1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * Gromacs 4.0 Copyright (c) 1991-2003
5 * David van der Spoel, Erik Lindahl, University of Groningen.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version 2
10 * of the License, or (at your option) any later version.
12 * To help us fund GROMACS development, we humbly ask that you cite
13 * the research papers on the package. Check out http://www.gromacs.org
16 * Gnomes, ROck Monsters And Chili Sauce
28 #include <mkl_service.h>
32 #include "gmx_fatal.h"
35 /* For MKL version (<10.0), we should define MKL_LONG. */
37 #define MKL_LONG long int
42 #define GMX_DFTI_PREC DFTI_DOUBLE
44 #define GMX_DFTI_PREC DFTI_SINGLE
47 /* Contents of the Intel MKL FFT fft datatype.
49 * Note that this is one of several possible implementations of gmx_fft_t.
51 * The MKL _API_ supports 1D,2D, and 3D transforms, including real-to-complex.
52 * Unfortunately the actual library implementation does not support 3D real
53 * transforms as of version 7.2, and versions before 7.0 don't support 2D real
54 * either. In addition, the multi-dimensional storage format for real data
55 * is not compatible with our padding.
57 * To work around this we roll our own 2D and 3D real-to-complex transforms,
58 * using separate X/Y/Z handles defined to perform (ny*nz), (nx*nz), and
59 * (nx*ny) transforms at once when necessary. To perform strided multiple
60 * transforms out-of-place (i.e., without padding in the last dimension)
61 * on the fly we also need to separate the forward and backward
62 * handles for real-to-complex/complex-to-real data permutation.
64 * This makes it necessary to define 3 handles for in-place FFTs, and 4 for
65 * the out-of-place transforms. Still, whenever possible we try to use
66 * a single 3D-transform handle instead.
68 * So, the handles are enumerated as follows:
70 * 1D FFT (real too): Index 0 is the handle for the entire FFT
71 * 2D complex FFT: Index 0 is the handle for the entire FFT
72 * 3D complex FFT: Index 0 is the handle for the entire FFT
73 * 2D, inplace real FFT: 0=FFTx, 1=FFTy handle
74 * 2D, ooplace real FFT: 0=FFTx, 1=real-to-complex FFTy, 2=complex-to-real FFTy
75 * 3D, inplace real FFT: 0=FFTx, 1=FFTy, 2=FFTz handle
76 * 3D, ooplace real FFT: 0=FFTx, 1=FFTy, 2=r2c FFTz, 3=c2r FFTz
78 * Intel people reading this: Learn from FFTW what a good interface looks like :-)
83 int ndim; /**< Number of dimensions in FFT */
84 int nx; /**< Length of X transform */
85 int ny; /**< Length of Y transform */
86 int nz; /**< Length of Z transform */
87 int real_fft; /**< 1 if real FFT, otherwise 0 */
88 DFTI_DESCRIPTOR * inplace[3]; /**< in-place FFT */
89 DFTI_DESCRIPTOR * ooplace[4]; /**< out-of-place FFT */
90 t_complex * work; /**< Enable out-of-place c2r FFT */
96 gmx_fft_init_1d(gmx_fft_t * pfft,
106 gmx_fatal(FARGS, "Invalid opaque FFT datatype pointer.");
111 if ( (fft = (gmx_fft_t)malloc(sizeof(struct gmx_fft))) == NULL)
116 /* Mark all handles invalid */
117 for (d = 0; d < 3; d++)
119 fft->inplace[d] = fft->ooplace[d] = NULL;
121 fft->ooplace[3] = NULL;
124 status = DftiCreateDescriptor(&fft->inplace[0], GMX_DFTI_PREC, DFTI_COMPLEX, 1, (MKL_LONG)nx);
128 status = DftiSetValue(fft->inplace[0], DFTI_PLACEMENT, DFTI_INPLACE);
133 status = DftiCommitDescriptor(fft->inplace[0]);
139 status = DftiCreateDescriptor(&fft->ooplace[0], GMX_DFTI_PREC, DFTI_COMPLEX, 1, (MKL_LONG)nx);
144 DftiSetValue(fft->ooplace[0], DFTI_PLACEMENT, DFTI_NOT_INPLACE);
149 DftiCommitDescriptor(fft->ooplace[0]);
155 gmx_fatal(FARGS, "Error initializing Intel MKL FFT; status=%d", status);
156 gmx_fft_destroy(fft);
172 gmx_fft_init_1d_real(gmx_fft_t * pfft,
182 gmx_fatal(FARGS, "Invalid opaque FFT datatype pointer.");
187 if ( (fft = (gmx_fft_t)malloc(sizeof(struct gmx_fft))) == NULL)
192 /* Mark all handles invalid */
193 for (d = 0; d < 3; d++)
195 fft->inplace[d] = fft->ooplace[d] = NULL;
197 fft->ooplace[3] = NULL;
199 status = DftiCreateDescriptor(&fft->inplace[0], GMX_DFTI_PREC, DFTI_REAL, 1, (MKL_LONG)nx);
203 status = DftiSetValue(fft->inplace[0], DFTI_PLACEMENT, DFTI_INPLACE);
208 status = DftiCommitDescriptor(fft->inplace[0]);
214 status = DftiCreateDescriptor(&fft->ooplace[0], GMX_DFTI_PREC, DFTI_REAL, 1, (MKL_LONG)nx);
219 status = DftiSetValue(fft->ooplace[0], DFTI_PLACEMENT, DFTI_NOT_INPLACE);
224 status = DftiCommitDescriptor(fft->ooplace[0]);
228 if (status == DFTI_UNIMPLEMENTED)
231 "The linked Intel MKL version (<6.0?) cannot do real FFTs.");
232 gmx_fft_destroy(fft);
239 gmx_fatal(FARGS, "Error initializing Intel MKL FFT; status=%d", status);
240 gmx_fft_destroy(fft);
256 gmx_fft_init_2d_real(gmx_fft_t * pfft,
269 gmx_fatal(FARGS, "Invalid opaque FFT datatype pointer.");
274 if ( (fft = (gmx_fft_t)malloc(sizeof(struct gmx_fft))) == NULL)
281 /* Mark all handles invalid */
282 for (d = 0; d < 3; d++)
284 fft->inplace[d] = fft->ooplace[d] = NULL;
286 fft->ooplace[3] = NULL;
288 /* Roll our own 2D real transform using multiple transforms in MKL,
289 * since the current MKL versions does not support our storage format,
290 * and all but the most recent don't even have 2D real FFTs.
294 status = DftiCreateDescriptor(&fft->inplace[0], GMX_DFTI_PREC, DFTI_COMPLEX, 1, (MKL_LONG)nx);
302 (DftiSetValue(fft->inplace[0], DFTI_PLACEMENT, DFTI_INPLACE) ||
303 DftiSetValue(fft->inplace[0], DFTI_NUMBER_OF_TRANSFORMS, nyc) ||
304 DftiSetValue(fft->inplace[0], DFTI_INPUT_DISTANCE, 1) ||
305 DftiSetValue(fft->inplace[0], DFTI_INPUT_STRIDES, stride) ||
306 DftiSetValue(fft->inplace[0], DFTI_OUTPUT_DISTANCE, 1) ||
307 DftiSetValue(fft->inplace[0], DFTI_OUTPUT_STRIDES, stride));
312 status = DftiCommitDescriptor(fft->inplace[0]);
315 /* Out-of-place X FFT */
318 status = DftiCreateDescriptor(&(fft->ooplace[0]), GMX_DFTI_PREC, DFTI_COMPLEX, 1, (MKL_LONG)nx);
327 (DftiSetValue(fft->ooplace[0], DFTI_PLACEMENT, DFTI_NOT_INPLACE) ||
328 DftiSetValue(fft->ooplace[0], DFTI_NUMBER_OF_TRANSFORMS, nyc) ||
329 DftiSetValue(fft->ooplace[0], DFTI_INPUT_DISTANCE, 1) ||
330 DftiSetValue(fft->ooplace[0], DFTI_INPUT_STRIDES, stride) ||
331 DftiSetValue(fft->ooplace[0], DFTI_OUTPUT_DISTANCE, 1) ||
332 DftiSetValue(fft->ooplace[0], DFTI_OUTPUT_STRIDES, stride));
337 status = DftiCommitDescriptor(fft->ooplace[0]);
344 status = DftiCreateDescriptor(&fft->inplace[1], GMX_DFTI_PREC, DFTI_REAL, 1, (MKL_LONG)ny);
353 (DftiSetValue(fft->inplace[1], DFTI_PLACEMENT, DFTI_INPLACE) ||
354 DftiSetValue(fft->inplace[1], DFTI_NUMBER_OF_TRANSFORMS, (MKL_LONG)nx) ||
355 DftiSetValue(fft->inplace[1], DFTI_INPUT_DISTANCE, 2*nyc) ||
356 DftiSetValue(fft->inplace[1], DFTI_INPUT_STRIDES, stride) ||
357 DftiSetValue(fft->inplace[1], DFTI_OUTPUT_DISTANCE, 2*nyc) ||
358 DftiSetValue(fft->inplace[1], DFTI_OUTPUT_STRIDES, stride) ||
359 DftiCommitDescriptor(fft->inplace[1]));
363 /* Out-of-place real-to-complex (affects output distance) Y FFT */
366 status = DftiCreateDescriptor(&fft->ooplace[1], GMX_DFTI_PREC, DFTI_REAL, 1, (MKL_LONG)ny);
375 (DftiSetValue(fft->ooplace[1], DFTI_PLACEMENT, DFTI_NOT_INPLACE) ||
376 DftiSetValue(fft->ooplace[1], DFTI_NUMBER_OF_TRANSFORMS, (MKL_LONG)nx) ||
377 DftiSetValue(fft->ooplace[1], DFTI_INPUT_DISTANCE, (MKL_LONG)ny) ||
378 DftiSetValue(fft->ooplace[1], DFTI_INPUT_STRIDES, stride) ||
379 DftiSetValue(fft->ooplace[1], DFTI_OUTPUT_DISTANCE, 2*nyc) ||
380 DftiSetValue(fft->ooplace[1], DFTI_OUTPUT_STRIDES, stride) ||
381 DftiCommitDescriptor(fft->ooplace[1]));
385 /* Out-of-place complex-to-real (affects output distance) Y FFT */
388 status = DftiCreateDescriptor(&fft->ooplace[2], GMX_DFTI_PREC, DFTI_REAL, 1, (MKL_LONG)ny);
397 (DftiSetValue(fft->ooplace[2], DFTI_PLACEMENT, DFTI_NOT_INPLACE) ||
398 DftiSetValue(fft->ooplace[2], DFTI_NUMBER_OF_TRANSFORMS, (MKL_LONG)nx) ||
399 DftiSetValue(fft->ooplace[2], DFTI_INPUT_DISTANCE, 2*nyc) ||
400 DftiSetValue(fft->ooplace[2], DFTI_INPUT_STRIDES, stride) ||
401 DftiSetValue(fft->ooplace[2], DFTI_OUTPUT_DISTANCE, (MKL_LONG)ny) ||
402 DftiSetValue(fft->ooplace[2], DFTI_OUTPUT_STRIDES, stride) ||
403 DftiCommitDescriptor(fft->ooplace[2]));
409 if ((fft->work = (t_complex *)malloc(sizeof(t_complex)*(nx*(ny/2+1)))) == NULL)
417 gmx_fatal(FARGS, "Error initializing Intel MKL FFT; status=%d", status);
418 gmx_fft_destroy(fft);
432 gmx_fft_1d(gmx_fft_t fft,
433 enum gmx_fft_direction dir,
437 int inplace = (in_data == out_data);
440 if ( (fft->real_fft == 1) || (fft->ndim != 1) ||
441 ((dir != GMX_FFT_FORWARD) && (dir != GMX_FFT_BACKWARD)) )
443 gmx_fatal(FARGS, "FFT plan mismatch - bad plan or direction.");
447 if (dir == GMX_FFT_FORWARD)
451 status = DftiComputeForward(fft->inplace[0], in_data);
455 status = DftiComputeForward(fft->ooplace[0], in_data, out_data);
462 status = DftiComputeBackward(fft->inplace[0], in_data);
466 status = DftiComputeBackward(fft->ooplace[0], in_data, out_data);
472 gmx_fatal(FARGS, "Error executing Intel MKL FFT.");
482 gmx_fft_1d_real(gmx_fft_t fft,
483 enum gmx_fft_direction dir,
487 int inplace = (in_data == out_data);
490 if ( (fft->real_fft != 1) || (fft->ndim != 1) ||
491 ((dir != GMX_FFT_REAL_TO_COMPLEX) && (dir != GMX_FFT_COMPLEX_TO_REAL)) )
493 gmx_fatal(FARGS, "FFT plan mismatch - bad plan or direction.");
497 if (dir == GMX_FFT_REAL_TO_COMPLEX)
501 status = DftiComputeForward(fft->inplace[0], in_data);
505 status = DftiComputeForward(fft->ooplace[0], in_data, out_data);
512 status = DftiComputeBackward(fft->inplace[0], in_data);
516 status = DftiComputeBackward(fft->ooplace[0], in_data, out_data);
522 gmx_fatal(FARGS, "Error executing Intel MKL FFT.");
531 gmx_fft_2d_real(gmx_fft_t fft,
532 enum gmx_fft_direction dir,
536 int inplace = (in_data == out_data);
539 if ( (fft->real_fft != 1) || (fft->ndim != 2) ||
540 ((dir != GMX_FFT_REAL_TO_COMPLEX) && (dir != GMX_FFT_COMPLEX_TO_REAL)) )
542 gmx_fatal(FARGS, "FFT plan mismatch - bad plan or direction.");
546 if (dir == GMX_FFT_REAL_TO_COMPLEX)
550 /* real-to-complex in Y dimension, in-place */
551 status = DftiComputeForward(fft->inplace[1], in_data);
553 /* complex-to-complex in X dimension, in-place */
556 status = DftiComputeForward(fft->inplace[0], in_data);
561 /* real-to-complex in Y dimension, in_data to out_data */
562 status = DftiComputeForward(fft->ooplace[1], in_data, out_data);
564 /* complex-to-complex in X dimension, in-place to out_data */
567 status = DftiComputeForward(fft->inplace[0], out_data);
573 /* prior implementation was incorrect. See fft.cpp unit test */
574 gmx_incons("Complex -> Real is not supported by MKL.");
579 gmx_fatal(FARGS, "Error executing Intel MKL FFT.");
587 gmx_fft_destroy(gmx_fft_t fft)
593 for (d = 0; d < 3; d++)
595 if (fft->inplace[d] != NULL)
597 DftiFreeDescriptor(&fft->inplace[d]);
599 if (fft->ooplace[d] != NULL)
601 DftiFreeDescriptor(&fft->ooplace[d]);
604 if (fft->ooplace[3] != NULL)
606 DftiFreeDescriptor(&fft->ooplace[3]);
608 if (fft->work != NULL)
616 void gmx_fft_cleanup()
624 #endif /* GMX_FFT_MKL */