1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-avx.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 <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_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_up16x25__fma3_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 __m256 vmax = _mm256_load_ps(params->avx.max);
33 const __m256 vmin = _mm256_load_ps(params->avx.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 >= 16; c -= 16) {
165 __m256 vacc01234567p0 = _mm256_load_ps(w);
166 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8);
167
168
169 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
170 const __m256 vi0x89ABCDEF = _mm256_loadu_ps(i0 + 8);
171 i0 += 16;
172
173 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
174 const __m256 vk0x89ABCDEF = _mm256_load_ps(w + 24);
175 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
176 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0);
177
178 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
179 const __m256 vi1x89ABCDEF = _mm256_loadu_ps(i1 + 8);
180 i1 += 16;
181
182 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
183 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40);
184 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
185 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF);
186
187 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
188 const __m256 vi2x89ABCDEF = _mm256_loadu_ps(i2 + 8);
189 i2 += 16;
190
191 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
192 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56);
193 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
194 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0);
195
196 const __m256 vi3x01234567 = _mm256_loadu_ps(i3);
197 const __m256 vi3x89ABCDEF = _mm256_loadu_ps(i3 + 8);
198 i3 += 16;
199
200 const __m256 vk3x01234567 = _mm256_load_ps(w + 64);
201 const __m256 vk3x89ABCDEF = _mm256_load_ps(w + 72);
202 vacc01234567p1 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1);
203 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp1);
204
205 const __m256 vi4x01234567 = _mm256_loadu_ps(i4);
206 const __m256 vi4x89ABCDEF = _mm256_loadu_ps(i4 + 8);
207 i4 += 16;
208
209 const __m256 vk4x01234567 = _mm256_load_ps(w + 80);
210 const __m256 vk4x89ABCDEF = _mm256_load_ps(w + 88);
211 vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);
212 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0);
213
214 const __m256 vi5x01234567 = _mm256_loadu_ps(i5);
215 const __m256 vi5x89ABCDEF = _mm256_loadu_ps(i5 + 8);
216 i5 += 16;
217
218 const __m256 vk5x01234567 = _mm256_load_ps(w + 96);
219 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104);
220 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1);
221 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp1);
222
223 const __m256 vi6x01234567 = _mm256_loadu_ps(i6);
224 const __m256 vi6x89ABCDEF = _mm256_loadu_ps(i6 + 8);
225 i6 += 16;
226
227 const __m256 vk6x01234567 = _mm256_load_ps(w + 112);
228 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120);
229 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);
230 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0);
231
232 const __m256 vi7x01234567 = _mm256_loadu_ps(i7);
233 const __m256 vi7x89ABCDEF = _mm256_loadu_ps(i7 + 8);
234 i7 += 16;
235
236 const __m256 vk7x01234567 = _mm256_load_ps(w + 128);
237 const __m256 vk7x89ABCDEF = _mm256_load_ps(w + 136);
238 vacc01234567p1 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1);
239 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi7x89ABCDEF, vk7x89ABCDEF, vacc89ABCDEFp1);
240
241 const __m256 vi8x01234567 = _mm256_loadu_ps(i8);
242 const __m256 vi8x89ABCDEF = _mm256_loadu_ps(i8 + 8);
243 i8 += 16;
244
245 const __m256 vk8x01234567 = _mm256_load_ps(w + 144);
246 const __m256 vk8x89ABCDEF = _mm256_load_ps(w + 152);
247 vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);
248 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi8x89ABCDEF, vk8x89ABCDEF, vacc89ABCDEFp0);
249
250 const __m256 vi9x01234567 = _mm256_loadu_ps(i9);
251 const __m256 vi9x89ABCDEF = _mm256_loadu_ps(i9 + 8);
252 i9 += 16;
253
254 const __m256 vk9x01234567 = _mm256_load_ps(w + 160);
255 const __m256 vk9x89ABCDEF = _mm256_load_ps(w + 168);
256 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1);
257 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi9x89ABCDEF, vk9x89ABCDEF, vacc89ABCDEFp1);
258
259 const __m256 vi10x01234567 = _mm256_loadu_ps(i10);
260 const __m256 vi10x89ABCDEF = _mm256_loadu_ps(i10 + 8);
261 i10 += 16;
262
263 const __m256 vk10x01234567 = _mm256_load_ps(w + 176);
264 const __m256 vk10x89ABCDEF = _mm256_load_ps(w + 184);
265 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);
266 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi10x89ABCDEF, vk10x89ABCDEF, vacc89ABCDEFp0);
267
268 const __m256 vi11x01234567 = _mm256_loadu_ps(i11);
269 const __m256 vi11x89ABCDEF = _mm256_loadu_ps(i11 + 8);
270 i11 += 16;
271
272 const __m256 vk11x01234567 = _mm256_load_ps(w + 192);
273 const __m256 vk11x89ABCDEF = _mm256_load_ps(w + 200);
274 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1);
275 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi11x89ABCDEF, vk11x89ABCDEF, vacc89ABCDEFp1);
276
277 const __m256 vi12x01234567 = _mm256_loadu_ps(i12);
278 const __m256 vi12x89ABCDEF = _mm256_loadu_ps(i12 + 8);
279 i12 += 16;
280
281 const __m256 vk12x01234567 = _mm256_load_ps(w + 208);
282 const __m256 vk12x89ABCDEF = _mm256_load_ps(w + 216);
283 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);
284 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi12x89ABCDEF, vk12x89ABCDEF, vacc89ABCDEFp0);
285
286 const __m256 vi13x01234567 = _mm256_loadu_ps(i13);
287 const __m256 vi13x89ABCDEF = _mm256_loadu_ps(i13 + 8);
288 i13 += 16;
289
290 const __m256 vk13x01234567 = _mm256_load_ps(w + 224);
291 const __m256 vk13x89ABCDEF = _mm256_load_ps(w + 232);
292 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1);
293 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi13x89ABCDEF, vk13x89ABCDEF, vacc89ABCDEFp1);
294
295 const __m256 vi14x01234567 = _mm256_loadu_ps(i14);
296 const __m256 vi14x89ABCDEF = _mm256_loadu_ps(i14 + 8);
297 i14 += 16;
298
299 const __m256 vk14x01234567 = _mm256_load_ps(w + 240);
300 const __m256 vk14x89ABCDEF = _mm256_load_ps(w + 248);
301 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);
302 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi14x89ABCDEF, vk14x89ABCDEF, vacc89ABCDEFp0);
303
304 const __m256 vi15x01234567 = _mm256_loadu_ps(i15);
305 const __m256 vi15x89ABCDEF = _mm256_loadu_ps(i15 + 8);
306 i15 += 16;
307
308 const __m256 vk15x01234567 = _mm256_load_ps(w + 256);
309 const __m256 vk15x89ABCDEF = _mm256_load_ps(w + 264);
310 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1);
311 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi15x89ABCDEF, vk15x89ABCDEF, vacc89ABCDEFp1);
312
313 const __m256 vi16x01234567 = _mm256_loadu_ps(i16);
314 const __m256 vi16x89ABCDEF = _mm256_loadu_ps(i16 + 8);
315 i16 += 16;
316
317 const __m256 vk16x01234567 = _mm256_load_ps(w + 272);
318 const __m256 vk16x89ABCDEF = _mm256_load_ps(w + 280);
319 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);
320 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi16x89ABCDEF, vk16x89ABCDEF, vacc89ABCDEFp0);
321
322 const __m256 vi17x01234567 = _mm256_loadu_ps(i17);
323 const __m256 vi17x89ABCDEF = _mm256_loadu_ps(i17 + 8);
324 i17 += 16;
325
326 const __m256 vk17x01234567 = _mm256_load_ps(w + 288);
327 const __m256 vk17x89ABCDEF = _mm256_load_ps(w + 296);
328 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1);
329 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi17x89ABCDEF, vk17x89ABCDEF, vacc89ABCDEFp1);
330
331 const __m256 vi18x01234567 = _mm256_loadu_ps(i18);
332 const __m256 vi18x89ABCDEF = _mm256_loadu_ps(i18 + 8);
333 i18 += 16;
334
335 const __m256 vk18x01234567 = _mm256_load_ps(w + 304);
336 const __m256 vk18x89ABCDEF = _mm256_load_ps(w + 312);
337 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);
338 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi18x89ABCDEF, vk18x89ABCDEF, vacc89ABCDEFp0);
339
340 const __m256 vi19x01234567 = _mm256_loadu_ps(i19);
341 const __m256 vi19x89ABCDEF = _mm256_loadu_ps(i19 + 8);
342 i19 += 16;
343
344 const __m256 vk19x01234567 = _mm256_load_ps(w + 320);
345 const __m256 vk19x89ABCDEF = _mm256_load_ps(w + 328);
346 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1);
347 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi19x89ABCDEF, vk19x89ABCDEF, vacc89ABCDEFp1);
348
349 const __m256 vi20x01234567 = _mm256_loadu_ps(i20);
350 const __m256 vi20x89ABCDEF = _mm256_loadu_ps(i20 + 8);
351 i20 += 16;
352
353 const __m256 vk20x01234567 = _mm256_load_ps(w + 336);
354 const __m256 vk20x89ABCDEF = _mm256_load_ps(w + 344);
355 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);
356 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi20x89ABCDEF, vk20x89ABCDEF, vacc89ABCDEFp0);
357
358 const __m256 vi21x01234567 = _mm256_loadu_ps(i21);
359 const __m256 vi21x89ABCDEF = _mm256_loadu_ps(i21 + 8);
360 i21 += 16;
361
362 const __m256 vk21x01234567 = _mm256_load_ps(w + 352);
363 const __m256 vk21x89ABCDEF = _mm256_load_ps(w + 360);
364 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1);
365 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi21x89ABCDEF, vk21x89ABCDEF, vacc89ABCDEFp1);
366
367 const __m256 vi22x01234567 = _mm256_loadu_ps(i22);
368 const __m256 vi22x89ABCDEF = _mm256_loadu_ps(i22 + 8);
369 i22 += 16;
370
371 const __m256 vk22x01234567 = _mm256_load_ps(w + 368);
372 const __m256 vk22x89ABCDEF = _mm256_load_ps(w + 376);
373 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);
374 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi22x89ABCDEF, vk22x89ABCDEF, vacc89ABCDEFp0);
375
376 const __m256 vi23x01234567 = _mm256_loadu_ps(i23);
377 const __m256 vi23x89ABCDEF = _mm256_loadu_ps(i23 + 8);
378 i23 += 16;
379
380 const __m256 vk23x01234567 = _mm256_load_ps(w + 384);
381 const __m256 vk23x89ABCDEF = _mm256_load_ps(w + 392);
382 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1);
383 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi23x89ABCDEF, vk23x89ABCDEF, vacc89ABCDEFp1);
384
385 const __m256 vi24x01234567 = _mm256_loadu_ps(i24);
386 const __m256 vi24x89ABCDEF = _mm256_loadu_ps(i24 + 8);
387 i24 += 16;
388
389 const __m256 vk24x01234567 = _mm256_load_ps(w + 400);
390 const __m256 vk24x89ABCDEF = _mm256_load_ps(w + 408);
391 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);
392 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi24x89ABCDEF, vk24x89ABCDEF, vacc89ABCDEFp0);
393
394 w += 416;
395
396 // Add up all accumulators to vacc0123456789ABCDEFp0
397 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
398 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1);
399
400 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
401 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
402 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
403 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
404
405 _mm256_storeu_ps(output, vacc01234567);
406 _mm256_storeu_ps(output + 8, vacc89ABCDEF);
407 output += 16;
408 }
409 for (; c >= 8; c -= 8) {
410 __m256 vacc01234567p0 = _mm256_load_ps(w);
411
412 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
413 i0 += 8;
414
415 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
416 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
417
418 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
419 i1 += 8;
420
421 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
422 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
423
424 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
425 i2 += 8;
426
427 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
428 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
429
430 const __m256 vi3x01234567 = _mm256_loadu_ps(i3);
431 i3 += 8;
432
433 const __m256 vk3x01234567 = _mm256_load_ps(w + 64);
434 vacc01234567p1 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1);
435
436 const __m256 vi4x01234567 = _mm256_loadu_ps(i4);
437 i4 += 8;
438
439 const __m256 vk4x01234567 = _mm256_load_ps(w + 80);
440 vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);
441
442 const __m256 vi5x01234567 = _mm256_loadu_ps(i5);
443 i5 += 8;
444
445 const __m256 vk5x01234567 = _mm256_load_ps(w + 96);
446 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1);
447
448 const __m256 vi6x01234567 = _mm256_loadu_ps(i6);
449 i6 += 8;
450
451 const __m256 vk6x01234567 = _mm256_load_ps(w + 112);
452 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);
453
454 const __m256 vi7x01234567 = _mm256_loadu_ps(i7);
455 i7 += 8;
456
457 const __m256 vk7x01234567 = _mm256_load_ps(w + 128);
458 vacc01234567p1 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1);
459
460 const __m256 vi8x01234567 = _mm256_loadu_ps(i8);
461 i8 += 8;
462
463 const __m256 vk8x01234567 = _mm256_load_ps(w + 144);
464 vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);
465
466 const __m256 vi9x01234567 = _mm256_loadu_ps(i9);
467 i9 += 8;
468
469 const __m256 vk9x01234567 = _mm256_load_ps(w + 160);
470 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1);
471
472 const __m256 vi10x01234567 = _mm256_loadu_ps(i10);
473 i10 += 8;
474
475 const __m256 vk10x01234567 = _mm256_load_ps(w + 176);
476 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);
477
478 const __m256 vi11x01234567 = _mm256_loadu_ps(i11);
479 i11 += 8;
480
481 const __m256 vk11x01234567 = _mm256_load_ps(w + 192);
482 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1);
483
484 const __m256 vi12x01234567 = _mm256_loadu_ps(i12);
485 i12 += 8;
486
487 const __m256 vk12x01234567 = _mm256_load_ps(w + 208);
488 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);
489
490 const __m256 vi13x01234567 = _mm256_loadu_ps(i13);
491 i13 += 8;
492
493 const __m256 vk13x01234567 = _mm256_load_ps(w + 224);
494 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1);
495
496 const __m256 vi14x01234567 = _mm256_loadu_ps(i14);
497 i14 += 8;
498
499 const __m256 vk14x01234567 = _mm256_load_ps(w + 240);
500 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);
501
502 const __m256 vi15x01234567 = _mm256_loadu_ps(i15);
503 i15 += 8;
504
505 const __m256 vk15x01234567 = _mm256_load_ps(w + 256);
506 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1);
507
508 const __m256 vi16x01234567 = _mm256_loadu_ps(i16);
509 i16 += 8;
510
511 const __m256 vk16x01234567 = _mm256_load_ps(w + 272);
512 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);
513
514 const __m256 vi17x01234567 = _mm256_loadu_ps(i17);
515 i17 += 8;
516
517 const __m256 vk17x01234567 = _mm256_load_ps(w + 288);
518 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1);
519
520 const __m256 vi18x01234567 = _mm256_loadu_ps(i18);
521 i18 += 8;
522
523 const __m256 vk18x01234567 = _mm256_load_ps(w + 304);
524 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);
525
526 const __m256 vi19x01234567 = _mm256_loadu_ps(i19);
527 i19 += 8;
528
529 const __m256 vk19x01234567 = _mm256_load_ps(w + 320);
530 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1);
531
532 const __m256 vi20x01234567 = _mm256_loadu_ps(i20);
533 i20 += 8;
534
535 const __m256 vk20x01234567 = _mm256_load_ps(w + 336);
536 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);
537
538 const __m256 vi21x01234567 = _mm256_loadu_ps(i21);
539 i21 += 8;
540
541 const __m256 vk21x01234567 = _mm256_load_ps(w + 352);
542 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1);
543
544 const __m256 vi22x01234567 = _mm256_loadu_ps(i22);
545 i22 += 8;
546
547 const __m256 vk22x01234567 = _mm256_load_ps(w + 368);
548 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);
549
550 const __m256 vi23x01234567 = _mm256_loadu_ps(i23);
551 i23 += 8;
552
553 const __m256 vk23x01234567 = _mm256_load_ps(w + 384);
554 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1);
555
556 const __m256 vi24x01234567 = _mm256_loadu_ps(i24);
557 i24 += 8;
558
559 const __m256 vk24x01234567 = _mm256_load_ps(w + 400);
560 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);
561
562 w += 8;
563
564 // Add up all accumulators to vacc01234567p0
565 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
566
567 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
568 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
569
570 _mm256_storeu_ps(output, vacc01234567);
571 output += 8;
572 }
573 if XNN_UNLIKELY(c != 0) {
574 assert(c >= 1);
575 assert(c <= 7);
576 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ¶ms->avx.mask_table[7 - c]);
577
578 __m256 vacc01234567p0 = _mm256_load_ps(w);
579
580 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);
581 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
582 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
583
584 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);
585 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
586 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
587
588 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);
589 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
590 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
591
592 const __m256 vi3x01234567 = _mm256_maskload_ps(i3, vmask);
593 const __m256 vk3x01234567 = _mm256_load_ps(w + 64);
594 vacc01234567p1 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1);
595
596 const __m256 vi4x01234567 = _mm256_maskload_ps(i4, vmask);
597 const __m256 vk4x01234567 = _mm256_load_ps(w + 80);
598 vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);
599
600 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask);
601 const __m256 vk5x01234567 = _mm256_load_ps(w + 96);
602 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1);
603
604 const __m256 vi6x01234567 = _mm256_maskload_ps(i6, vmask);
605 const __m256 vk6x01234567 = _mm256_load_ps(w + 112);
606 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);
607
608 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask);
609 const __m256 vk7x01234567 = _mm256_load_ps(w + 128);
610 vacc01234567p1 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1);
611
612 const __m256 vi8x01234567 = _mm256_maskload_ps(i8, vmask);
613 const __m256 vk8x01234567 = _mm256_load_ps(w + 144);
614 vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);
615
616 const __m256 vi9x01234567 = _mm256_maskload_ps(i9, vmask);
617 const __m256 vk9x01234567 = _mm256_load_ps(w + 160);
618 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1);
619
620 const __m256 vi10x01234567 = _mm256_maskload_ps(i10, vmask);
621 const __m256 vk10x01234567 = _mm256_load_ps(w + 176);
622 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);
623
624 const __m256 vi11x01234567 = _mm256_maskload_ps(i11, vmask);
625 const __m256 vk11x01234567 = _mm256_load_ps(w + 192);
626 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1);
627
628 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask);
629 const __m256 vk12x01234567 = _mm256_load_ps(w + 208);
630 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);
631
632 const __m256 vi13x01234567 = _mm256_maskload_ps(i13, vmask);
633 const __m256 vk13x01234567 = _mm256_load_ps(w + 224);
634 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1);
635
636 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask);
637 const __m256 vk14x01234567 = _mm256_load_ps(w + 240);
638 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);
639
640 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask);
641 const __m256 vk15x01234567 = _mm256_load_ps(w + 256);
642 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1);
643
644 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask);
645 const __m256 vk16x01234567 = _mm256_load_ps(w + 272);
646 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);
647
648 const __m256 vi17x01234567 = _mm256_maskload_ps(i17, vmask);
649 const __m256 vk17x01234567 = _mm256_load_ps(w + 288);
650 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1);
651
652 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask);
653 const __m256 vk18x01234567 = _mm256_load_ps(w + 304);
654 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);
655
656 const __m256 vi19x01234567 = _mm256_maskload_ps(i19, vmask);
657 const __m256 vk19x01234567 = _mm256_load_ps(w + 320);
658 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1);
659
660 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask);
661 const __m256 vk20x01234567 = _mm256_load_ps(w + 336);
662 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);
663
664 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask);
665 const __m256 vk21x01234567 = _mm256_load_ps(w + 352);
666 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1);
667
668 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask);
669 const __m256 vk22x01234567 = _mm256_load_ps(w + 368);
670 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);
671
672 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask);
673 const __m256 vk23x01234567 = _mm256_load_ps(w + 384);
674 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1);
675
676 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask);
677 const __m256 vk24x01234567 = _mm256_load_ps(w + 400);
678 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);
679
680 // Add up all accumulators to vacc01234567p0
681 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
682
683 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
684 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
685
686 __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
687 if (c & 4) {
688 _mm_storeu_ps(output, vacc0123);
689 vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
690 output += 4;
691 }
692 if (c & 2) {
693 _mm_storel_pi((__m64*) output, vacc0123);
694 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
695 output += 2;
696 }
697 if (c & 1) {
698 _mm_store_ss(output, vacc0123);
699 output += 1;
700 }
701 }
702
703 output = (float*) ((uintptr_t) output + output_increment);
704 } while (--output_width != 0);
705 }
706