1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20
21
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16(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)])22 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16(
23 size_t channels,
24 size_t output_width,
25 const int8_t** input,
26 const void* weights,
27 int8_t* output,
28 size_t input_stride,
29 size_t output_increment,
30 size_t input_offset,
31 const int8_t* zero,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 for (; c >= 8; c -= 8) {
88 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
89 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
90
91
92 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
93 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
94 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
95 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
96 i0 += 8;
97
98
99 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
100
101
102 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
103 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
104 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
105 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
106 i1 += 8;
107
108
109 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
110
111 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
112 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
113
114 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
115 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
116 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
117 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
118 i2 += 8;
119
120
121 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
122
123
124 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
125 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
126 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
127 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
128 i3 += 8;
129
130
131 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
132
133 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
134 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
135
136 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
137 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
138 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
139 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
140 i4 += 8;
141
142
143 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
144
145
146 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
147 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
148 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
149 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
150 i5 += 8;
151
152
153 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
154
155 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
156 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
157
158 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
159 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
160 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
161 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
162 i6 += 8;
163
164
165 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
166
167
168 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
169 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
170 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
171 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
172 i7 += 8;
173
174
175 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
176
177 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
178 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
179
180 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
181 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
182 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
183 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
184 i8 += 8;
185
186
187 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
188
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
191
192 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
193
194 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
195 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
196
197 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
198 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
199 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
200
201 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
202 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
203 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
204
205 vacc0123 = _mm_cvtps_epi32(vscaled0123);
206 vacc4567 = _mm_cvtps_epi32(vscaled4567);
207
208 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
209 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
210
211
212 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
213
214 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
215 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
216
217 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
218 output += 8;
219 }
220 if XNN_UNLIKELY(c != 0) {
221 {
222 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
223 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
224
225
226 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
227 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
228 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
229 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
230
231
232 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
233
234
235 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
236 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
237 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
238 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
239
240
241 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
242
243 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
244 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
245
246 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
247 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
248 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
249 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
250
251
252 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
253
254
255 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
256 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
257 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
258 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
259
260
261 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
262
263 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
264 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
265
266 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
267 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
268 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
269 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
270
271
272 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
273
274
275 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
276 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
277 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
278 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
279
280
281 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
282
283 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
284 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
285
286 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
287 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
288 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
289 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
290
291
292 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
293
294
295 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
296 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
297 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
298 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
299
300
301 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
302
303 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
304 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
305
306 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
307 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
308 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
309 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
310
311
312 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
313
314 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
315 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
316
317
318 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
319 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
320
321 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
322 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
323 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
324
325 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
326 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
327 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
328
329 vacc0123 = _mm_cvtps_epi32(vscaled0123);
330 vacc4567 = _mm_cvtps_epi32(vscaled4567);
331
332
333 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
334 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
335
336
337 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
338
339 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
340
341 if (c & 4) {
342 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
343 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
344 output += 4;
345 }
346 if (c & 2) {
347 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
348 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
349 output += 2;
350 }
351 if (c & 1) {
352 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
353 output += 1;
354 }
355 }
356 }
357
358 output = (int8_t*) ((uintptr_t) output + output_increment);
359 } while (--output_width != 0);
360 }
361