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
478
479
480
481
482
483
484
485
486
487
488
489
490
491
|
// Copyright (C) 2016 - 2023 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
#pragma once
#ifndef FFTWTRANSFORM_H
#define FFTWTRANSFORM_H
#include "test_params.h"
#include <fftw3.h>
#include <vector>
// Function to return maximum error for float and double types.
//
// Following Schatzman (1996; Accuracy of the Discrete Fourier
// Transform and the Fast Fourier Transform), the shape of relative
// l_2 error vs length should look like
//
// epsilon * sqrt(log2(length)).
//
// The magic epsilon constants below were chosen so that we get a
// reasonable upper bound for (all of) our tests.
//
// For rocFFT, prime lengths result in the highest error. As such,
// the epsilons below are perhaps too loose for pow2 lengths; but they
// are appropriate for prime lengths.
template <typename Tfloat>
inline double type_epsilon();
template <>
inline double type_epsilon<_Float16>()
{
return half_epsilon;
}
template <>
inline double type_epsilon<float>()
{
return single_epsilon;
}
template <>
inline double type_epsilon<double>()
{
return double_epsilon;
}
// C++ traits to translate float->fftwf_complex and
// double->fftw_complex.
// The correct FFTW complex type can be accessed via, for example,
// using complex_t = typename fftw_complex_trait<Tfloat>::complex_t;
template <typename Tfloat>
struct fftw_trait;
template <>
struct fftw_trait<_Float16>
{
// fftw does not support half precision, so use single precision and convert
using fftw_complex_type = fftwf_complex;
using fftw_plan_type = fftwf_plan;
};
template <>
struct fftw_trait<float>
{
using fftw_complex_type = fftwf_complex;
using fftw_plan_type = fftwf_plan;
};
template <>
struct fftw_trait<double>
{
using fftw_complex_type = fftw_complex;
using fftw_plan_type = fftw_plan;
};
// Copies the half-precision input buffer to a single-precision
// buffer. Note that the input buffer is already sized like it's a
// single-precision buffer (but only half of it is filled), because
// we allocate a single-precision buffer for FFTW to plan with.
static hostbuf half_to_single_copy(const hostbuf& in)
{
auto out = in.copy();
auto in_begin = reinterpret_cast<const _Float16*>(in.data());
std::copy_n(in_begin, in.size() / sizeof(_Float16) / 2, reinterpret_cast<float*>(out.data()));
return out;
}
// converts a wider precision buffer to a narrower precision, in-place
template <typename TfloatIn, typename TfloatOut>
void narrow_precision_inplace(hostbuf& in)
{
// ensure we're actually shrinking the data
static_assert(sizeof(TfloatIn) > sizeof(TfloatOut));
auto readPtr = reinterpret_cast<const TfloatIn*>(in.data());
auto writePtr = reinterpret_cast<TfloatOut*>(in.data());
std::copy_n(readPtr, in.size() / sizeof(TfloatIn), writePtr);
in.shrink(in.size() / (sizeof(TfloatIn) / sizeof(TfloatOut)));
}
static void single_to_half_inplace(hostbuf& in)
{
narrow_precision_inplace<float, _Float16>(in);
}
// Template wrappers for real-valued FFTW allocators:
template <typename Tfloat>
inline Tfloat* fftw_alloc_real_type(size_t n);
template <>
inline float* fftw_alloc_real_type<float>(size_t n)
{
return fftwf_alloc_real(n);
}
template <>
inline double* fftw_alloc_real_type<double>(size_t n)
{
return fftw_alloc_real(n);
}
// Template wrappers for complex-valued FFTW allocators:
template <typename Tfloat>
inline typename fftw_trait<Tfloat>::fftw_complex_type* fftw_alloc_complex_type(size_t n);
template <>
inline typename fftw_trait<float>::fftw_complex_type* fftw_alloc_complex_type<float>(size_t n)
{
return fftwf_alloc_complex(n);
}
template <>
inline typename fftw_trait<double>::fftw_complex_type* fftw_alloc_complex_type<double>(size_t n)
{
return fftw_alloc_complex(n);
}
template <typename fftw_type>
inline fftw_type* fftw_alloc_type(size_t n);
template <>
inline float* fftw_alloc_type<float>(size_t n)
{
return fftw_alloc_real_type<float>(n);
}
template <>
inline double* fftw_alloc_type<double>(size_t n)
{
return fftw_alloc_real_type<double>(n);
}
template <>
inline fftwf_complex* fftw_alloc_type<fftwf_complex>(size_t n)
{
return fftw_alloc_complex_type<float>(n);
}
template <>
inline fftw_complex* fftw_alloc_type<fftw_complex>(size_t n)
{
return fftw_alloc_complex_type<double>(n);
}
template <>
inline rocfft_complex<float>* fftw_alloc_type<rocfft_complex<float>>(size_t n)
{
return (rocfft_complex<float>*)fftw_alloc_complex_type<float>(n);
}
template <>
inline rocfft_complex<double>* fftw_alloc_type<rocfft_complex<double>>(size_t n)
{
return (rocfft_complex<double>*)fftw_alloc_complex_type<double>(n);
}
// Template wrappers for FFTW plan executors:
template <typename Tfloat>
inline void fftw_execute_type(typename fftw_trait<Tfloat>::fftw_plan_type plan);
template <>
inline void fftw_execute_type<float>(typename fftw_trait<float>::fftw_plan_type plan)
{
return fftwf_execute(plan);
}
template <>
inline void fftw_execute_type<double>(typename fftw_trait<double>::fftw_plan_type plan)
{
return fftw_execute(plan);
}
// Template wrappers for FFTW plan destroyers:
template <typename Tfftw_plan>
inline void fftw_destroy_plan_type(Tfftw_plan plan);
template <>
inline void fftw_destroy_plan_type<fftwf_plan>(fftwf_plan plan)
{
return fftwf_destroy_plan(plan);
}
template <>
inline void fftw_destroy_plan_type<fftw_plan>(fftw_plan plan)
{
return fftw_destroy_plan(plan);
}
// Template wrappers for FFTW c2c planners:
template <typename Tfloat>
inline typename fftw_trait<Tfloat>::fftw_plan_type
fftw_plan_guru64_dft(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<Tfloat>::fftw_complex_type* in,
typename fftw_trait<Tfloat>::fftw_complex_type* out,
int sign,
unsigned flags);
template <>
inline typename fftw_trait<_Float16>::fftw_plan_type
fftw_plan_guru64_dft<_Float16>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<_Float16>::fftw_complex_type* in,
typename fftw_trait<_Float16>::fftw_complex_type* out,
int sign,
unsigned flags)
{
return fftwf_plan_guru64_dft(rank, dims, howmany_rank, howmany_dims, in, out, sign, flags);
}
template <>
inline typename fftw_trait<float>::fftw_plan_type
fftw_plan_guru64_dft<float>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<float>::fftw_complex_type* in,
typename fftw_trait<float>::fftw_complex_type* out,
int sign,
unsigned flags)
{
return fftwf_plan_guru64_dft(rank, dims, howmany_rank, howmany_dims, in, out, sign, flags);
}
template <>
inline typename fftw_trait<double>::fftw_plan_type
fftw_plan_guru64_dft<double>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<double>::fftw_complex_type* in,
typename fftw_trait<double>::fftw_complex_type* out,
int sign,
unsigned flags)
{
return fftw_plan_guru64_dft(rank, dims, howmany_rank, howmany_dims, in, out, sign, flags);
}
// Template wrappers for FFTW c2c executors:
template <typename Tfloat>
inline void fftw_plan_execute_c2c(typename fftw_trait<Tfloat>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out);
template <>
inline void fftw_plan_execute_c2c<_Float16>(typename fftw_trait<_Float16>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
// since FFTW does not natively support half precision, convert
// input to single, execute, then convert output back to half
auto in_single = half_to_single_copy(in.front());
fftwf_execute_dft(plan,
reinterpret_cast<fftwf_complex*>(in_single.data()),
reinterpret_cast<fftwf_complex*>(out.front().data()));
single_to_half_inplace(out.front());
}
template <>
inline void fftw_plan_execute_c2c<float>(typename fftw_trait<float>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftwf_execute_dft(plan,
reinterpret_cast<fftwf_complex*>(in.front().data()),
reinterpret_cast<fftwf_complex*>(out.front().data()));
}
template <>
inline void fftw_plan_execute_c2c<double>(typename fftw_trait<double>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftw_execute_dft(plan,
reinterpret_cast<fftw_complex*>(in.front().data()),
reinterpret_cast<fftw_complex*>(out.front().data()));
}
// Template wrappers for FFTW r2c planners:
template <typename Tfloat>
inline typename fftw_trait<Tfloat>::fftw_plan_type
fftw_plan_guru64_r2c(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
Tfloat* in,
typename fftw_trait<Tfloat>::fftw_complex_type* out,
unsigned flags);
template <>
inline typename fftw_trait<_Float16>::fftw_plan_type
fftw_plan_guru64_r2c<_Float16>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
_Float16* in,
typename fftw_trait<_Float16>::fftw_complex_type* out,
unsigned flags)
{
return fftwf_plan_guru64_dft_r2c(
rank, dims, howmany_rank, howmany_dims, reinterpret_cast<float*>(in), out, flags);
}
template <>
inline typename fftw_trait<float>::fftw_plan_type
fftw_plan_guru64_r2c<float>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
float* in,
typename fftw_trait<float>::fftw_complex_type* out,
unsigned flags)
{
return fftwf_plan_guru64_dft_r2c(rank, dims, howmany_rank, howmany_dims, in, out, flags);
}
template <>
inline typename fftw_trait<double>::fftw_plan_type
fftw_plan_guru64_r2c<double>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
double* in,
typename fftw_trait<double>::fftw_complex_type* out,
unsigned flags)
{
return fftw_plan_guru64_dft_r2c(rank, dims, howmany_rank, howmany_dims, in, out, flags);
}
// Template wrappers for FFTW r2c executors:
template <typename Tfloat>
inline void fftw_plan_execute_r2c(typename fftw_trait<Tfloat>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out);
template <>
inline void fftw_plan_execute_r2c<_Float16>(typename fftw_trait<float>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
// since FFTW does not natively support half precision, convert
// input to single, execute, then convert output back to half
auto in_single = half_to_single_copy(in.front());
fftwf_execute_dft_r2c(plan,
reinterpret_cast<float*>(in_single.data()),
reinterpret_cast<fftwf_complex*>(out.front().data()));
single_to_half_inplace(out.front());
}
template <>
inline void fftw_plan_execute_r2c<float>(typename fftw_trait<float>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftwf_execute_dft_r2c(plan,
reinterpret_cast<float*>(in.front().data()),
reinterpret_cast<fftwf_complex*>(out.front().data()));
}
template <>
inline void fftw_plan_execute_r2c<double>(typename fftw_trait<double>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftw_execute_dft_r2c(plan,
reinterpret_cast<double*>(in.front().data()),
reinterpret_cast<fftw_complex*>(out.front().data()));
}
// Template wrappers for FFTW c2r planners:
template <typename Tfloat>
inline typename fftw_trait<Tfloat>::fftw_plan_type
fftw_plan_guru64_c2r(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<Tfloat>::fftw_complex_type* in,
Tfloat* out,
unsigned flags);
template <>
inline typename fftw_trait<_Float16>::fftw_plan_type
fftw_plan_guru64_c2r<_Float16>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<_Float16>::fftw_complex_type* in,
_Float16* out,
unsigned flags)
{
return fftwf_plan_guru64_dft_c2r(
rank, dims, howmany_rank, howmany_dims, in, reinterpret_cast<float*>(out), flags);
}
template <>
inline typename fftw_trait<float>::fftw_plan_type
fftw_plan_guru64_c2r<float>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<float>::fftw_complex_type* in,
float* out,
unsigned flags)
{
return fftwf_plan_guru64_dft_c2r(rank, dims, howmany_rank, howmany_dims, in, out, flags);
}
template <>
inline typename fftw_trait<double>::fftw_plan_type
fftw_plan_guru64_c2r<double>(int rank,
const fftw_iodim64* dims,
int howmany_rank,
const fftw_iodim64* howmany_dims,
typename fftw_trait<double>::fftw_complex_type* in,
double* out,
unsigned flags)
{
return fftw_plan_guru64_dft_c2r(rank, dims, howmany_rank, howmany_dims, in, out, flags);
}
// Template wrappers for FFTW c2r executors:
template <typename Tfloat>
inline void fftw_plan_execute_c2r(typename fftw_trait<Tfloat>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out);
template <>
inline void fftw_plan_execute_c2r<_Float16>(typename fftw_trait<float>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
// since FFTW does not natively support half precision, convert
// input to single, execute, then convert output back to half
auto in_single = half_to_single_copy(in.front());
fftwf_execute_dft_c2r(plan,
reinterpret_cast<fftwf_complex*>(in_single.data()),
reinterpret_cast<float*>(out.front().data()));
single_to_half_inplace(out.front());
}
template <>
inline void fftw_plan_execute_c2r<float>(typename fftw_trait<float>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftwf_execute_dft_c2r(plan,
reinterpret_cast<fftwf_complex*>(in.front().data()),
reinterpret_cast<float*>(out.front().data()));
}
template <>
inline void fftw_plan_execute_c2r<double>(typename fftw_trait<double>::fftw_plan_type plan,
std::vector<hostbuf>& in,
std::vector<hostbuf>& out)
{
fftw_execute_dft_c2r(plan,
reinterpret_cast<fftw_complex*>(in.front().data()),
reinterpret_cast<double*>(out.front().data()));
}
#ifdef FFTW_HAVE_SPRINT_PLAN
// Template wrappers for FFTW print plan:
template <typename Tfloat>
inline char* fftw_sprint_plan(const typename fftw_trait<Tfloat>::fftw_plan_type plan);
template <>
inline char* fftw_sprint_plan<_Float16>(const typename fftw_trait<_Float16>::fftw_plan_type plan)
{
return fftwf_sprint_plan(plan);
}
template <>
inline char* fftw_sprint_plan<float>(const typename fftw_trait<float>::fftw_plan_type plan)
{
return fftwf_sprint_plan(plan);
}
template <>
inline char* fftw_sprint_plan<double>(const typename fftw_trait<double>::fftw_plan_type plan)
{
return fftw_sprint_plan(plan);
}
#endif
#endif
|
