22); i22 += 16;
351       const int8x16_t vk22x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
352 
353       vprod01234567 = vmull_s8(vget_low_s8(vi22x0123456789ABCDEF), vget_low_s8(vk22x0123456789ABCDEF));
354       vprod89ABCDEF = vmull_s8(vget_high_s8(vi22x0123456789ABCDEF), vget_high_s8(vk22x0123456789ABCDEF));
355 
356       const int8x16_t vi23x0123456789ABCDEF = vld1q_s8(i23); i23 += 16;
357       const int8x16_t vk23x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
358 
359       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi23x0123456789ABCDEF), vget_low_s8(vk23x0123456789ABCDEF));
360       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi23x0123456789ABCDEF), vget_high_s8(vk23x0123456789ABCDEF));
361 
362       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
363       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
364       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
365       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
366       const int8x16_t vi24x0123456789ABCDEF = vld1q_s8(i24); i24 += 16;
367       const int8x16_t vk24x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
368 
369       vprod01234567 = vmull_s8(vget_low_s8(vi24x0123456789ABCDEF), vget_low_s8(vk24x0123456789ABCDEF));
370       vprod89ABCDEF = vmull_s8(vget_high_s8(vi24x0123456789ABCDEF), vget_high_s8(vk24x0123456789ABCDEF));
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 
377       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
378       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
379       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
380       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
381 
382       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
383       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
384       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
385       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
386 
387       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
388       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
389       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
390       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
391 
392 #if XNN_ARCH_ARM64
393       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
394       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
395 
396       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
397       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
398 
399       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
400 #else  // !XNN_ARCH_ARM64
401       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
402       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
403 
404       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
405       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
406 
407       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
408 #endif  // !XNN_ARCH_ARM64
409 
410       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
411 
412       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
413 
414       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
415     }
416     if XNN_UNLIKELY(c != 0) {
417       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
418       do {
419         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
420         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
421 
422         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
423         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
424 
425         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
426 
427         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
428         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
429 
430         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
431 
432         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
433         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
434         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
435         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
436 
437         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
438 
439         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
440         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
441 
442         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
443 
444         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
445         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
446         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
447         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
448 
449         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
450 
451         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
452         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
453 
454         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
455 
456         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
457         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
458         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
459         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
460 
461         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
462 
463         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
464         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
465 
466         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
467 
468         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
469         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
470         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
471         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
472 
473         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
474 
475         const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
476         const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
477 
478         vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
479 
480         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
481         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
482         const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
483         const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
484 
485         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
486 
487         const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
488         const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
489 
490         vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
491 
492         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
493         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
494         const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
495         const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
496 
497         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
498 
499         const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
500         const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
501 
502         vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
503 
504         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
505         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
506         const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
507         const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
508 
509         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
510 
511         const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
512         const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
513 
514         vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
515 
516         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
517         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
518         const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
519         const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
520 
521         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
522 
523         const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
524         const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
525 
526         vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
527 
528         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
529         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
530         const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
531         const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
532 
533         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
534 
535         const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
536         const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
537 
538         vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
539 
540         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
541         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
542         const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
543         const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
544 
545         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
546 
547         const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
548         const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
549 
550         vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
551 
552         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
553         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
554         const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
555         const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
556 
557         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
558 
559         const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
560         const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
561 
562         vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
563 
564         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
565         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
566         const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
567         const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
568 
569         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
570 
571         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
572         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
573 
574         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
575         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
576 
577         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
578         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
579 
580         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
581         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
582 
583 #if XNN_ARCH_ARM64
584         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
585 #else
586         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
587 #endif
588         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
589 
590         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
591         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
592         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
593 
594         if XNN_LIKELY(c >= 8) {
595           vst1_s8(output, vout01234567); output += 8;
596           c -= 8;
597         } else {
598           if (c & 4) {
599             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
600             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
601           }
602           if (c & 2) {
603             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
604             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
605           }
606           if (c & 1) {
607             vst1_lane_s8(output, vout01234567, 0); output += 1;
608           }
609           c = 0;
610         }
611       } while (c != 0);
612     }
613 
614     output = (int8_t*) ((uintptr_t) output + output_increment);
615   } while (--output_width != 0);
616 }
617