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 #include <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_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)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16(
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 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const int8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 for (; c >= 8; c -= 8) {
83 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
84 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
85
86
87 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
88 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
89 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
90 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
91 i0 += 8;
92
93
94 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
95
96
97 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
98 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
99 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
100 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
101 i1 += 8;
102
103
104 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
105
106 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
107 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
108
109 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
110 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
111 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
112 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
113 i2 += 8;
114
115
116 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
117
118
119 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
120 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
121 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
122 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
123 i3 += 8;
124
125
126 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
127
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
130
131 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
132 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
133 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
134 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
135 i4 += 8;
136
137
138 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
139
140
141 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
142 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
143 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
144 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
145 i5 += 8;
146
147
148 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
149
150 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
151 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
152
153 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
154 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
155 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
156 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
157 i6 += 8;
158
159
160 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
161
162
163 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
164 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
165 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
166 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
167 i7 += 8;
168
169
170 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
171
172 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
173 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
174
175 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
176 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
177 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
178 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
179 i8 += 8;
180
181
182 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
183
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
186
187 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
188
189 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
190 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
191
192 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
193 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
194 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
195
196 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
197 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
198 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
199
200 vacc0123 = _mm_cvtps_epi32(vscaled0123);
201 vacc4567 = _mm_cvtps_epi32(vscaled4567);
202
203 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
204 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
205
206
207 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
208
209 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
210 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
211
212 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
213 output += 8;
214 }
215 if XNN_UNLIKELY(c != 0) {
216 {
217 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
218 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
219
220
221 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
222 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
223 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
224 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
225
226
227 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
228
229
230 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
231 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
232 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
233 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
234
235
236 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
237
238 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
239 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
240
241 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
242 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
243 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
244 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
245
246
247 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
248
249
250 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
251 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
252 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
253 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
254
255
256 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
257
258 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
259 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
260
261 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
262 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
263 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
264 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
265
266
267 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
268
269
270 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
271 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
272 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
273 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
274
275
276 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
277
278 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
279 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
280
281 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
282 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
283 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
284 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
285
286
287 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
288
289
290 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
291 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
292 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
293 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
294
295
296 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
300
301 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
302 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
303 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
304 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
305
306
307 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
311
312
313 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
314 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
315
316 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
317 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
318 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
319
320 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
321 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
322 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
323
324 vacc0123 = _mm_cvtps_epi32(vscaled0123);
325 vacc4567 = _mm_cvtps_epi32(vscaled4567);
326
327
328 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
329 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
330
331
332 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
333
334 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
335
336 if (c & 4) {
337 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
338 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
339 output += 4;
340 }
341 if (c & 2) {
342 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
343 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
344 output += 2;
345 }
346 if (c & 1) {
347 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
348 output += 1;
349 }
350 }
351 }
352
353 output = (int8_t*) ((uintptr_t) output + output_increment);
354 } while (--output_width != 0);
355 }
356