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_up8x25__neonv8_mla8_ld64(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
34 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
35 const int8x8_t voutput_max = vld1_dup_s8(¶ms->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 >= 8; c -= 8) {
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
170 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
171 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
172
173 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
174
175 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
176 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
177
178 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
179
180 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
181 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
182 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
183 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
184
185 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
186
187 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
188 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
189
190 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
191
192 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
193 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
194 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
195 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
196
197 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
198
199 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
200 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
201
202 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
203
204 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
205 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
206 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
207 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
208
209 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
210
211 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
212 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
213
214 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
215
216 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
217 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
218 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
219 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
220
221 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
222
223 const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
224 const int8x8_t vk9x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
225
226 vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
227
228 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
229 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
230 const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
231 const int8x8_t vk10x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
232
233 vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
234
235 const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
236 const int8x8_t vk11x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
237
238 vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
239
240 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
241 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
242 const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
243 const int8x8_t vk12x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
244
245 vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
246
247 const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
248 const int8x8_t vk13x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
249
250 vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
251
252 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
253 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
254 const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
255 const int8x8_t vk14x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
256
257 vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
258
259 const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
260 const int8x8_t vk15x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
261
262 vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
263
264 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
265 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
266 const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
267 const int8x8_t vk16x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
268
269 vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
270
271 const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
272 const int8x8_t vk17x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
273
274 vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
275
276 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
277 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
278 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
279 const int8x8_t vk18x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
280
281 vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
282
283 const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
284 const int8x8_t vk19x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
285
286 vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
287
288 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
289 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
290 const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
291 const int8x8_t vk20x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
292
293 vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
294
295 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
296 const int8x8_t vk21x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
297
298 vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
299
300 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
301 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
302 const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
303 const int8x8_t vk22x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
304
305 vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
306
307 const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
308 const int8x8_t vk23x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
309
310 vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
311
312 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
313 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
314 const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
315 const int8x8_t vk24x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
316
317 vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
318
319 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
320 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
321
322 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
323 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
324
325 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
326 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
327
328 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
329 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
330
331 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
332 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
333
334 #if XNN_ARCH_ARM64
335 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
336
337 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
338
339 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
340 #else // !XNN_ARCH_ARM64
341 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
342
343 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
344
345 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
346 #endif // !XNN_ARCH_ARM64
347
348 vout01234567 = vmax_s8(vout01234567, voutput_min);
349
350 vout01234567 = vmin_s8(vout01234567, voutput_max);
351
352 vst1_s8(output, vout01234567); output += 8;
353 }
354 if XNN_UNLIKELY(c != 0) {
355 {
356 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
357 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
358
359 const int8x8_t vi0x01234567 = vld1_s8(i0);
360 const int8x8_t vk0x01234567 = vld1_s8(w);
361
362 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
363
364 const int8x8_t vi1x01234567 = vld1_s8(i1);
365 const int8x8_t vk1x01234567 = vld1_s8((const void*) ((const int8_t*) w + 8));
366
367 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
368
369 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
370 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
371 const int8x8_t vi2x01234567 = vld1_s8(i2);
372 const int8x8_t vk2x01234567 = vld1_s8((const void*) ((const int8_t*) w + 16));
373
374 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
375
376 const int8x8_t vi3x01234567 = vld1_s8(i3);
377 const int8x8_t vk3x01234567 = vld1_s8((const void*) ((const int8_t*) w + 24));
378
379 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
380
381 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
382 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
383 const int8x8_t vi4x01234567 = vld1_s8(i4);
384 const int8x8_t vk4x01234567 = vld1_s8((const void*) ((const int8_t*) w + 32));
385
386 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
387
388 const int8x8_t vi5x01234567 = vld1_s8(i5);
389 const int8x8_t vk5x01234567 = vld1_s8((const void*) ((const int8_t*) w + 40));
390
391 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
392
393 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
394 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
395 const int8x8_t vi6x01234567 = vld1_s8(i6);
396 const int8x8_t vk6x01234567 = vld1_s8((const void*) ((const int8_t*) w + 48));
397
398 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
399
400 const int8x8_t vi7x01234567 = vld1_s8(i7);
401 const int8x8_t vk7x01234567 = vld1_s8((const void*) ((const int8_t*) w + 56));
402
403 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
404
405 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
406 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
407 const int8x8_t vi8x01234567 = vld1_s8(i8);
408 const int8x8_t vk8x01234567 = vld1_s8((const void*) ((const int8_t*) w + 64));
409
410 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
411
412 const int8x8_t vi9x01234567 = vld1_s8(i9);
413 const int8x8_t vk9x01234567 = vld1_s8((const void*) ((const int8_t*) w + 72));
414
415 vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
416
417 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
418 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
419 const int8x8_t vi10x01234567 = vld1_s8(i10);
420 const int8x8_t vk10x01234567 = vld1_s8((const void*) ((const int8_t*) w + 80));
421
422 vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
423
424 const int8x8_t vi11x01234567 = vld1_s8(i11);
425 const int8x8_t vk11x01234567 = vld1_s8((const void*) ((const int8_t*) w + 88));
426
427 vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
428
429 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
430 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
431 const int8x8_t vi12x01234567 = vld1_s8(i12);
432 const int8x8_t vk12x01234567 = vld1_s8((const void*) ((const int8_t*) w + 96));
433
434 vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
435
436 const int8x8_t vi13x01234567 = vld1_s8(i13);
437 const int8x8_t vk13x01234567 = vld1_s8((const void*) ((const int8_t*) w + 104));
438
439 vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
440
441 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
442 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
443 const int8x8_t vi14x01234567 = vld1_s8(i14);
444 const int8x8_t vk14x01234567 = vld1_s8((const void*) ((const int8_t*) w + 112));
445
446 vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
447
448 const int8x8_t vi15x01234567 = vld1_s8(i15);
449 const int8x8_t vk15x01234567 = vld1_s8((const void*) ((const int8_t*) w + 120));
450
451 vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
452
453 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
454 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
455 const int8x8_t vi16x01234567 = vld1_s8(i16);
456 const int8x8_t vk16x01234567 = vld1_s8((const void*) ((const int8_t*) w + 128));
457
458 vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
459
460 const int8x8_t vi17x01234567 = vld1_s8(i17);
461 const int8x8_t vk17x01234567 = vld1_s8((const void*) ((const int8_t*) w + 136));
462
463 vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
464
465 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
466 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
467 const int8x8_t vi18x01234567 = vld1_s8(i18);
468 const int8x8_t vk18x01234567 = vld1_s8((const void*) ((const int8_t*) w + 144));
469
470 vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
471
472 const int8x8_t vi19x01234567 = vld1_s8(i19);
473 const int8x8_t vk19x01234567 = vld1_s8((const void*) ((const int8_t*) w + 152));
474
475 vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
476
477 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
478 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
479 const int8x8_t vi20x01234567 = vld1_s8(i20);
480 const int8x8_t vk20x01234567 = vld1_s8((const void*) ((const int8_t*) w + 160));
481
482 vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
483
484 const int8x8_t vi21x01234567 = vld1_s8(i21);
485 const int8x8_t vk21x01234567 = vld1_s8((const void*) ((const int8_t*) w + 168));
486
487 vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
488
489 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
490 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
491 const int8x8_t vi22x01234567 = vld1_s8(i22);
492 const int8x8_t vk22x01234567 = vld1_s8((const void*) ((const int8_t*) w + 176));
493
494 vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
495
496 const int8x8_t vi23x01234567 = vld1_s8(i23);
497 const int8x8_t vk23x01234567 = vld1_s8((const void*) ((const int8_t*) w + 184));
498
499 vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
500
501 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
502 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
503 const int8x8_t vi24x01234567 = vld1_s8(i24);
504 const int8x8_t vk24x01234567 = vld1_s8((const void*) ((const int8_t*) w + 192));
505
506 vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
507
508 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
509 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
510
511 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
512 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
513
514 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 200 * sizeof(int8_t)));
515 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
516 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
517 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
518
519 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
520 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
521
522 #if XNN_ARCH_ARM64
523 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
524 #else
525 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
526 #endif
527 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
528
529 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
530 vout01234567 = vmax_s8(vout01234567, voutput_min);
531 vout01234567 = vmin_s8(vout01234567, voutput_max);
532
533 if (c & 4) {
534 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
535 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
536 }
537 if (c & 2) {
538 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
539 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
540 }
541 if (c & 1) {
542 vst1_lane_s8(output, vout01234567, 0); output += 1;
543 }
544 }
545 }
546
547 output = (int8_t*) ((uintptr_t) output + output_increment);
548 } while (--output_width != 0);
549 }
550