Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c8s8r-minmax-rndnu-neon-mla8-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-neon-mul8.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_dwconv_minmax_rndnu_ukernel_8f8m9l16c8s8r__neon_mla8_ld128(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 8);
36
37 const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
38 const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
39 const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
40 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
41 const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
42 const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
43 do {
44 const void* w = weights;
45
46 // First pass to process 8 inputs.
47 {
48 int32_t* b = buffer;
49 const int8_t* i0 = input[0];
50 assert(i0 != NULL);
51 if XNN_UNPREDICTABLE(i0 != zero) {
52 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
53 }
54 const int8_t* i1 = input[1];
55 assert(i1 != NULL);
56 if XNN_UNPREDICTABLE(i1 != zero) {
57 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
58 }
59 const int8_t* i2 = input[2];
60 assert(i2 != NULL);
61 if XNN_UNPREDICTABLE(i2 != zero) {
62 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
63 }
64 const int8_t* i3 = input[3];
65 assert(i3 != NULL);
66 if XNN_UNPREDICTABLE(i3 != zero) {
67 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
68 }
69 const int8_t* i4 = input[4];
70 assert(i4 != NULL);
71 if XNN_UNPREDICTABLE(i4 != zero) {
72 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
73 }
74 const int8_t* i5 = input[5];
75 assert(i5 != NULL);
76 if XNN_UNPREDICTABLE(i5 != zero) {
77 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
78 }
79 const int8_t* i6 = input[6];
80 assert(i6 != NULL);
81 if XNN_UNPREDICTABLE(i6 != zero) {
82 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
83 }
84 const int8_t* i7 = input[7];
85 assert(i7 != NULL);
86 if XNN_UNPREDICTABLE(i7 != zero) {
87 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
88 }
89 input += 8;
90
91 size_t c = round_up_po2(channels, 8);
92
93 for (; c >= 16; c -= 16) {
94 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
95 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
96 int32x4_t vacc89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
97 int32x4_t vaccCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
98
99 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
100 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101
102 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
103 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
104
105
106 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
107 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108
109 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
110 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
111
112 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
113 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
114 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
115 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
116
117 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
118 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119
120 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
121 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
122
123
124 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
125 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126
127 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
128 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
129
130 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
131 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
132 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
133 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
134
135 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
136 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
137
138 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
139 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
140
141
142 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
143 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
144
145 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
146 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
147
148 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
149 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
150 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
151 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
152
153 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
154 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
155
156 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
157 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
158
159
160 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
161 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162
163 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
164 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
165
166 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
167 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
168 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
169 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
170
171 vst1q_s32(b, vacc0123); b += 4;
172 vst1q_s32(b, vacc4567); b += 4;
173 vst1q_s32(b, vacc89AB); b += 4;
174 vst1q_s32(b, vaccCDEF); b += 4;
175 }
176
177 if XNN_UNLIKELY(c != 0) {
178 do {
179 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
180 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
181
182 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
183 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
184
185 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
186
187
188 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
189 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
190
191 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
192
193 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
194 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
195
196 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
197 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
198
199 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
200
201
202 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
203 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
204
205 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
206
207 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
208 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
209
210 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
211 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
212
213 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
214
215
216 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
217 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
218
219 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
220
221 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
222 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
223
224 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
225 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
226
227 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
228
229
230 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
231 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
232
233 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
234
235 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
236 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
237
238 vst1q_s32(b, vacc0123); b += 4;
239 vst1q_s32(b, vacc4567); b += 4;
240 c -= 8;
241 } while (c != 0);
242 }
243 }
244
245 // Middle pass to process 8 inputs in each iteration.
246 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
247 int32_t* b = buffer;
248
249 const int8_t* i0 = input[0];
250 assert(i0 != NULL);
251 if XNN_UNPREDICTABLE(i0 != zero) {
252 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
253 }
254 const int8_t* i1 = input[1];
255 assert(i1 != NULL);
256 if XNN_UNPREDICTABLE(i1 != zero) {
257 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
258 }
259 const int8_t* i2 = input[2];
260 assert(i2 != NULL);
261 if XNN_UNPREDICTABLE(i2 != zero) {
262 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
263 }
264 const int8_t* i3 = input[3];
265 assert(i3 != NULL);
266 if XNN_UNPREDICTABLE(i3 != zero) {
267 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
268 }
269 const int8_t* i4 = input[4];
270 assert(i4 != NULL);
271 if XNN_UNPREDICTABLE(i4 != zero) {
272 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
273 }
274 const int8_t* i5 = input[5];
275 assert(i5 != NULL);
276 if XNN_UNPREDICTABLE(i5 != zero) {
277 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
278 }
279 const int8_t* i6 = input[6];
280 assert(i6 != NULL);
281 if XNN_UNPREDICTABLE(i6 != zero) {
282 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
283 }
284 const int8_t* i7 = input[7];
285 assert(i7 != NULL);
286 if XNN_UNPREDICTABLE(i7 != zero) {
287 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
288 }
289 input += 8;
290
291 size_t c = round_up_po2(channels, 8);
292
293 for (; c >= 16; c -= 16) {
294 int32x4_t vacc0123 = vld1q_s32(b);
295 int32x4_t vacc4567 = vld1q_s32(b + 4);
296 int32x4_t vacc89AB = vld1q_s32(b + 8);
297 int32x4_t vaccCDEF = vld1q_s32(b + 12);
298
299 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
300 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
301
302 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
303 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
304
305
306 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
307 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
308
309 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
310 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
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
317 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
318 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
319
320 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
321 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
322
323
324 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
325 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
326
327 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
328 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
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
335 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
336 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
337
338 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
339 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
340
341
342 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
343 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
344
345 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
346 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
347
348 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
349 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
350 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
351 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
352
353 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
354 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
355
356 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
357 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
358
359
360 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
361 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
362
363 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
364 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
365
366 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
367 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
368 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
369 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
370
371 vst1q_s32(b, vacc0123); b += 4;
372 vst1q_s32(b, vacc4567); b += 4;
373 vst1q_s32(b, vacc89AB); b += 4;
374 vst1q_s32(b, vaccCDEF); b += 4;
375 }
376
377 if XNN_UNLIKELY(c != 0) {
378 do {
379 int32x4_t vacc0123 = vld1q_s32(b);
380 int32x4_t vacc4567 = vld1q_s32(b + 4);
381
382 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
383 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
384
385 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
386
387
388 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
389 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
390
391 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
392
393 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
394 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
395
396 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
397 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
398
399 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
400
401
402 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
403 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
404
405 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
406
407 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
408 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
409
410 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
411 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
412
413 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
414
415
416 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
417 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
418
419 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
420
421 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
422 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
423
424 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
425 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
426
427 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
428
429
430 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
431 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
432
433 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
434
435 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
436 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
437
438 vst1q_s32(b, vacc0123); b += 4;
439 vst1q_s32(b, vacc4567); b += 4;
440 c -= 8;
441 } while (c != 0);
442 }
443 }
444
445 // Last pass to process up to 9 inputs.
446 {
447 const int32_t* b = buffer;
448 const int8_t* i0 = input[0];
449 assert(i0 != NULL);
450 if XNN_UNPREDICTABLE(i0 != zero) {
451 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
452 }
453 const int8_t* i1 = input[1];
454 assert(i1 != NULL);
455 if XNN_UNPREDICTABLE(i1 != zero) {
456 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
457 }
458 const int8_t* i2 = input[2];
459 assert(i2 != NULL);
460 if XNN_UNPREDICTABLE(i2 != zero) {
461 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
462 }
463 const int8_t* i3 = input[3];
464 assert(i3 != NULL);
465 if XNN_UNPREDICTABLE(i3 != zero) {
466 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
467 }
468 const int8_t* i4 = input[4];
469 assert(i4 != NULL);
470 if XNN_UNPREDICTABLE(i4 != zero) {
471 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
472 }
473 const int8_t* i5 = input[5];
474 assert(i5 != NULL);
475 if XNN_UNPREDICTABLE(i5 != zero) {
476 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
477 }
478 const int8_t* i6 = input[6];
479 assert(i6 != NULL);
480 if XNN_UNPREDICTABLE(i6 != zero) {
481 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
482 }
483 const int8_t* i7 = input[7];
484 assert(i7 != NULL);
485 if XNN_UNPREDICTABLE(i7 != zero) {
486 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
487 }
488 const int8_t* i8 = input[8];
489 assert(i8 != NULL);
490 if XNN_UNPREDICTABLE(i8 != zero) {
491 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
492 }
493
494 size_t c = channels;
495 for (; c >= 16; c -= 16) {
496 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
497 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
498 int32x4_t vacc89AB = vld1q_s32(b); b += 4;
499 int32x4_t vaccCDEF = vld1q_s32(b); b += 4;
500
501 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
502 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
503
504 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
505 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
506
507
508 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
509 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
510
511 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
512 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
513
514 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
515 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
516 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
517 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
518
519 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
520 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
521
522 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
523 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
524
525
526 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
527 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
528
529 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
530 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
531
532 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
533 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
534 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
535 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
536
537 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
538 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
539
540 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
541 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
542
543
544 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
545 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
546
547 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
548 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
549
550 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
551 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
552 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
553 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
554
555 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
556 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
557
558 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
559 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
560
561
562 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
563 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
564
565 vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
566 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
567
568 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
569 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
570 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
571 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
572
573 const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
574 const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
575
576 vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
577 vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
578
579 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
580 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
581 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
582 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
583
584 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
585 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
586 vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
587 vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
588
589 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
590 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
591 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
592 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
593
594 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
595 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
596 vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
597 vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
598
599 #if XNN_ARCH_ARM64
600 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
601 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
602
603 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
604 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
605
606 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
607 #else // !XNN_ARCH_ARM64
608 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
609 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
610
611 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
612 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
613
614 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
615 #endif // !XNN_ARCH_ARM64
616
617 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
618
619 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
620
621 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
622 }
623
624
625 if XNN_UNLIKELY(c != 0) {
626 do {
627 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
628 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
629
630 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
631 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
632
633 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
634
635 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
636 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
637
638 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
639
640 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
641 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
642 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
643 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
644
645 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
646
647 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
648 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
649
650 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
651
652 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
653 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
654 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
655 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
656
657 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
658
659 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
660 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
661
662 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
663
664 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
665 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
666 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
667 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
668
669 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
670
671 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
672 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
673
674 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
675
676 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
677 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
678 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
679 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
680
681 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
682
683 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
684 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
685
686 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
687 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
688
689 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
690 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
691
692 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
693 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
694
695 #if XNN_ARCH_ARM64
696 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
697 #else
698 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
699 #endif
700 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
701
702 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
703 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
704 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
705
706 if XNN_LIKELY(c >= 8) {
707 vst1_s8(output, vout01234567); output += 8;
708 c -= 8;
709 } else {
710 if (c & 4) {
711 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
712 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
713 }
714 if (c & 2) {
715 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
716 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
717 }
718 if (c & 1) {
719 vst1_lane_s8(output, vout01234567, 0); output += 1;
720 }
721 c = 0;
722 }
723
724 } while (c != 0);
725 }
726 }
727
728 input = (const int8_t**) ((uintptr_t) input + input_stride);
729 output = (int8_t*) ((uintptr_t) output + output_increment);
730 } while (--output_width != 0);
731 }