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 #include <xnnpack/unaligned.h>
21
22
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)])23 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul16_add16(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
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 input = (const int8_t**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const void* w = weights;
88 for (; c >= 8; c -= 8) {
89 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
90 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
91
92
93 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
94 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
95 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
96 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
97 i0 += 8;
98
99
100 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
101
102
103 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
104 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
105 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
106 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
107 i1 += 8;
108
109
110 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
111
112 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
113 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
114
115 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
116 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
117 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
118 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
119 i2 += 8;
120
121
122 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
123
124
125 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
126 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
127 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
128 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
129 i3 += 8;
130
131
132 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
133
134 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
135 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
136
137 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
138 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
139 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
140 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
141 i4 += 8;
142
143
144 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
145
146
147 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
148 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
149 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
150 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
151 i5 += 8;
152
153
154 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
155
156 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
157 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
158
159 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
160 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
161 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
162 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
163 i6 += 8;
164
165
166 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
167
168
169 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
170 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
171 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
172 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
173 i7 += 8;
174
175
176 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
177
178 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
179 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
180
181 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
182 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
183 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
184 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
185 i8 += 8;
186
187
188 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
192
193 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
194
195 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
196 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
197
198 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
199 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
200 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
201
202 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
203 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
204 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
205
206 vacc0123 = _mm_cvtps_epi32(vscaled0123);
207 vacc4567 = _mm_cvtps_epi32(vscaled4567);
208
209 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
210 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
211
212
213 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
214
215 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
216 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
217
218 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
219 output += 8;
220 }
221 if XNN_UNLIKELY(c != 0) {
222 {
223 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
224 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
225
226
227 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
228 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
229 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
230 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
231
232
233 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
234
235
236 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
237 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
238 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
239 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
240
241
242 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
243
244 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
245 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
246
247 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
248 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
249 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
250 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
251
252
253 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
254
255
256 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
257 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
258 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
259 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
260
261
262 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
263
264 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
265 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
266
267 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
268 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
269 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
270 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
271
272
273 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
274
275
276 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
277 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
278 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
279 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
280
281
282 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
286
287 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
288 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
289 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
290 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
291
292
293 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
294
295
296 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
297 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
298 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
299 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
300
301
302 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
303
304 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
305 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
306
307 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
308 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
309 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
310 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
311
312
313 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
314
315 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
316 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
317
318
319 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
320 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
321
322 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
323 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
324 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
325
326 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
327 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
328 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
329
330 vacc0123 = _mm_cvtps_epi32(vscaled0123);
331 vacc4567 = _mm_cvtps_epi32(vscaled4567);
332
333
334 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
335 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
336
337
338 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
339
340 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
341
342 if (c & 4) {
343 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
344 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
345 output += 4;
346 }
347 if (c & 2) {
348 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
349 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
350 output += 2;
351 }
352 if (c & 1) {
353 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
354 output += 1;
355 }
356 }
357 }
358
359 output = (int8_t*) ((uintptr_t) output + output_increment);
360 } while (--output_width != 0);
361 }
362