Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p8c-minmax-fp32-sse41-mul16.c

raw
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 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p8c__sse41_mul16(
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 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35
36 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
39 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42
43 do {
44 const int8_t* i0 = input[0];
45 assert(i0 != NULL);
46 if XNN_UNPREDICTABLE(i0 != zero) {
47 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
48 }
49 const int8_t* i1 = input[1];
50 assert(i1 != NULL);
51 if XNN_UNPREDICTABLE(i1 != zero) {
52 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
53 }
54 const int8_t* i2 = input[2];
55 assert(i2 != NULL);
56 if XNN_UNPREDICTABLE(i2 != zero) {
57 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
58 }
59 const int8_t* i3 = input[3];
60 assert(i3 != NULL);
61 if XNN_UNPREDICTABLE(i3 != zero) {
62 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
63 }
64 const int8_t* i4 = input[4];
65 assert(i4 != NULL);
66 if XNN_UNPREDICTABLE(i4 != zero) {
67 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
68 }
69 const int8_t* i5 = input[5];
70 assert(i5 != NULL);
71 if XNN_UNPREDICTABLE(i5 != zero) {
72 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
73 }
74 const int8_t* i6 = input[6];
75 assert(i6 != NULL);
76 if XNN_UNPREDICTABLE(i6 != zero) {
77 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
78 }
79 const int8_t* i7 = input[7];
80 assert(i7 != NULL);
81 if XNN_UNPREDICTABLE(i7 != zero) {
82 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
83 }
84 const int8_t* i8 = input[8];
85 assert(i8 != NULL);
86 if XNN_UNPREDICTABLE(i8 != zero) {
87 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
88 }
89 input = (const int8_t**) ((uintptr_t) input + input_stride);
90
91 size_t c = channels;
92 const void* w = weights;
93 for (; c >= 8; c -= 8) {
94 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
95 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
96
97
98 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
99 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
100 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
101 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
102 i0 += 8;
103
104
105 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
106
107 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
109
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
111 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
112 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
113 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
114 i1 += 8;
115
116
117 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
118
119 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
120 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
121
122 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
123 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
124 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
125 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
126 i2 += 8;
127
128
129 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
130
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
133
134 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
135 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
136 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
137 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
138 i3 += 8;
139
140
141 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
142
143 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
144 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
145
146 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
147 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
148 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
149 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
150 i4 += 8;
151
152
153 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
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 vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
159 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
160 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
161 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
162 i5 += 8;
163
164
165 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
166
167 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
168 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
169
170 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
171 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
172 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
173 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
174 i6 += 8;
175
176
177 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
178
179 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
180 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
181
182 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
183 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
184 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
185 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
186 i7 += 8;
187
188
189 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
193
194 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
195 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
196 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
197 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
198 i8 += 8;
199
200
201 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
205
206 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
207
208 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
209 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
210
211 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
212 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
213
214 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
215 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
216
217 vacc0123 = _mm_cvtps_epi32(vscaled0123);
218 vacc4567 = _mm_cvtps_epi32(vscaled4567);
219
220 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
221
222
223 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
224
225 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
226
227 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
228 output += 8;
229 }
230 if XNN_UNLIKELY(c != 0) {
231 {
232 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
233 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
234
235
236 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
237 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
238 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
239 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
240
241
242 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
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 vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
248 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
249 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
250 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
251
252
253 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
254
255 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
256 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
257
258 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
259 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
260 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
261 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
262
263
264 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
268
269 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
270 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
271 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
272 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
273
274
275 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
276
277 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
278 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
279
280 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
281 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
282 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
283 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
284
285
286 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
287
288 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
289 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
290
291 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
292 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
293 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
294 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
295
296
297 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
298
299 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
300 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
301
302 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
303 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
304 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
305 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
306
307
308 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
311 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
312
313 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
314 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
315 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
316 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
317
318
319 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
323
324 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
325 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
326 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
327 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
328
329
330 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
331
332 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
333 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
334
335
336 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
337 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
338
339 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
340 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
341
342 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
343 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
344
345 vacc0123 = _mm_cvtps_epi32(vscaled0123);
346 vacc4567 = _mm_cvtps_epi32(vscaled4567);
347
348
349 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
350
351
352 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
353
354 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
355
356 if (c & 4) {
357 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
358 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
359 output += 4;
360 }
361 if (c & 2) {
362 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
363 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
364 output += 2;
365 }
366 if (c & 1) {
367 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
368 output += 1;
369 }
370 }
371 }
372
373 output = (int8_t*) ((uintptr_t) output + output_increment);
374 } while (--output_width != 0);
375 }