fftw.py 18.7 KB
Newer Older
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
import warnings

import numpy as np
from d2o import distributed_data_object, STRATEGIES
from nifty.config import about, dependency_injector as gdi
import nifty.nifty_utilities as utilities
from transform import FFT

pyfftw = gdi.get('pyfftw')


class FFTW(FFT):

    """
        The pyfftw pendant of a fft object.
    """
    # The plan_dict stores the FFTWTransformInfo objects which correspond
    # to a certain set of (field_val, domain, codomain) sets.
    info_dict = {}

    # initialize the dictionary which stores the values from
    # get_centering_mask
    centering_mask_dict = {}

    def __init__(self, domain, codomain):
        self.domain = domain
        self.codomain = codomain

        if 'pyfftw' not in gdi:
            raise ImportError("The module pyfftw is needed but not available.")

        self.name = 'pyfftw'

        # Enable caching for pyfftw.interfaces
        pyfftw.interfaces.cache.enable()

    @classmethod
    def get_centering_mask(cls, to_center_input, dimensions_input,
                           offset_input=False):
        """
            Computes the mask, used to (de-)zerocenter domain and target
            fields.

            Parameters
            ----------
            to_center_input : tuple, list, numpy.ndarray
                A tuple of booleans which dimensions should be
                zero-centered.

            dimensions_input : tuple, list, numpy.ndarray
                A tuple containing the mask's desired shape.

            offset_input : int, boolean
                Specifies whether the zero-th dimension starts with an odd
                or and even index, i.e. if it is shifted.

            Returns
            -------
            result : np.ndarray
                A 1/-1-alternating mask.
        """
        # cast input
        to_center = np.array(to_center_input)
        dimensions = np.array(dimensions_input)

        # if none of the dimensions are zero centered, return a 1
        if np.all(to_center == 0):
            return 1

        if np.all(dimensions == np.array(1)) or \
                np.all(dimensions == np.array([1])):
            return dimensions
        # The dimensions of size 1 must be sorted out for computing the
        # centering_mask. The depth of the array will be restored in the
        # end.
        size_one_dimensions = []
        temp_dimensions = []
        temp_to_center = []
        for i in range(len(dimensions)):
            if dimensions[i] == 1:
                size_one_dimensions += [True]
            else:
                size_one_dimensions += [False]
                temp_dimensions += [dimensions[i]]
                temp_to_center += [to_center[i]]
        dimensions = np.array(temp_dimensions)
        to_center = np.array(temp_to_center)
        # cast the offset_input into the shape of to_center
        offset = np.zeros(to_center.shape, dtype=int)
        offset[0] = int(offset_input)
        # check for dimension match
        if to_center.size != dimensions.size:
            raise TypeError(
                'The length of the supplied lists does not match.')

        # build up the value memory
        # compute an identifier for the parameter set
        temp_id = tuple(
            (tuple(to_center), tuple(dimensions), tuple(offset)))
        if temp_id not in cls.centering_mask_dict:
            # use np.tile in order to stack the core alternation scheme
            # until the desired format is constructed.
            core = np.fromfunction(
                lambda *args: (-1) **
                (np.tensordot(to_center,
                              args +
                              offset.reshape(offset.shape +
                                             (1,) *
                                             (np.array(args).ndim - 1)),
                              1)),
                (2,) * to_center.size)
            # Cast the core to the smallest integers we can get
            core = core.astype(np.int8)

            centering_mask = np.tile(core, dimensions // 2)
            # for the dimensions of odd size corresponding slices must be
            # added
            for i in range(centering_mask.ndim):
                # check if the size of the certain dimension is odd or even
                if (dimensions % 2)[i] == 0:
                    continue
                # prepare the slice object
                temp_slice = (slice(None),) * i + (slice(-2, -1, 1),) + \
                             (slice(None),) * (centering_mask.ndim - 1 - i)
                # append the slice to the centering_mask
                centering_mask = np.append(centering_mask,
                                           centering_mask[temp_slice],
                                           axis=i)
            # Add depth to the centering_mask where the length of a
            # dimension was one
            temp_slice = ()
            for i in range(len(size_one_dimensions)):
                if size_one_dimensions[i]:
                    temp_slice += (None,)
                else:
                    temp_slice += (slice(None),)
            centering_mask = centering_mask[temp_slice]
            cls.centering_mask_dict[temp_id] = centering_mask
        return cls.centering_mask_dict[temp_id]

    @classmethod
    def _get_transform_info(cls, domain, codomain, local_shape,
                            local_offset_Q, is_local, transform_shape=None,
                            **kwargs):
        # generate a id-tuple which identifies the domain-codomain setting
        temp_id = (domain.__hash__() ^
                   (101 * codomain.__hash__()) ^
                   (211 * transform_shape.__hash__()))

        # generate the plan_and_info object if not already there
        if temp_id not in cls.info_dict:
            if is_local:
                cls.info_dict[temp_id] = FFTWLocalTransformInfo(
                    domain, codomain, local_shape,
                    local_offset_Q, **kwargs
                )
            else:
                cls.info_dict[temp_id] = FFTWMPITransfromInfo(
                    domain, codomain, local_shape,
                    local_offset_Q, transform_shape, **kwargs
                )

        return cls.info_dict[temp_id]

    def _apply_mask(self, val, mask, axes):
        """
            Apply centering mask to an array.

            Parameters
            ----------
            val: distributed_data_object or numpy.ndarray
                The value-array on which the mask should be applied.

            mask: numpy.ndarray
                The mask to be applied.

            axes: tuple
                The axes which are to be transformed.

            Returns
            -------
            distributed_data_object or np.nd_array
                Mask input array by multiplying it with the mask.
        """
        # reshape mask if necessary
        if axes:
            mask = mask.reshape(
                [y if x in axes else 1
                    for x, y in enumerate(val.shape)]
            )

        return val * mask

    def _atomic_mpi_transform(self, val, info, axes):
        # Apply codomain centering mask
        if reduce(lambda x, y: x+y, self.codomain.paradict['zerocenter']):
            temp_val = np.copy(val)
            val = self._apply_mask(temp_val, info.cmask_codomain, axes)

        p = info.plan
        # Load the value into the plan
        if p.has_input:
            p.input_array[:] = val
        # Execute the plan
        p()

        if p.has_output:
            result = p.output_array
        else:
            return None

        # Apply domain centering mask
        if reduce(lambda x, y: x+y, self.domain.paradict['zerocenter']):
            result = self._apply_mask(result, info.cmask_domain, axes)

        # Correct the sign if needed
        result *= info.sign

        return result

    def _local_transform(self, val, axes, **kwargs):
        ####
        # val must be numpy array or d2o with slicing distributor
        ###

        local_offset_Q = False
        try:
            local_val = val.get_local_data(copy=False)
            if axes is None or 0 in axes:
                local_offset_Q = val.distributor.local_shape[0] % 2
        except(AttributeError):
            local_val = val
        current_info = self._get_transform_info(self.domain,
                                                self.codomain,
                                                local_shape=local_val.shape,
                                                local_offset_Q=local_offset_Q,
                                                is_local=True,
                                                **kwargs)

        # Apply codomain centering mask
        if reduce(lambda x, y: x+y, self.codomain.paradict['zerocenter']):
            temp_val = np.copy(local_val)
            local_val = self._apply_mask(temp_val,
                                         current_info.cmask_codomain, axes)

        local_result = current_info.fftw_interface(
            local_val,
            axes=axes,
            planner_effort='FFTW_ESTIMATE'
        )

        # Apply domain centering mask
        if reduce(lambda x, y: x+y, self.domain.paradict['zerocenter']):
            local_result = self._apply_mask(local_result,
                                            current_info.cmask_domain, axes)

        # Correct the sign if needed
        if current_info.sign != 1:
            local_result *= current_info.sign

        try:
            # Create return object and insert results inplace
            return_val = val.copy_empty(global_shape=val.shape,
                                        dtype=self.codomain.dtype)
            return_val.set_local_data(data=local_result, copy=False)
        except(AttributeError):
            return_val = local_result

        return return_val

    def _repack_to_fftw_and_transform(self, val, axes, **kwargs):
        temp_val = val.copy_empty(distribution_strategy='fftw')
        about.warnings.cprint('WARNING: Repacking d2o to fftw \
                                distribution strategy')
        temp_val.set_full_data(val, copy=False)

        # Recursive call to transform
        result = self.transform(temp_val, axes, **kwargs)

        return_val = result.copy_empty(
            distribution_strategy=val.distribution_strategy)
        return_val.set_full_data(data=result, copy=False)

        return return_val

    def _mpi_transform(self, val, axes, **kwargs):

        if axes is None or 0 in axes:
            local_offset_list = np.cumsum(
                np.concatenate([[0, ], val.distributor.all_local_slices[:, 2]])
            )
            local_offset_Q = bool(
                local_offset_list[val.distributor.comm.rank] % 2)
        else:
            local_offset_Q = False

        return_val = val.copy_empty(global_shape=val.shape,
                                    dtype=self.codomain.dtype)

        # Extract local data
        local_val = val.get_local_data(copy=False)

        # Create temporary storage for slices
        temp_val = None

        # If axes tuple includes all axes, set it to None
        if axes is not None:
            if set(axes) == set(range(len(val.shape))):
                axes = None

        current_info = None
        for slice_list in utilities.get_slice_list(local_val.shape, axes):
            if slice_list == [slice(None, None)]:
                inp = local_val
            else:
                if temp_val is None:
                    temp_val = np.empty_like(local_val)
                inp = local_val[slice_list]

            # This is in order to make FFTW behave properly when slicing input
            # over MPI ranks when the input is 1-dimensional. The default
            # behaviour is to optimize to take advantage of byte-alignment,
            # which doesn't match the slicing strategy for multi-dimensional
            # data.
            original_shape = None
            if len(inp.shape) == 1:
                original_shape = inp.shape
                inp = inp.reshape(inp.shape[0], 1)

            if current_info is None:
                current_info = self._get_transform_info(
                    self.domain,
                    self.codomain,
                    local_shape=val.local_shape,
                    local_offset_Q=local_offset_Q,
                    is_local=False,
                    transform_shape=inp.shape,
                    **kwargs
                )

            with warnings.catch_warnings():
                warnings.simplefilter("ignore")
                result = self._atomic_mpi_transform(inp, current_info, axes)

            if result is None:
                temp_val = np.empty_like(local_val)
            elif slice_list == [slice(None, None)]:
                temp_val = result
            else:
                # Reverting to the original shape i.e. before the input was
                # augmented with 1 to make FFTW behave properly.
                if original_shape is not None:
                    result = result.reshape(original_shape)
                temp_val[slice_list] = result

        return_val.set_local_data(data=temp_val, copy=False)

        return return_val

    def transform(self, val, axes=None, **kwargs):
        """
            The pyfftw transform function.

            Parameters
            ----------
            val : distributed_data_object or numpy.ndarray
                The value-array of the field which is supposed to
                be transformed.

            domain : nifty.rg.nifty_rg.rg_space
                The domain of the space which should be transformed.

            codomain : nifty.rg.nifty_rg.rg_space
                The target into which the field should be transformed.

            axes: tuple, None
                The axes which should be transformed.

            **kwargs : *optional*
                Further kwargs are passed to the create_mpi_plan routine.

            Returns
            -------
            result : np.ndarray or distributed_data_object
                Fourier-transformed pendant of the input field.
        """
        # Check if the axes provided are valid given the shape
        if axes is not None and \
                not all(axis in range(len(val.shape)) for axis in axes):
            raise ValueError("ERROR: Provided axes does not match array shape")

        # If the input is a numpy array we transform it locally
        if not isinstance(val, distributed_data_object):
            # Cast to a np.ndarray
            temp_val = np.asarray(val)

            # local transform doesn't apply transforms inplace
            return_val = self._local_transform(temp_val, axes)
        else:
            if val.distribution_strategy in STRATEGIES['slicing']:
                if axes is None or 0 in axes:
                    if val.distribution_strategy != 'fftw':
                        return_val = \
                            self._repack_to_fftw_and_transform(
                                val, axes, **kwargs
                            )
                    else:
                        return_val = self._mpi_transform(
                            val, axes, **kwargs
                        )
                else:
                    return_val = self._local_transform(
                        val, axes, **kwargs
                    )
            else:
                return_val = self._repack_to_fftw_and_transform(
                    val, axes, **kwargs
                )

            # If domain is purely real, the result of the FFT is hermitian
            if self.domain.paradict['complexity'] == 0:
                return_val.hermitian = True

        return return_val


class FFTWTransformInfo(object):

    def __init__(self, domain, codomain, local_shape,
                 local_offset_Q, **kwargs):
        if pyfftw is None:
            raise ImportError("The module pyfftw is needed but not available.")

        self.cmask_domain = FFTW.get_centering_mask(
            domain.paradict['zerocenter'],
            local_shape,
            local_offset_Q)

        self.cmask_codomain = FFTW.get_centering_mask(
            codomain.paradict['zerocenter'],
            local_shape,
            local_offset_Q)

        # If both domain and codomain are zero-centered the result,
        # will get a global minus. Store the sign to correct it.
        self.sign = (-1) ** np.sum(np.array(domain.paradict['zerocenter']) *
                                   np.array(codomain.paradict['zerocenter']) *
                                   (np.array(domain.shape) // 2 % 2))

    @property
    def cmask_domain(self):
        return self._domain_centering_mask

    @cmask_domain.setter
    def cmask_domain(self, cmask):
        self._domain_centering_mask = cmask

    @property
    def cmask_codomain(self):
        return self._codomain_centering_mask

    @cmask_codomain.setter
    def cmask_codomain(self, cmask):
        self._codomain_centering_mask = cmask

    @property
    def sign(self):
        return self._sign

    @sign.setter
    def sign(self, sign):
        self._sign = sign


class FFTWLocalTransformInfo(FFTWTransformInfo):

    def __init__(self, domain, codomain, local_shape,
                 local_offset_Q, **kwargs):
        super(FFTWLocalTransformInfo, self).__init__(domain,
                                                     codomain,
                                                     local_shape,
                                                     local_offset_Q,
                                                     **kwargs)
        if codomain.harmonic:
            self._fftw_interface = pyfftw.interfaces.numpy_fft.fftn
        else:
            self._fftw_interface = pyfftw.interfaces.numpy_fft.ifftn

    @property
    def fftw_interface(self):
        return self._fftw_interface

    @fftw_interface.setter
    def fftw_interface(self, interface):
        about.warnings.cprint('WARNING: FFTWLocalTransformInfo fftw_interface \
                               cannot be modified')


class FFTWMPITransfromInfo(FFTWTransformInfo):

    def __init__(self, domain, codomain, local_shape,
                 local_offset_Q, transform_shape, **kwargs):
        super(FFTWMPITransfromInfo, self).__init__(domain,
                                                   codomain,
                                                   local_shape,
                                                   local_offset_Q,
                                                   **kwargs)
        self._plan = pyfftw.create_mpi_plan(
            input_shape=transform_shape,
            input_dtype='complex128',
            output_dtype='complex128',
            direction='FFTW_FORWARD' if codomain.harmonic else 'FFTW_BACKWARD',
            flags=["FFTW_ESTIMATE"],
            **kwargs
        )

    @property
    def plan(self):
        return self._plan

    @plan.setter
    def plan(self, plan):
        about.warnings.cprint('WARNING: FFTWMPITransfromInfo plan \
                               cannot be modified')