Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p16c-minmax-fp32-avx2-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul32.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 <immintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p16c__avx2_mul32(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36
37 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
38 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
39 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
40 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
41 XNN_FORCE_REALIZATION(voutput_zero_point);
42 XNN_FORCE_REALIZATION(voutput_min);
43
44 do {
45 const int8_t* i0 = input[0];
46 assert(i0 != NULL);
47 if XNN_UNPREDICTABLE(i0 != zero) {
48 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
49 }
50 const int8_t* i1 = input[1];
51 assert(i1 != NULL);
52 if XNN_UNPREDICTABLE(i1 != zero) {
53 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
54 }
55 const int8_t* i2 = input[2];
56 assert(i2 != NULL);
57 if XNN_UNPREDICTABLE(i2 != zero) {
58 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
59 }
60 const int8_t* i3 = input[3];
61 assert(i3 != NULL);
62 if XNN_UNPREDICTABLE(i3 != zero) {
63 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
64 }
65 const int8_t* i4 = input[4];
66 assert(i4 != NULL);
67 if XNN_UNPREDICTABLE(i4 != zero) {
68 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
69 }
70 const int8_t* i5 = input[5];
71 assert(i5 != NULL);
72 if XNN_UNPREDICTABLE(i5 != zero) {
73 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
74 }
75 const int8_t* i6 = input[6];
76 assert(i6 != NULL);
77 if XNN_UNPREDICTABLE(i6 != zero) {
78 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
79 }
80 const int8_t* i7 = input[7];
81 assert(i7 != NULL);
82 if XNN_UNPREDICTABLE(i7 != zero) {
83 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
84 }
85 const int8_t* i8 = input[8];
86 assert(i8 != NULL);
87 if XNN_UNPREDICTABLE(i8 != zero) {
88 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
89 }
90 input = (const int8_t**) ((uintptr_t) input + input_stride);
91
92 size_t c = channels;
93 const void* w = weights;
94 for (; c >= 16; c -= 16) {
95 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
96 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
97
98
99 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
100 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
101 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
102 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
103 i0 += 16;
104
105 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
106 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
107
108 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
109 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
110 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
111 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
112 i1 += 16;
113
114 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
115 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
116
117 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
118 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
119 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
120 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
121 i2 += 16;
122
123 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
124 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
125
126 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
127 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
128 const __m256i vi3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
129 const __m256i vk3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t))));
130 i3 += 16;
131
132 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
133 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
134
135 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
136 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
137 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
138 const __m256i vk4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t))));
139 i4 += 16;
140
141 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
142 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
143
144 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
145 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
146 const __m256i vi5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
147 const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t))));
148 i5 += 16;
149
150 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
151 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
152
153 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
154 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
155 const __m256i vi6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
156 const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t))));
157 i6 += 16;
158
159 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
160 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
161
162 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
163 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
164 const __m256i vi7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
165 const __m256i vk7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t))));
166 i7 += 16;
167
168 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
169 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
170
171 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
172 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
173 const __m256i vi8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 8)));
174 const __m256i vk8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t))));
175 i8 += 16;
176
177 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
178 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi8x89ABCDEF, vk8x89ABCDEF));
179
180 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
181
182 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
183 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
184
185 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
186 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale);
187
188 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
189 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
190
191 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
192 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
193
194 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
195
196 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
197
198 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
199
200 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
201 output += 16;
202 }
203 if XNN_UNLIKELY(c != 0) {
204 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
205 do {
206 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
207
208
209 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
210 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) k));
211 i0 += 8;
212
213 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
214
215 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
216 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 16)));
217 i1 += 8;
218
219 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
220
221 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
222 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 32)));
223 i2 += 8;
224
225 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
226
227 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
228 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 48)));
229 i3 += 8;
230
231 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
232
233 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
234 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 64)));
235 i4 += 8;
236
237 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
238
239 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
240 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 80)));
241 i5 += 8;
242
243 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
244
245 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
246 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 96)));
247 i6 += 8;
248
249 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
250
251 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
252 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 112)));
253 i7 += 8;
254
255 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
256
257 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
258 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 128)));
259 i8 += 8;
260
261 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
262
263 k += 8;
264
265 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
266 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
267 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
268 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
269
270 w = (const void*) ((const int32_t*) w + 8);
271
272 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
273
274 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
275
276 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
277
278 if XNN_LIKELY(c >= 8) {
279 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
280 output += 8;
281 c -= 8;
282 } else {
283 if (c & 4) {
284 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
285 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
286 output += 4;
287 }
288 if (c & 2) {
289 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
290 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
291 output += 2;
292 }
293 if (c & 1) {
294 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
295 output += 1;
296 }
297 c = 0;
298 }
299 } while (c != 0);
300 }
301
302 output = (int8_t*) ((uintptr_t) output + output_increment);
303 } while (--output_width != 0);
304 }