1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-neon-mul16.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_up8x25__neon_mul16(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_up8x25__neon_mul16(
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(¶ms->rndnu_neon.right_pre_shift);
33 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
34 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
35 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
36 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
37 const int8x8_t voutput_max = vld1_dup_s8(¶ms->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 >= 8; c -= 8) {
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
172
173 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
174 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
175
176 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
177 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
178
179 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
180 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
181
182 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
183 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
184
185 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
186 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
187
188 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
189 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
190
191 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
192 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
193
194 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
195 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
196
197 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
198 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
199
200 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
201 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
202
203 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
204 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
205
206 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
207 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
208
209 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
210 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
211
212 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
213 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
214
215 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
216 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
217
218 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
219 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
220
221 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
222 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
223
224 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
225 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
226
227 const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
228 const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
229
230 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
231 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
232
233 const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
234 const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
235
236 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
237 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
238
239 const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
240 const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
241
242 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
243 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
244
245 const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
246 const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
247
248 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
249 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
250
251 const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
252 const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
253
254 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
255 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
256
257 const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
258 const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
259
260 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
261 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
262
263 const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
264 const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
265
266 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
267 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
268
269 const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
270 const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
271
272 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
273 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
274
275 const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
276 const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
277
278 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
279 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
280
281 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
282 const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
283
284 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
285 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
286
287 const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
288 const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
289
290 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
291 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
292
293 const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
294 const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
295
296 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
297 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
298
299 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
300 const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
301
302 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
303 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
304
305 const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
306 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
307
308 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
309 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
310
311 const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
312 const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
313
314 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
315 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
316
317 const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
318 const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
319
320 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
321 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
322
323 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
324 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
325
326 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
327 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
328
329 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
330 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
331
332 #if XNN_ARCH_ARM64
333 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
334
335 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
336
337 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
338 #else // !XNN_ARCH_ARM64
339 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
340
341 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
342
343 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
344 #endif // !XNN_ARCH_ARM64
345
346 vout01234567 = vmax_s8(vout01234567, voutput_min);
347
348 vout01234567 = vmin_s8(vout01234567, voutput_max);
349
350 vst1_s8(output, vout01234567); output += 8;
351 }
352 if XNN_UNLIKELY(c != 0) {
353 {
354 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
355 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
356
357 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0));
358 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w));
359
360 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
361 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
362 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1));
363 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 8)));
364
365 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
366 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
367 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2));
368 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 16)));
369
370 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
371 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
372 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3));
373 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 24)));
374
375 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
376 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
377 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4));
378 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 32)));
379
380 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
381 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
382 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5));
383 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 40)));
384
385 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
386 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
387 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6));
388 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 48)));
389
390 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
391 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
392 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7));
393 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 56)));
394
395 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
396 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
397 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8));
398 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 64)));
399
400 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
401 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
402 const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9));
403 const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 72)));
404
405 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
406 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
407 const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10));
408 const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 80)));
409
410 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
411 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
412 const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11));
413 const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 88)));
414
415 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
416 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
417 const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12));
418 const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 96)));
419
420 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
421 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
422 const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13));
423 const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 104)));
424
425 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
426 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
427 const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14));
428 const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 112)));
429
430 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
431 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
432 const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15));
433 const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 120)));
434
435 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
436 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
437 const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16));
438 const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 128)));
439
440 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
441 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
442 const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17));
443 const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 136)));
444
445 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
446 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
447 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18));
448 const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 144)));
449
450 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
451 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
452 const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19));
453 const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 152)));
454
455 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
456 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
457 const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20));
458 const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 160)));
459
460 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
461 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
462 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21));
463 const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 168)));
464
465 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
466 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
467 const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22));
468 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176)));
469
470 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
471 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
472 const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23));
473 const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 184)));
474
475 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
476 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
477 const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24));
478 const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 192)));
479
480 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
481 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
482
483 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
484 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
485
486 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
487 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
488
489 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
490 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
491
492 #if XNN_ARCH_ARM64
493 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
494 #else
495 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
496 #endif
497 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
498
499 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
500 vout01234567 = vmax_s8(vout01234567, voutput_min);
501 vout01234567 = vmin_s8(vout01234567, voutput_max);
502
503 if (c & 4) {
504 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
505 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
506 }
507 if (c & 2) {
508 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
509 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
510 }
511 if (c & 1) {
512 vst1_lane_s8(output, vout01234567, 0); output += 1;
513 }
514 }
515 }
516
517 output = (int8_t*) ((uintptr_t) output + output_increment);
518 } while (--output_width != 0);
519 }
520