1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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 <xmmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const __m128 vmax = _mm_load_ps(params->sse.max);
33 const __m128 vmin = _mm_load_ps(params->sse.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 const float* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const float*) ((uintptr_t) i4 + input_offset);
59 }
60 const float* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const float*) ((uintptr_t) i5 + input_offset);
64 }
65 const float* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const float*) ((uintptr_t) i6 + input_offset);
69 }
70 const float* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const float*) ((uintptr_t) i7 + input_offset);
74 }
75 const float* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const float*) ((uintptr_t) i8 + input_offset);
79 }
80 const float* i9 = input[9];
81 assert(i9 != NULL);
82 if XNN_UNPREDICTABLE(i9 != zero) {
83 i9 = (const float*) ((uintptr_t) i9 + input_offset);
84 }
85 const float* i10 = input[10];
86 assert(i10 != NULL);
87 if XNN_UNPREDICTABLE(i10 != zero) {
88 i10 = (const float*) ((uintptr_t) i10 + input_offset);
89 }
90 const float* i11 = input[11];
91 assert(i11 != NULL);
92 if XNN_UNPREDICTABLE(i11 != zero) {
93 i11 = (const float*) ((uintptr_t) i11 + input_offset);
94 }
95 const float* i12 = input[12];
96 assert(i12 != NULL);
97 if XNN_UNPREDICTABLE(i12 != zero) {
98 i12 = (const float*) ((uintptr_t) i12 + input_offset);
99 }
100 const float* i13 = input[13];
101 assert(i13 != NULL);
102 if XNN_UNPREDICTABLE(i13 != zero) {
103 i13 = (const float*) ((uintptr_t) i13 + input_offset);
104 }
105 const float* i14 = input[14];
106 assert(i14 != NULL);
107 if XNN_UNPREDICTABLE(i14 != zero) {
108 i14 = (const float*) ((uintptr_t) i14 + input_offset);
109 }
110 const float* i15 = input[15];
111 assert(i15 != NULL);
112 if XNN_UNPREDICTABLE(i15 != zero) {
113 i15 = (const float*) ((uintptr_t) i15 + input_offset);
114 }
115 const float* i16 = input[16];
116 assert(i16 != NULL);
117 if XNN_UNPREDICTABLE(i16 != zero) {
118 i16 = (const float*) ((uintptr_t) i16 + input_offset);
119 }
120 const float* i17 = input[17];
121 assert(i17 != NULL);
122 if XNN_UNPREDICTABLE(i17 != zero) {
123 i17 = (const float*) ((uintptr_t) i17 + input_offset);
124 }
125 const float* i18 = input[18];
126 assert(i18 != NULL);
127 if XNN_UNPREDICTABLE(i18 != zero) {
128 i18 = (const float*) ((uintptr_t) i18 + input_offset);
129 }
130 const float* i19 = input[19];
131 assert(i19 != NULL);
132 if XNN_UNPREDICTABLE(i19 != zero) {
133 i19 = (const float*) ((uintptr_t) i19 + input_offset);
134 }
135 const float* i20 = input[20];
136 assert(i20 != NULL);
137 if XNN_UNPREDICTABLE(i20 != zero) {
138 i20 = (const float*) ((uintptr_t) i20 + input_offset);
139 }
140 const float* i21 = input[21];
141 assert(i21 != NULL);
142 if XNN_UNPREDICTABLE(i21 != zero) {
143 i21 = (const float*) ((uintptr_t) i21 + input_offset);
144 }
145 const float* i22 = input[22];
146 assert(i22 != NULL);
147 if XNN_UNPREDICTABLE(i22 != zero) {
148 i22 = (const float*) ((uintptr_t) i22 + input_offset);
149 }
150 const float* i23 = input[23];
151 assert(i23 != NULL);
152 if XNN_UNPREDICTABLE(i23 != zero) {
153 i23 = (const float*) ((uintptr_t) i23 + input_offset);
154 }
155 const float* i24 = input[24];
156 assert(i24 != NULL);
157 if XNN_UNPREDICTABLE(i24 != zero) {
158 i24 = (const float*) ((uintptr_t) i24 + input_offset);
159 }
160 input = (const float**) ((uintptr_t) input + input_stride);
161
162 size_t c = channels;
163 const float* w = weights;
164 for (; c >= 4; c -= 4) {
165 __m128 vacc0123p0 = _mm_load_ps(w);
166
167
168 const __m128 vi0x0123 = _mm_loadu_ps(i0);
169 i0 += 4;
170
171 const __m128 vk0x0123 = _mm_load_ps(w + 4);
172 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
173
174 const __m128 vi1x0123 = _mm_loadu_ps(i1);
175 i1 += 4;
176
177 const __m128 vk1x0123 = _mm_load_ps(w + 8);
178 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
179
180 const __m128 vi2x0123 = _mm_loadu_ps(i2);
181 i2 += 4;
182
183 const __m128 vk2x0123 = _mm_load_ps(w + 12);
184 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
185
186 const __m128 vi3x0123 = _mm_loadu_ps(i3);
187 i3 += 4;
188
189 const __m128 vk3x0123 = _mm_load_ps(w + 16);
190 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
191
192 const __m128 vi4x0123 = _mm_loadu_ps(i4);
193 i4 += 4;
194
195 const __m128 vk4x0123 = _mm_load_ps(w + 20);
196 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
197
198 const __m128 vi5x0123 = _mm_loadu_ps(i5);
199 i5 += 4;
200
201 const __m128 vk5x0123 = _mm_load_ps(w + 24);
202 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
203
204 const __m128 vi6x0123 = _mm_loadu_ps(i6);
205 i6 += 4;
206
207 const __m128 vk6x0123 = _mm_load_ps(w + 28);
208 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
209
210 const __m128 vi7x0123 = _mm_loadu_ps(i7);
211 i7 += 4;
212
213 const __m128 vk7x0123 = _mm_load_ps(w + 32);
214 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
215
216 const __m128 vi8x0123 = _mm_loadu_ps(i8);
217 i8 += 4;
218
219 const __m128 vk8x0123 = _mm_load_ps(w + 36);
220 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
221
222 const __m128 vi9x0123 = _mm_loadu_ps(i9);
223 i9 += 4;
224
225 const __m128 vk9x0123 = _mm_load_ps(w + 40);
226 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123));
227
228 const __m128 vi10x0123 = _mm_loadu_ps(i10);
229 i10 += 4;
230
231 const __m128 vk10x0123 = _mm_load_ps(w + 44);
232 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123));
233
234 const __m128 vi11x0123 = _mm_loadu_ps(i11);
235 i11 += 4;
236
237 const __m128 vk11x0123 = _mm_load_ps(w + 48);
238 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123));
239
240 const __m128 vi12x0123 = _mm_loadu_ps(i12);
241 i12 += 4;
242
243 const __m128 vk12x0123 = _mm_load_ps(w + 52);
244 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123));
245
246 const __m128 vi13x0123 = _mm_loadu_ps(i13);
247 i13 += 4;
248
249 const __m128 vk13x0123 = _mm_load_ps(w + 56);
250 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123));
251
252 const __m128 vi14x0123 = _mm_loadu_ps(i14);
253 i14 += 4;
254
255 const __m128 vk14x0123 = _mm_load_ps(w + 60);
256 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123));
257
258 const __m128 vi15x0123 = _mm_loadu_ps(i15);
259 i15 += 4;
260
261 const __m128 vk15x0123 = _mm_load_ps(w + 64);
262 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123));
263
264 const __m128 vi16x0123 = _mm_loadu_ps(i16);
265 i16 += 4;
266
267 const __m128 vk16x0123 = _mm_load_ps(w + 68);
268 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123));
269
270 const __m128 vi17x0123 = _mm_loadu_ps(i17);
271 i17 += 4;
272
273 const __m128 vk17x0123 = _mm_load_ps(w + 72);
274 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123));
275
276 const __m128 vi18x0123 = _mm_loadu_ps(i18);
277 i18 += 4;
278
279 const __m128 vk18x0123 = _mm_load_ps(w + 76);
280 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123));
281
282 const __m128 vi19x0123 = _mm_loadu_ps(i19);
283 i19 += 4;
284
285 const __m128 vk19x0123 = _mm_load_ps(w + 80);
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123));
287
288 const __m128 vi20x0123 = _mm_loadu_ps(i20);
289 i20 += 4;
290
291 const __m128 vk20x0123 = _mm_load_ps(w + 84);
292 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123));
293
294 const __m128 vi21x0123 = _mm_loadu_ps(i21);
295 i21 += 4;
296
297 const __m128 vk21x0123 = _mm_load_ps(w + 88);
298 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123));
299
300 const __m128 vi22x0123 = _mm_loadu_ps(i22);
301 i22 += 4;
302
303 const __m128 vk22x0123 = _mm_load_ps(w + 92);
304 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123));
305
306 const __m128 vi23x0123 = _mm_loadu_ps(i23);
307 i23 += 4;
308
309 const __m128 vk23x0123 = _mm_load_ps(w + 96);
310 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123));
311
312 const __m128 vi24x0123 = _mm_loadu_ps(i24);
313 i24 += 4;
314
315 const __m128 vk24x0123 = _mm_load_ps(w + 100);
316 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123));
317
318 w += 104;
319
320 // Add up all accumulators to vacc0123p0
321 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
322
323 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
324 vacc0123 = _mm_min_ps(vacc0123, vmax);
325
326 _mm_storeu_ps(output, vacc0123);
327 output += 4;
328 }
329 if XNN_UNLIKELY(c != 0) {
330 __m128 vacc0123p0 = _mm_load_ps(w);
331
332 const __m128 vi0x0123 = _mm_loadu_ps(i0);
333 const __m128 vk0x0123 = _mm_load_ps(w + 4);
334 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
335
336 const __m128 vi1x0123 = _mm_loadu_ps(i1);
337 const __m128 vk1x0123 = _mm_load_ps(w + 8);
338 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
339
340 const __m128 vi2x0123 = _mm_loadu_ps(i2);
341 const __m128 vk2x0123 = _mm_load_ps(w + 12);
342 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
343
344 const __m128 vi3x0123 = _mm_loadu_ps(i3);
345 const __m128 vk3x0123 = _mm_load_ps(w + 16);
346 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
347
348 const __m128 vi4x0123 = _mm_loadu_ps(i4);
349 const __m128 vk4x0123 = _mm_load_ps(w + 20);
350 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));
351
352 const __m128 vi5x0123 = _mm_loadu_ps(i5);
353 const __m128 vk5x0123 = _mm_load_ps(w + 24);
354 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123));
355
356 const __m128 vi6x0123 = _mm_loadu_ps(i6);
357 const __m128 vk6x0123 = _mm_load_ps(w + 28);
358 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));
359
360 const __m128 vi7x0123 = _mm_loadu_ps(i7);
361 const __m128 vk7x0123 = _mm_load_ps(w + 32);
362 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123));
363
364 const __m128 vi8x0123 = _mm_loadu_ps(i8);
365 const __m128 vk8x0123 = _mm_load_ps(w + 36);
366 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));
367
368 const __m128 vi9x0123 = _mm_loadu_ps(i9);
369 const __m128 vk9x0123 = _mm_load_ps(w + 40);
370 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123));
371
372 const __m128 vi10x0123 = _mm_loadu_ps(i10);
373 const __m128 vk10x0123 = _mm_load_ps(w + 44);
374 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123));
375
376 const __m128 vi11x0123 = _mm_loadu_ps(i11);
377 const __m128 vk11x0123 = _mm_load_ps(w + 48);
378 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123));
379
380 const __m128 vi12x0123 = _mm_loadu_ps(i12);
381 const __m128 vk12x0123 = _mm_load_ps(w + 52);
382 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123));
383
384 const __m128 vi13x0123 = _mm_loadu_ps(i13);
385 const __m128 vk13x0123 = _mm_load_ps(w + 56);
386 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123));
387
388 const __m128 vi14x0123 = _mm_loadu_ps(i14);
389 const __m128 vk14x0123 = _mm_load_ps(w + 60);
390 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi14x0123, vk14x0123));
391
392 const __m128 vi15x0123 = _mm_loadu_ps(i15);
393 const __m128 vk15x0123 = _mm_load_ps(w + 64);
394 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123));
395
396 const __m128 vi16x0123 = _mm_loadu_ps(i16);
397 const __m128 vk16x0123 = _mm_load_ps(w + 68);
398 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123));
399
400 const __m128 vi17x0123 = _mm_loadu_ps(i17);
401 const __m128 vk17x0123 = _mm_load_ps(w + 72);
402 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123));
403
404 const __m128 vi18x0123 = _mm_loadu_ps(i18);
405 const __m128 vk18x0123 = _mm_load_ps(w + 76);
406 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi18x0123, vk18x0123));
407
408 const __m128 vi19x0123 = _mm_loadu_ps(i19);
409 const __m128 vk19x0123 = _mm_load_ps(w + 80);
410 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123));
411
412 const __m128 vi20x0123 = _mm_loadu_ps(i20);
413 const __m128 vk20x0123 = _mm_load_ps(w + 84);
414 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi20x0123, vk20x0123));
415
416 const __m128 vi21x0123 = _mm_loadu_ps(i21);
417 const __m128 vk21x0123 = _mm_load_ps(w + 88);
418 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi21x0123, vk21x0123));
419
420 const __m128 vi22x0123 = _mm_loadu_ps(i22);
421 const __m128 vk22x0123 = _mm_load_ps(w + 92);
422 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123));
423
424 const __m128 vi23x0123 = _mm_loadu_ps(i23);
425 const __m128 vk23x0123 = _mm_load_ps(w + 96);
426 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi23x0123, vk23x0123));
427
428 const __m128 vi24x0123 = _mm_loadu_ps(i24);
429 const __m128 vk24x0123 = _mm_load_ps(w + 100);
430 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123));
431
432 // Add up all accumulators to vacc0123p0
433 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
434
435 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
436 vacc0123 = _mm_min_ps(vacc0123, vmax);
437
438 if (c & 2) {
439 _mm_storel_pi((__m64*) output, vacc0123);
440 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
441 output += 2;
442 }
443 if (c & 1) {
444 _mm_store_ss(output, vacc0123);
445 output += 1;
446 }
447 }
448
449 output = (float*) ((uintptr_t) output + output_increment);
450 } while (--output_width != 0);
451 }
452