File size: 14,999 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
// Auto-generated file. Do not edit!
//   Template: src/f32-dwconv/unipass-neon.c.in
//   Generator: tools/xngen
//
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/dwconv.h>


void xnn_f32_dwconv_minmax_ukernel_25p4c__neonfma_acc2(
    size_t channels,
    size_t output_width,
    const float** input,
    const float* weights,
    float* output,
    intptr_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)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
  const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
  do {
    const float* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const float*) ((uintptr_t) i0 + input_offset);
    }
    const float* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const float*) ((uintptr_t) i1 + input_offset);
    }
    const float* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const float*) ((uintptr_t) i2 + input_offset);
    }
    const float* i3 = input[3];
    assert(i3 != NULL);
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const float*) ((uintptr_t) i3 + input_offset);
    }
    const float* i4 = input[4];
    assert(i4 != NULL);
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const float*) ((uintptr_t) i4 + input_offset);
    }
    const float* i5 = input[5];
    assert(i5 != NULL);
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const float*) ((uintptr_t) i5 + input_offset);
    }
    const float* i6 = input[6];
    assert(i6 != NULL);
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const float*) ((uintptr_t) i6 + input_offset);
    }
    const float* i7 = input[7];
    assert(i7 != NULL);
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const float*) ((uintptr_t) i7 + input_offset);
    }
    const float* i8 = input[8];
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const float*) ((uintptr_t) i8 + input_offset);
    }
    const float* i9 = input[9];
    assert(i9 != NULL);
    if XNN_UNPREDICTABLE(i9 != zero) {
      i9 = (const float*) ((uintptr_t) i9 + input_offset);
    }
    const float* i10 = input[10];
    assert(i10 != NULL);
    if XNN_UNPREDICTABLE(i10 != zero) {
      i10 = (const float*) ((uintptr_t) i10 + input_offset);
    }
    const float* i11 = input[11];
    assert(i11 != NULL);
    if XNN_UNPREDICTABLE(i11 != zero) {
      i11 = (const float*) ((uintptr_t) i11 + input_offset);
    }
    const float* i12 = input[12];
    assert(i12 != NULL);
    if XNN_UNPREDICTABLE(i12 != zero) {
      i12 = (const float*) ((uintptr_t) i12 + input_offset);
    }
    const float* i13 = input[13];
    assert(i13 != NULL);
    if XNN_UNPREDICTABLE(i13 != zero) {
      i13 = (const float*) ((uintptr_t) i13 + input_offset);
    }
    const float* i14 = input[14];
    assert(i14 != NULL);
    if XNN_UNPREDICTABLE(i14 != zero) {
      i14 = (const float*) ((uintptr_t) i14 + input_offset);
    }
    const float* i15 = input[15];
    assert(i15 != NULL);
    if XNN_UNPREDICTABLE(i15 != zero) {
      i15 = (const float*) ((uintptr_t) i15 + input_offset);
    }
    const float* i16 = input[16];
    assert(i16 != NULL);
    if XNN_UNPREDICTABLE(i16 != zero) {
      i16 = (const float*) ((uintptr_t) i16 + input_offset);
    }
    const float* i17 = input[17];
    assert(i17 != NULL);
    if XNN_UNPREDICTABLE(i17 != zero) {
      i17 = (const float*) ((uintptr_t) i17 + input_offset);
    }
    const float* i18 = input[18];
    assert(i18 != NULL);
    if XNN_UNPREDICTABLE(i18 != zero) {
      i18 = (const float*) ((uintptr_t) i18 + input_offset);
    }
    const float* i19 = input[19];
    assert(i19 != NULL);
    if XNN_UNPREDICTABLE(i19 != zero) {
      i19 = (const float*) ((uintptr_t) i19 + input_offset);
    }
    const float* i20 = input[20];
    assert(i20 != NULL);
    if XNN_UNPREDICTABLE(i20 != zero) {
      i20 = (const float*) ((uintptr_t) i20 + input_offset);
    }
    const float* i21 = input[21];
    assert(i21 != NULL);
    if XNN_UNPREDICTABLE(i21 != zero) {
      i21 = (const float*) ((uintptr_t) i21 + input_offset);
    }
    const float* i22 = input[22];
    assert(i22 != NULL);
    if XNN_UNPREDICTABLE(i22 != zero) {
      i22 = (const float*) ((uintptr_t) i22 + input_offset);
    }
    const float* i23 = input[23];
    assert(i23 != NULL);
    if XNN_UNPREDICTABLE(i23 != zero) {
      i23 = (const float*) ((uintptr_t) i23 + input_offset);
    }
    const float* i24 = input[24];
    assert(i24 != NULL);
    if XNN_UNPREDICTABLE(i24 != zero) {
      i24 = (const float*) ((uintptr_t) i24 + input_offset);
    }

    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const float* w = weights;
    for (; c >= 4; c -= 4) {
      float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;


      const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
      const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123);

      const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
      const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
      float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);

      const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
      const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123);

      const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
      const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123);

      const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
      const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123);

      const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
      const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123);

      const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
      const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123);

      const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
      const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123);

      const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
      const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123);

      const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4;
      const float32x4_t vk9x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123);

      const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4;
      const float32x4_t vk10x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123);

      const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4;
      const float32x4_t vk11x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123);

      const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4;
      const float32x4_t vk12x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123);

      const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4;
      const float32x4_t vk13x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123);

      const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4;
      const float32x4_t vk14x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123);

      const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4;
      const float32x4_t vk15x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123);

      const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4;
      const float32x4_t vk16x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123);

      const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4;
      const float32x4_t vk17x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123);

      const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4;
      const float32x4_t vk18x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123);

      const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4;
      const float32x4_t vk19x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123);

      const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4;
      const float32x4_t vk20x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123);

      const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4;
      const float32x4_t vk21x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123);

      const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4;
      const float32x4_t vk22x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123);

      const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4;
      const float32x4_t vk23x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123);

      const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4;
      const float32x4_t vk24x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123);

      // Add up all accumulators to vacc0123p0
      vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);

      float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
      vacc0123 = vminq_f32(vacc0123, vmax);

      vst1q_f32(output, vacc0123); output += 4;
    }
    if XNN_UNLIKELY(c != 0) {
      float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;


      const float32x4_t vi0x0123 = vld1q_f32(i0);
      const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123);

      const float32x4_t vi1x0123 = vld1q_f32(i1);
      const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
      float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);

      const float32x4_t vi2x0123 = vld1q_f32(i2);
      const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123);

      const float32x4_t vi3x0123 = vld1q_f32(i3);
      const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123);

      const float32x4_t vi4x0123 = vld1q_f32(i4);
      const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123);

      const float32x4_t vi5x0123 = vld1q_f32(i5);
      const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123);

      const float32x4_t vi6x0123 = vld1q_f32(i6);
      const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123);

      const float32x4_t vi7x0123 = vld1q_f32(i7);
      const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123);

      const float32x4_t vi8x0123 = vld1q_f32(i8);
      const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123);

      const float32x4_t vi9x0123 = vld1q_f32(i9);
      const float32x4_t vk9x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123);

      const float32x4_t vi10x0123 = vld1q_f32(i10);
      const float32x4_t vk10x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123);

      const float32x4_t vi11x0123 = vld1q_f32(i11);
      const float32x4_t vk11x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123);

      const float32x4_t vi12x0123 = vld1q_f32(i12);
      const float32x4_t vk12x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123);

      const float32x4_t vi13x0123 = vld1q_f32(i13);
      const float32x4_t vk13x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123);

      const float32x4_t vi14x0123 = vld1q_f32(i14);
      const float32x4_t vk14x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123);

      const float32x4_t vi15x0123 = vld1q_f32(i15);
      const float32x4_t vk15x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123);

      const float32x4_t vi16x0123 = vld1q_f32(i16);
      const float32x4_t vk16x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123);

      const float32x4_t vi17x0123 = vld1q_f32(i17);
      const float32x4_t vk17x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123);

      const float32x4_t vi18x0123 = vld1q_f32(i18);
      const float32x4_t vk18x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123);

      const float32x4_t vi19x0123 = vld1q_f32(i19);
      const float32x4_t vk19x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123);

      const float32x4_t vi20x0123 = vld1q_f32(i20);
      const float32x4_t vk20x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123);

      const float32x4_t vi21x0123 = vld1q_f32(i21);
      const float32x4_t vk21x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123);

      const float32x4_t vi22x0123 = vld1q_f32(i22);
      const float32x4_t vk22x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123);

      const float32x4_t vi23x0123 = vld1q_f32(i23);
      const float32x4_t vk23x0123 = vld1q_f32(w); w += 4;
      vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123);

      const float32x4_t vi24x0123 = vld1q_f32(i24);
      const float32x4_t vk24x0123 = vld1q_f32(w); w += 4;
      vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123);

      // Add up all accumulators to vacc0123p0
      vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);

      float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
      vacc0123 = vminq_f32(vacc0123, vmax);

      float32x2_t vacc01 = vget_low_f32(vacc0123);
      if (c & 2) {
        vst1_f32(output, vacc01); output += 2;
        vacc01 = vget_high_f32(vacc0123);
      }
      if (c & 1) {
        vst1_lane_f32(output, vacc01, 0); output += 1;
      }
    }

    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}