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_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_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)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mla8_ld64(
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_qc8_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 float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
33 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
34 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
35 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
36 do {
37 const int8_t* i0 = input[0];
38 assert(i0 != NULL);
39 if XNN_UNPREDICTABLE(i0 != zero) {
40 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41 }
42 const int8_t* i1 = input[1];
43 assert(i1 != NULL);
44 if XNN_UNPREDICTABLE(i1 != zero) {
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 }
47 const int8_t* i2 = input[2];
48 assert(i2 != NULL);
49 if XNN_UNPREDICTABLE(i2 != zero) {
50 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51 }
52 const int8_t* i3 = input[3];
53 assert(i3 != NULL);
54 if XNN_UNPREDICTABLE(i3 != zero) {
55 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56 }
57 const int8_t* i4 = input[4];
58 assert(i4 != NULL);
59 if XNN_UNPREDICTABLE(i4 != zero) {
60 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61 }
62 const int8_t* i5 = input[5];
63 assert(i5 != NULL);
64 if XNN_UNPREDICTABLE(i5 != zero) {
65 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66 }
67 const int8_t* i6 = input[6];
68 assert(i6 != NULL);
69 if XNN_UNPREDICTABLE(i6 != zero) {
70 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71 }
72 const int8_t* i7 = input[7];
73 assert(i7 != NULL);
74 if XNN_UNPREDICTABLE(i7 != zero) {
75 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76 }
77 const int8_t* i8 = input[8];
78 assert(i8 != NULL);
79 if XNN_UNPREDICTABLE(i8 != zero) {
80 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81 }
82 const int8_t* i9 = input[9];
83 assert(i9 != NULL);
84 if XNN_UNPREDICTABLE(i9 != zero) {
85 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86 }
87 const int8_t* i10 = input[10];
88 assert(i10 != NULL);
89 if XNN_UNPREDICTABLE(i10 != zero) {
90 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91 }
92 const int8_t* i11 = input[11];
93 assert(i11 != NULL);
94 if XNN_UNPREDICTABLE(i11 != zero) {
95 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96 }
97 const int8_t* i12 = input[12];
98 assert(i12 != NULL);
99 if XNN_UNPREDICTABLE(i12 != zero) {
100 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101 }
102 const int8_t* i13 = input[13];
103 assert(i13 != NULL);
104 if XNN_UNPREDICTABLE(i13 != zero) {
105 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106 }
107 const int8_t* i14 = input[14];
108 assert(i14 != NULL);
109 if XNN_UNPREDICTABLE(i14 != zero) {
110 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111 }
112 const int8_t* i15 = input[15];
113 assert(i15 != NULL);
114 if XNN_UNPREDICTABLE(i15 != zero) {
115 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116 }
117 const int8_t* i16 = input[16];
118 assert(i16 != NULL);
119 if XNN_UNPREDICTABLE(i16 != zero) {
120 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121 }
122 const int8_t* i17 = input[17];
123 assert(i17 != NULL);
124 if XNN_UNPREDICTABLE(i17 != zero) {
125 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126 }
127 const int8_t* i18 = input[18];
128 assert(i18 != NULL);
129 if XNN_UNPREDICTABLE(i18 != zero) {
130 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131 }
132 const int8_t* i19 = input[19];
133 assert(i19 != NULL);
134 if XNN_UNPREDICTABLE(i19 != zero) {
135 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136 }
137 const int8_t* i20 = input[20];
138 assert(i20 != NULL);
139 if XNN_UNPREDICTABLE(i20 != zero) {
140 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141 }
142 const int8_t* i21 = input[21];
143 assert(i21 != NULL);
144 if XNN_UNPREDICTABLE(i21 != zero) {
145 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146 }
147 const int8_t* i22 = input[22];
148 assert(i22 != NULL);
149 if XNN_UNPREDICTABLE(i22 != zero) {
150 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151 }
152 const int8_t* i23 = input[23];
153 assert(i23 != NULL);
154 if XNN_UNPREDICTABLE(i23 != zero) {
155 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156 }
157 const int8_t* i24 = input[24];
158 assert(i24 != NULL);
159 if XNN_UNPREDICTABLE(i24 != zero) {
160 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161 }
162 input = (const int8_t**) ((uintptr_t) input + input_stride);
163
164 size_t c = channels;
165 const void* w = weights;
166 for (; c >= 16; c -= 16) {
167 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
168 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169 int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170 int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171
172 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
173 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
174 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
175 const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176
177 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
178 int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
179
180 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
181 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
182 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
183 const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
184
185 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
186 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi1x89ABCDEF, vk1x89ABCDEF);
187
188 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
189 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
190 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
191 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
192 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
193 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
194 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
195 const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
196
197 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
198 vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
199
200 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
201 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
202 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
203 const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
204
205 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
206 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi3x89ABCDEF, vk3x89ABCDEF);
207
208 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
209 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
210 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
211 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
212 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
213 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
214 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
215 const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
216
217 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
218 vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
219
220 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
221 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
222 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
223 const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
224
225 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
226 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi5x89ABCDEF, vk5x89ABCDEF);
227
228 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
229 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
230 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
231 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
232 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
233 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
234 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
235 const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
236
237 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
238 vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
239
240 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
241 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
242 const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
243 const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
244
245 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
246 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi7x89ABCDEF, vk7x89ABCDEF);
247
248 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
249 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
250 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
251 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
252 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
253 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
254 const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
255 const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
256
257 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
258 vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
259
260 const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
261 const int8x8_t vk9x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
262 const int8x8_t vi9x89ABCDEF = vld1_s8(i9); i9 += 8;
263 const int8x8_t vk9x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
264
265 vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
266 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi9x89ABCDEF, vk9x89ABCDEF);
267
268 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
269 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
270 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
271 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
272 const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
273 const int8x8_t vk10x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
274 const int8x8_t vi10x89ABCDEF = vld1_s8(i10); i10 += 8;
275 const int8x8_t vk10x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
276
277 vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
278 vprod89ABCDEF = vmull_s8(vi10x89ABCDEF, vk10x89ABCDEF);
279
280 const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
281 const int8x8_t vk11x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
282 const int8x8_t vi11x89ABCDEF = vld1_s8(i11); i11 += 8;
283 const int8x8_t vk11x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
284
285 vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
286 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi11x89ABCDEF, vk11x89ABCDEF);
287
288 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
289 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
290 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
291 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
292 const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
293 const int8x8_t vk12x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
294 const int8x8_t vi12x89ABCDEF = vld1_s8(i12); i12 += 8;
295 const int8x8_t vk12x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
296
297 vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
298 vprod89ABCDEF = vmull_s8(vi12x89ABCDEF, vk12x89ABCDEF);
299
300 const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
301 const int8x8_t vk13x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
302 const int8x8_t vi13x89ABCDEF = vld1_s8(i13); i13 += 8;
303 const int8x8_t vk13x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
304
305 vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
306 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi13x89ABCDEF, vk13x89ABCDEF);
307
308 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
309 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
310 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
311 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
312 const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
313 const int8x8_t vk14x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
314 const int8x8_t vi14x89ABCDEF = vld1_s8(i14); i14 += 8;
315 const int8x8_t vk14x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
316
317 vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
318 vprod89ABCDEF = vmull_s8(vi14x89ABCDEF, vk14x89ABCDEF);
319
320 const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
321 const int8x8_t vk15x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
322 const int8x8_t vi15x89ABCDEF = vld1_s8(i15); i15 += 8;
323 const int8x8_t vk15x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
324
325 vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
326 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi15x89ABCDEF, vk15x89ABCDEF);
327
328 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
329 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
330 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
331 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
332 const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
333 const int8x8_t vk16x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
334 const int8x8_t vi16x89ABCDEF = vld1_s8(i16); i16 += 8;
335 const int8x8_t vk16x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
336
337 vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
338 vprod89ABCDEF = vmull_s8(vi16x89ABCDEF, vk16x89ABCDEF);
339
340 const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
341 const int8x8_t vk17x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
342 const int8x8_t vi17x89ABCDEF = vld1_s8(i17); i17 += 8;
343 const int8x8_t vk17x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
344
345 vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
346 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi17x89ABCDEF, vk17x89ABCDEF);
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 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
353 const int8x8_t vk18x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
354 const int8x8_t vi18x89ABCDEF = vld1_s8(i18); i18 += 8;
355 const int8x8_t vk18x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
356
357 vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
358 vprod89ABCDEF = vmull_s8(vi18x89ABCDEF, vk18x89ABCDEF);
359
360 const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
361 const int8x8_t vk19x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
362 const int8x8_t vi19x89ABCDEF = vld1_s8(i19); i19 += 8;
363 const int8x8_t vk19x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
364
365 vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
366 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi19x89ABCDEF, vk19x89ABCDEF);
367
368 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
369 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
370 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
371 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
372 const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
373 const int8x8_t vk20x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
374 const int8x8_t vi20x89ABCDEF = vld1_s8(i20); i20 += 8;
375 const int8x8_t vk20x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
376
377 vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
378 vprod89ABCDEF = vmull_s8(vi20x89ABCDEF, vk20x89ABCDEF);
379
380 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
381 const int8x8_t vk21x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
382 const int8x8_t vi21x89ABCDEF = vld1_s8(i21); i21 += 8;
383 const int8x8_t vk21x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
384
385 vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
386 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi21x89ABCDEF, vk21x89ABCDEF);
387
388 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
389 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
390 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
391 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
392 const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
393 const int8x8_t vk22x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
394 const int8x8_t vi22x89ABCDEF = vld1_s8(i22); i22 += 8;
395 const int8x8_t vk22x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
396
397 vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
398 vprod89ABCDEF = vmull_s8(vi22x89ABCDEF, vk22x89ABCDEF);
399
400 const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
401 const int8x8_t vk23x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
402 const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8;
403 const int8x8_t vk23x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
404
405 vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
406 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi23x89ABCDEF, vk23x89ABCDEF);
407
408 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
409 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
410 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
411 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
412 const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
413 const int8x8_t vk24x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
414 const int8x8_t vi24x89ABCDEF = vld1_s8(i24); i24 += 8;
415 const int8x8_t vk24x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
416
417 vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
418 vprod89ABCDEF = vmull_s8(vi24x89ABCDEF, vk24x89ABCDEF);
419
420 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
421 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
422 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
423 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
424
425 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
426 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
427 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
428 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
429
430 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
431 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
432 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
433 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
434
435 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
436 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
437 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
438 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
439
440 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
441 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
442 vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
443 vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
444
445 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
446 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
447 vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
448 vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
449
450 #if XNN_ARCH_ARM64
451 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
452 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
453
454
455 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
456 #else // !XNN_ARCH_ARM64
457 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
458 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
459
460
461 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
462 #endif // !XNN_ARCH_ARM64
463
464 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
465
466 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
467
468 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
469 }
470 if XNN_UNLIKELY(c != 0) {
471 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
472 do {
473 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
474 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
475
476 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
477 const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
478
479 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
480
481 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
482 const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
483
484 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
485
486 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
487 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
488 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
489 const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
490
491 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
492
493 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
494 const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
495
496 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
497
498 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
499 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
500 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
501 const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
502
503 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
504
505 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
506 const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
507
508 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
509
510 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
511 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
512 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
513 const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
514
515 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
516
517 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
518 const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
519
520 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
521
522 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
523 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
524 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
525 const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
526
527 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
528
529 const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
530 const int8x8_t vk9x01234567 = vld1_s8((const void*) (k + 136));
531
532 vprod01234567 = vmlal_s8(vprod01234567, vi9x01234567, vk9x01234567);
533
534 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
535 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
536 const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
537 const int8x8_t vk10x01234567 = vld1_s8((const void*) (k + 152));
538
539 vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
540
541 const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
542 const int8x8_t vk11x01234567 = vld1_s8((const void*) (k + 168));
543
544 vprod01234567 = vmlal_s8(vprod01234567, vi11x01234567, vk11x01234567);
545
546 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
547 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
548 const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
549 const int8x8_t vk12x01234567 = vld1_s8((const void*) (k + 184));
550
551 vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
552
553 const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
554 const int8x8_t vk13x01234567 = vld1_s8((const void*) (k + 200));
555
556 vprod01234567 = vmlal_s8(vprod01234567, vi13x01234567, vk13x01234567);
557
558 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
559 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
560 const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
561 const int8x8_t vk14x01234567 = vld1_s8((const void*) (k + 216));
562
563 vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
564
565 const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
566 const int8x8_t vk15x01234567 = vld1_s8((const void*) (k + 232));
567
568 vprod01234567 = vmlal_s8(vprod01234567, vi15x01234567, vk15x01234567);
569
570 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
571 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
572 const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
573 const int8x8_t vk16x01234567 = vld1_s8((const void*) (k + 248));
574
575 vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
576
577 const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
578 const int8x8_t vk17x01234567 = vld1_s8((const void*) (k + 264));
579
580 vprod01234567 = vmlal_s8(vprod01234567, vi17x01234567, vk17x01234567);
581
582 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
583 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
584 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
585 const int8x8_t vk18x01234567 = vld1_s8((const void*) (k + 280));
586
587 vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
588
589 const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
590 const int8x8_t vk19x01234567 = vld1_s8((const void*) (k + 296));
591
592 vprod01234567 = vmlal_s8(vprod01234567, vi19x01234567, vk19x01234567);
593
594 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
595 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
596 const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
597 const int8x8_t vk20x01234567 = vld1_s8((const void*) (k + 312));
598
599 vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
600
601 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
602 const int8x8_t vk21x01234567 = vld1_s8((const void*) (k + 328));
603
604 vprod01234567 = vmlal_s8(vprod01234567, vi21x01234567, vk21x01234567);
605
606 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
607 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
608 const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
609 const int8x8_t vk22x01234567 = vld1_s8((const void*) (k + 344));
610
611 vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
612
613 const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
614 const int8x8_t vk23x01234567 = vld1_s8((const void*) (k + 360));
615
616 vprod01234567 = vmlal_s8(vprod01234567, vi23x01234567, vk23x01234567);
617
618 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
619 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
620 const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
621 const int8x8_t vk24x01234567 = vld1_s8((const void*) (k + 376));
622
623 vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
624
625 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
626 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
627
628 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
629 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
630
631 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t)));
632 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 400 * sizeof(int8_t) + 4 * sizeof(float)));
633 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
634 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
635
636 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
637 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
638
639 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
640 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
641
642 #if XNN_ARCH_ARM64
643 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
644 #else
645 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
646 #endif
647
648 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
649 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
650 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
651
652 if XNN_LIKELY(c >= 8) {
653 vst1_s8(output, vout01234567); output += 8;
654 c -= 8;
655 } else {
656 if (c & 4) {
657 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
658 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
659 }
660 if (c & 2) {
661 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
662 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
663 }
664 if (c & 1) {
665 vst1_lane_s8(output, vout01234567, 0); output += 1;
666 }
667 c = 0;
668 }
669 } while (c != 0);
670 }
671
672 output = (int8_t*) ((uintptr_t) output + output_increment);
673 } while (--output_width != 0);
674 }
675