-
Notifications
You must be signed in to change notification settings - Fork 550
Expand file tree
/
Copy pathconvolve.cpp
More file actions
477 lines (419 loc) · 17.3 KB
/
convolve.cpp
File metadata and controls
477 lines (419 loc) · 17.3 KB
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
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
/*******************************************************
* Copyright (c) 2014, ArrayFire
* All rights reserved.
*
* This file is distributed under 3-clause BSD license.
* The complete license agreement can be obtained at:
* http://arrayfire.com/licenses/BSD-3-Clause
********************************************************/
#include <convolve.hpp>
#include <arith.hpp>
#include <backend.hpp>
#include <common/cast.hpp>
#include <common/err_common.hpp>
#include <common/half.hpp>
#include <common/tile.hpp>
#include <fftconvolve.hpp>
#include <handle.hpp>
#include <af/data.h>
#include <af/defines.h>
#include <af/dim4.hpp>
#include <af/ml.h>
#include <af/signal.h>
#include <cstdio>
using af::dim4;
using arrayfire::common::cast;
using arrayfire::common::half;
using detail::arithOp;
using detail::Array;
using detail::cdouble;
using detail::cfloat;
using detail::convolve;
using detail::intl;
using detail::schar;
using detail::uchar;
using detail::uint;
using detail::uintl;
using detail::ushort;
template<typename T, typename accT>
inline af_array convolve(const af_array &s, const af_array &f,
AF_BATCH_KIND kind, const int rank,
const bool expand) {
return getHandle(convolve<T, accT>(getArray<T>(s), castArray<accT>(f), kind,
rank, expand));
}
template<typename T, typename accT>
inline af_array convolve2(const af_array &s, const af_array &c_f,
const af_array &r_f, const bool expand) {
const Array<accT> colFilter = castArray<accT>(c_f);
const Array<accT> rowFilter = castArray<accT>(r_f);
const Array<accT> signal = castArray<accT>(s);
if (colFilter.isScalar() && rowFilter.isScalar()) {
Array<accT> colArray =
arrayfire::common::tile(colFilter, signal.dims());
Array<accT> rowArray =
arrayfire::common::tile(rowFilter, signal.dims());
Array<accT> filter =
arithOp<accT, af_mul_t>(colArray, rowArray, signal.dims());
return getHandle(cast<T, accT>(
arithOp<accT, af_mul_t>(signal, filter, signal.dims())));
}
ARG_ASSERT(2, colFilter.isVector());
ARG_ASSERT(3, rowFilter.isVector());
return getHandle(
convolve2<T, accT>(getArray<T>(s), colFilter, rowFilter, expand));
}
AF_BATCH_KIND identifyBatchKind(const int rank, const dim4 &sDims,
const dim4 &fDims) {
dim_t sn = sDims.ndims();
dim_t fn = fDims.ndims();
if (sn == rank && fn == rank) { return AF_BATCH_NONE; }
if (sn == rank && (fn > rank && fn <= AF_MAX_DIMS)) { return AF_BATCH_RHS; }
if ((sn > rank && sn <= AF_MAX_DIMS) && fn == rank) { return AF_BATCH_LHS; }
if ((sn > rank && sn <= AF_MAX_DIMS) && (fn > rank && fn <= AF_MAX_DIMS)) {
bool doesDimensionsMatch = true;
bool isInterleaved = true;
for (dim_t i = rank; i < AF_MAX_DIMS; i++) {
doesDimensionsMatch &= (sDims[i] == fDims[i]);
isInterleaved &=
(sDims[i] == 1 || fDims[i] == 1 || sDims[i] == fDims[i]);
}
if (doesDimensionsMatch) { return AF_BATCH_SAME; }
return (isInterleaved ? AF_BATCH_DIFF : AF_BATCH_UNSUPPORTED);
}
return AF_BATCH_UNSUPPORTED;
}
bool isFreqDomain(const int rank, const af_array &signal, const af_array filter,
af_conv_domain domain) {
if (domain == AF_CONV_FREQ) { return true; }
if (domain != AF_CONV_AUTO) { return false; }
const ArrayInfo &sInfo = getInfo(signal);
const ArrayInfo &fInfo = getInfo(filter);
const dim4 &sdims = sInfo.dims();
dim4 fdims = fInfo.dims();
if (identifyBatchKind(rank, sdims, fdims) == AF_BATCH_DIFF) { return true; }
int kbatch = 1;
for (int i = 3; i >= rank; i--) { kbatch *= fdims[i]; }
if (kbatch >= 10) { return true; }
if (rank == 1) {
if (fdims[0] > 128) { return true; }
}
if (rank == 2) {
// maximum supported size in 2D domain
if (fdims[0] > 17 || fdims[1] > 17) { return true; }
// Maximum supported non square size
if (fdims[0] != fdims[1] && fdims[0] > 5) { return true; }
}
if (rank == 3) {
if (fdims[0] > 5 || fdims[1] > 5 || fdims[2] > 5) { return true; }
}
return false;
}
af_err convolve(af_array *out, const af_array signal, const af_array filter,
const af_conv_mode mode, const int rank) {
try {
const ArrayInfo &sInfo = getInfo(signal);
const ArrayInfo &fInfo = getInfo(filter);
af_dtype stype = sInfo.getType();
dim4 sdims = sInfo.dims();
dim4 fdims = fInfo.dims();
if (fdims.ndims() == 0 || sdims.ndims() == 0) {
return af_retain_array(out, signal);
}
AF_BATCH_KIND convBT = identifyBatchKind(rank, sdims, fdims);
ARG_ASSERT(1,
(convBT != AF_BATCH_UNSUPPORTED && convBT != AF_BATCH_DIFF));
const bool expand = mode == AF_CONV_EXPAND;
af_array output;
switch (stype) {
case c32:
output = convolve<cfloat, cfloat>(signal, filter, convBT, rank,
expand);
break;
case c64:
output = convolve<cdouble, cdouble>(signal, filter, convBT,
rank, expand);
break;
case f32:
output = convolve<float, float>(signal, filter, convBT, rank,
expand);
break;
case f64:
output = convolve<double, double>(signal, filter, convBT, rank,
expand);
break;
case u32:
output =
convolve<uint, float>(signal, filter, convBT, rank, expand);
break;
case s32:
output =
convolve<int, float>(signal, filter, convBT, rank, expand);
break;
case u16:
output = convolve<ushort, float>(signal, filter, convBT, rank,
expand);
break;
case s16:
output = convolve<short, float>(signal, filter, convBT, rank,
expand);
break;
case u64:
output = convolve<uintl, float>(signal, filter, convBT, rank,
expand);
break;
case s64:
output =
convolve<intl, float>(signal, filter, convBT, rank, expand);
break;
case u8:
output = convolve<uchar, float>(signal, filter, convBT, rank,
expand);
break;
case s8:
output = convolve<schar, float>(signal, filter, convBT, rank,
expand);
break;
case b8:
output =
convolve<char, float>(signal, filter, convBT, rank, expand);
break;
default: TYPE_ERROR(1, stype);
}
std::swap(*out, output);
}
CATCHALL;
return AF_SUCCESS;
}
af_err af_convolve1(af_array *out, const af_array signal, const af_array filter,
const af_conv_mode mode, af_conv_domain domain) {
try {
if (isFreqDomain(1, signal, filter, domain)) {
return af_fft_convolve1(out, signal, filter, mode);
}
return convolve(out, signal, filter, mode, 1);
}
CATCHALL;
}
af_err af_convolve2(af_array *out, const af_array signal, const af_array filter,
const af_conv_mode mode, af_conv_domain domain) {
try {
if (getInfo(signal).dims().ndims() < 2 ||
getInfo(filter).dims().ndims() < 2) {
return af_convolve1(out, signal, filter, mode, domain);
}
if (isFreqDomain(2, signal, filter, domain)) {
return af_fft_convolve2(out, signal, filter, mode);
}
return convolve(out, signal, filter, mode, 2);
}
CATCHALL;
}
af_err af_convolve3(af_array *out, const af_array signal, const af_array filter,
const af_conv_mode mode, af_conv_domain domain) {
try {
if (getInfo(signal).dims().ndims() < 3 ||
getInfo(filter).dims().ndims() < 3) {
return af_convolve2(out, signal, filter, mode, domain);
}
if (isFreqDomain(3, signal, filter, domain)) {
return af_fft_convolve3(out, signal, filter, mode);
}
return convolve(out, signal, filter, mode, 3);
}
CATCHALL;
}
af_err af_convolve2_sep(af_array *out, const af_array col_filter,
const af_array row_filter, const af_array signal,
const af_conv_mode mode) {
try {
const ArrayInfo &sInfo = getInfo(signal);
const dim4 &sdims = sInfo.dims();
const af_dtype signalType = sInfo.getType();
ARG_ASSERT(1, (sdims.ndims() >= 2));
af_array output = 0;
const bool expand = mode == AF_CONV_EXPAND;
switch (signalType) {
case c32:
output = convolve2<cfloat, cfloat>(signal, col_filter,
row_filter, expand);
break;
case c64:
output = convolve2<cdouble, cdouble>(signal, col_filter,
row_filter, expand);
break;
case f32:
output = convolve2<float, float>(signal, col_filter, row_filter,
expand);
break;
case f64:
output = convolve2<double, double>(signal, col_filter,
row_filter, expand);
break;
case u32:
output = convolve2<uint, float>(signal, col_filter, row_filter,
expand);
break;
case s32:
output = convolve2<int, float>(signal, col_filter, row_filter,
expand);
break;
case u16:
output = convolve2<ushort, float>(signal, col_filter,
row_filter, expand);
break;
case s16:
output = convolve2<short, float>(signal, col_filter, row_filter,
expand);
break;
case u64:
output = convolve2<uintl, float>(signal, col_filter, row_filter,
expand);
break;
case s64:
output = convolve2<intl, float>(signal, col_filter, row_filter,
expand);
break;
case u8:
output = convolve2<uchar, float>(signal, col_filter, row_filter,
expand);
break;
case s8:
output = convolve2<schar, float>(signal, col_filter, row_filter,
expand);
break;
case b8:
output = convolve2<char, float>(signal, col_filter, row_filter,
expand);
break;
default: TYPE_ERROR(1, signalType);
}
std::swap(*out, output);
}
CATCHALL;
return AF_SUCCESS;
}
template<typename T>
inline af_array convolve2Strided(const af_array &s, const af_array &f,
const dim4 stride, const dim4 padding,
const dim4 dilation) {
return getHandle(convolve2<T>(getArray<T>(s), getArray<T>(f), stride,
padding, dilation));
}
af_err af_convolve2_nn(af_array *out, const af_array signal,
const af_array filter, const unsigned stride_dims,
const dim_t *strides, const unsigned padding_dims,
const dim_t *paddings, const unsigned dilation_dims,
const dim_t *dilations) {
try {
const ArrayInfo &sInfo = getInfo(signal);
const ArrayInfo &fInfo = getInfo(filter);
af::dim4 sDims = sInfo.dims();
af::dim4 fDims = fInfo.dims();
const af_dtype signalType = sInfo.getType();
dim4 stride(stride_dims, strides);
dim4 padding(padding_dims, paddings);
dim4 dilation(dilation_dims, dilations);
size_t stride_ndims = stride.ndims();
size_t padding_ndims = padding.ndims();
size_t dilation_ndims = dilation.ndims();
ARG_ASSERT(3, stride_ndims > 0 && stride_ndims <= 2);
ARG_ASSERT(5, padding_ndims >= 0 && padding_ndims <= 2);
ARG_ASSERT(7, dilation_ndims > 0 && dilation_ndims <= 2);
// assert number of features matches between signal and filter
DIM_ASSERT(1, sDims[2] == fDims[2]);
af_array output;
switch (signalType) {
case f32:
output = convolve2Strided<float>(signal, filter, stride,
padding, dilation);
break;
case f64:
output = convolve2Strided<double>(signal, filter, stride,
padding, dilation);
break;
case f16:
output = convolve2Strided<half>(signal, filter, stride, padding,
dilation);
break;
default: TYPE_ERROR(1, signalType);
}
std::swap(*out, output);
}
CATCHALL;
return AF_SUCCESS;
}
template<typename T>
af_array conv2GradCall(const af_array incoming_gradient,
const af_array original_signal,
const af_array original_filter,
const af_array convolved_output, const dim4 &stride,
const dim4 &padding, const dim4 &dilation,
af_conv_gradient_type grad_type) {
if (grad_type == AF_CONV_GRADIENT_FILTER) {
return getHandle(detail::conv2FilterGradient<T>(
getArray<T>(incoming_gradient), getArray<T>(original_signal),
getArray<T>(original_filter), getArray<T>(convolved_output), stride,
padding, dilation));
} else {
return getHandle(detail::conv2DataGradient<T>(
getArray<T>(incoming_gradient), getArray<T>(original_signal),
getArray<T>(original_filter), getArray<T>(convolved_output), stride,
padding, dilation));
}
}
af_err af_convolve2_gradient_nn(
af_array *out, const af_array incoming_gradient,
const af_array original_signal, const af_array original_filter,
const af_array convolved_output, const unsigned stride_dims,
const dim_t *strides, const unsigned padding_dims, const dim_t *paddings,
const unsigned dilation_dims, const dim_t *dilations,
af_conv_gradient_type grad_type) {
try {
const ArrayInfo &iinfo = getInfo(incoming_gradient);
const af::dim4 &iDims = iinfo.dims();
const ArrayInfo &sinfo = getInfo(original_signal);
af::dim4 sDims = sinfo.dims();
const ArrayInfo &finfo = getInfo(original_filter);
af::dim4 fDims = finfo.dims();
const ArrayInfo &oinfo = getInfo(convolved_output);
af::dim4 oDims = oinfo.dims();
DIM_ASSERT(1, iDims == oDims);
DIM_ASSERT(3, oDims[2] == fDims[3]);
DIM_ASSERT(3, oDims[3] == sDims[3]);
DIM_ASSERT(2, sDims[2] == fDims[2]);
af_array output;
af::dim4 stride(stride_dims, strides);
af::dim4 padding(padding_dims, paddings);
af::dim4 dilation(dilation_dims, dilations);
size_t stride_ndims = stride.ndims();
size_t padding_ndims = padding.ndims();
size_t dilation_ndims = dilation.ndims();
ARG_ASSERT(3, stride_ndims > 0 && stride_ndims <= 2);
ARG_ASSERT(5, padding_ndims >= 0 && padding_ndims <= 2);
ARG_ASSERT(7, dilation_ndims > 0 && dilation_ndims <= 2);
af_dtype type = oinfo.getType();
switch (type) {
case f32:
output = conv2GradCall<float>(
incoming_gradient, original_signal, original_filter,
convolved_output, stride, padding, dilation, grad_type);
break;
case f64:
output = conv2GradCall<double>(
incoming_gradient, original_signal, original_filter,
convolved_output, stride, padding, dilation, grad_type);
break;
case f16:
output = conv2GradCall<half>(
incoming_gradient, original_signal, original_filter,
convolved_output, stride, padding, dilation, grad_type);
break;
default: TYPE_ERROR(1, type);
}
// output array is pooled array
std::swap(output, *out);
}
CATCHALL;
return AF_SUCCESS;
}