def __init__(self, shape, in_dtype, out_dtype, batch=1, stream=None,
mode=0x01):
if np.isscalar(shape):
self.shape = (shape, )
else:
self.shape = shape
self.in_dtype = in_dtype
self.out_dtype = out_dtype
if batch <= 0:
raise ValueError('batch size must be greater than 0')
self.batch = batch
# Determine type of transformation:
if in_dtype == np.float32 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_R2C
self.fft_func = cufft.cufftExecR2C
elif in_dtype == np.complex64 and out_dtype == np.float32:
self.fft_type = cufft.CUFFT_C2R
self.fft_func = cufft.cufftExecC2R
elif in_dtype == np.complex64 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_C2C
self.fft_func = cufft.cufftExecC2C
elif in_dtype == np.float64 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_D2Z
self.fft_func = cufft.cufftExecD2Z
elif in_dtype == np.complex128 and out_dtype == np.float64:
self.fft_type = cufft.CUFFT_Z2D
self.fft_func = cufft.cufftExecZ2D
elif in_dtype == np.complex128 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_Z2Z
self.fft_func = cufft.cufftExecZ2Z
else:
raise ValueError('unsupported input/output type combination')
# Check for double precision support:
capability = misc.get_compute_capability(misc.get_current_device())
if capability < 1.3 and \
(misc.isdoubletype(in_dtype) or misc.isdoubletype(out_dtype)):
raise RuntimeError('double precision requires compute capability '
'>= 1.3 (you have %g)' % capability)
# Set up plan:
if len(self.shape) > 0:
n = np.asarray(self.shape, np.int32)
self.handle = cufft.cufftPlanMany(len(self.shape), n.ctypes.data,
None, 1, 0, None, 1, 0,
self.fft_type, self.batch)
else:
raise ValueError('invalid transform size')
# Set FFTW compatibility mode:
cufft.cufftSetCompatibilityMode(self.handle, mode)
# Associate stream with plan:
if stream != None:
cufft.cufftSetStream(self.handle, stream.handle)
def __init__(self, shape, in_dtype, out_dtype, batch=1, stream=None,
mode=0x01):
if np.isscalar(shape):
self.shape = (shape, )
else:
self.shape = shape
self.in_dtype = in_dtype
self.out_dtype = out_dtype
if batch <= 0:
raise ValueError('batch size must be greater than 0')
self.batch = batch
# Determine type of transformation:
if in_dtype == np.float32 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_R2C
self.fft_func = cufft.cufftExecR2C
elif in_dtype == np.complex64 and out_dtype == np.float32:
self.fft_type = cufft.CUFFT_C2R
self.fft_func = cufft.cufftExecC2R
elif in_dtype == np.complex64 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_C2C
self.fft_func = cufft.cufftExecC2C
elif in_dtype == np.float64 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_D2Z
self.fft_func = cufft.cufftExecD2Z
elif in_dtype == np.complex128 and out_dtype == np.float64:
self.fft_type = cufft.CUFFT_Z2D
self.fft_func = cufft.cufftExecZ2D
elif in_dtype == np.complex128 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_Z2Z
self.fft_func = cufft.cufftExecZ2Z
else:
raise ValueError('unsupported input/output type combination')
# Check for double precision support:
capability = misc.get_compute_capability(misc.get_current_device())
if capability < 1.3 and \
(misc.isdoubletype(in_dtype) or misc.isdoubletype(out_dtype)):
raise RuntimeError('double precision requires compute capability '
'>= 1.3 (you have %g)' % capability)
# Set up plan:
if len(self.shape) > 0:
n = np.asarray(self.shape, np.int32)
self.handle = cufft.cufftPlanMany(len(self.shape), n.ctypes.data,
None, 1, 0, None, 1, 0,
self.fft_type, self.batch)
else:
raise ValueError('invalid transform size')
# Set FFTW compatibility mode:
cufft.cufftSetCompatibilityMode(self.handle, mode)
# Associate stream with plan:
if stream != None:
cufft.cufftSetStream(self.handle, stream.handle)
def __init__(self, shape, in_dtype, out_dtype, batch=1, stream=None,
mode=0x01):
if np.isscalar(shape):
self.shape = (shape, )
else:
self.shape = shape
self.in_dtype = in_dtype
self.out_dtype = out_dtype
if batch <= 0:
raise ValueError('batch size must be greater than 0')
self.batch = batch
# Determine type of transformation:
if in_dtype == np.float32 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_R2C
self.fft_func = cufft.cufftExecR2C
elif in_dtype == np.complex64 and out_dtype == np.float32:
self.fft_type = cufft.CUFFT_C2R
self.fft_func = cufft.cufftExecC2R
elif in_dtype == np.complex64 and out_dtype == np.complex64:
self.fft_type = cufft.CUFFT_C2C
self.fft_func = cufft.cufftExecC2C
elif in_dtype == np.float64 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_D2Z
self.fft_func = cufft.cufftExecD2Z
elif in_dtype == np.complex128 and out_dtype == np.float64:
self.fft_type = cufft.CUFFT_Z2D
self.fft_func = cufft.cufftExecZ2D
elif in_dtype == np.complex128 and out_dtype == np.complex128:
self.fft_type = cufft.CUFFT_Z2Z
self.fft_func = cufft.cufftExecZ2Z
else:
raise ValueError('unsupported input/output type combination')
# Set up plan:
if len(self.shape) > 0:
n = np.asarray(self.shape, np.int32)
self.handle = cufft.cufftPlanMany(len(self.shape), n.ctypes.data,
None, 1, 0, None, 1, 0,
self.fft_type, self.batch)
else:
raise ValueError('invalid transform size')
# Set FFTW compatibility mode:
cufft.cufftSetCompatibilityMode(self.handle, mode)
# Associate stream with plan:
if stream != None:
cufft.cufftSetStream(self.handle, stream.handle)