def prepare_video_cuda():
# Things go *a lot* faster when you have the memory structures pre-allocated
cs['plan1'] = fft.Plan(blocklen, np.float32, np.complex64)
cs['plan1i'] = fft.Plan(blocklen, np.complex64, np.complex64)
cs['fft1_out'] = gpuarray.empty((blocklen // 2) + 1, np.complex64)
cs['filtered1'] = gpuarray.empty(blocklen, np.complex64)
cs['fm_demod'] = gpuarray.empty(blocklen, np.float32)
cs['postlpf'] = gpuarray.empty(blocklen, np.float32)
cs['fft2_out'] = gpuarray.empty((blocklen // 2) + 1, np.complex64)
cs['clipped_gpu'] = gpuarray.empty(blocklen, np.uint16)
cs['plan2'] = fft.Plan(blocklen, np.float32, np.complex64)
cs['plan2i'] = fft.Plan(blocklen, np.complex64, np.float32)
# CUDA functions. The fewer setups we need, the faster it goes.
cs['doclamp16'] = mod.get_function("clamp16")
cs['doanglediff'] = mod.get_function("anglediff")
# GPU-stored frequency-fomain filters
cs['filt_post'] = FFTtoGPU(SysParams['fft_post'])
cs['filt_video'] = FFTtoGPU(SysParams['fft_video'])
cs['filt_video_inner'] = FFTtoGPU(SysParams['fft_video_inner'])
def prepare_audio_cuda():
cs['plan1'] = fft.Plan(blocklen, np.float32, np.complex64)
cs['plan1i'] = fft.Plan(ablocklen, np.complex64, np.complex64)
cs['fft1_out'] = gpuarray.empty(blocklen, np.complex64)
cs['ifft1_out'] = gpuarray.empty(ablocklen, np.complex64)
cs['fm_left'] = gpuarray.empty(ablocklen, np.complex64)
cs['fm_right'] = gpuarray.empty(ablocklen, np.complex64)
cs['left_clipped'] = gpuarray.empty(ablocklen, np.float32)
cs['right_clipped'] = gpuarray.empty(ablocklen, np.float32)
cs['left_fft1'] = gpuarray.empty(blocklen // 2 + 1, np.complex64)
cs['right_fft1'] = gpuarray.empty(blocklen // 2 + 1, np.complex64)
cs['left_fft2'] = gpuarray.empty(ablocklen // 2 + 1, np.complex64)
cs['right_fft2'] = gpuarray.empty(ablocklen // 2 + 1, np.complex64)
cs['left_out'] = gpuarray.empty(ablocklen, np.float32)
cs['right_out'] = gpuarray.empty(ablocklen, np.float32)
cs['plan2'] = fft.Plan(ablocklen, np.float32, np.complex64)
cs['plan2i'] = fft.Plan(ablocklen, np.complex64, np.float32)
cs['outlen'] = outlen = ablocklen // 20
cs['scaledout'] = gpuarray.empty(outlen * 2, np.float32)
cs['left_scaledout'] = gpuarray.empty(outlen, np.float32)
cs['right_scaledout'] = gpuarray.empty(outlen, np.float32)
cs['doanglediff_mac'] = mod.get_function("anglediff_mac")
cs['doaudioscale'] = mod.get_function("audioscale")
cs['filt_audiolpf'] = FFTtoGPU(SysParams['fft_audiolpf'])
cs['filt_audio_left'] = FFTtoGPU(SysParams['fft_audio_left'])
cs['filt_audio_right'] = FFTtoGPU(SysParams['fft_audio_right'])
def cufft_conv(x, y):
x = x.astype(np.complex64)
y = y.astype(np.complex64)
if (x.shape != y.shape):
return -1
plan = fft.Plan(x.shape, np.complex64, np.complex64)
inverse_plan = fft.Plan(x.shape, np.complex64, np.complex64)
x_gpu = gpuarray.to_gpu(x)
y_gpu = gpuarray.to_gpu(y)
x_fft = gpuarray.empty_like(x_gpu, dtype=np.complex64)
y_fft = gpuarray.empty_like(y_gpu, dtype=np.complex64)
out_gpu = gpuarray.empty_like(x_gpu, dtype=np.complex64)
fft.fft(x_gpu, x_fft, plan)
fft.fft(y_gpu, y_fft, plan)
linalg.multiply(x_fft, y_fft, overwrite=True)
fft.ifft(y_fft, out_gpu, inverse_plan, scale=True)
conv_out = out_gpu.get()
x_gpu.gpudata.free()
y_gpu.gpudata.free()
x_fft.gpudata.free()
y_fft.gpudata.free()
out_gpu.gpudata.free()
return conv_out
def propagate_eager(self, wavelength, wavefront):
"""
'Not-Too-Good' version of the propagation on the GPU (lots of Memory issues...)
Remove in the future
:param wavelength:
:param wavefront:
:return:
"""
N = self.N_PIX
# free, total = cuda.mem_get_info()
free, total = cuda.mem_get_info()
print("Free: %.2f percent" % (free / total * 100))
# Pupil Plane -> Image Slicer
complex_pupil = self.pupil_masks[wavelength] * np.exp(
1j * 2 * np.pi * self.pupil_masks[wavelength] / wavelength)
complex_pupil_gpu = gpuarray.to_gpu(
np.asarray(complex_pupil, np.complex64))
plan = cu_fft.Plan(complex_pupil_gpu.shape, np.complex64, np.complex64)
cu_fft.fft(complex_pupil_gpu, complex_pupil_gpu, plan, scale=True)
# Add N_slices copies to be Masked
complex_slicer_cpu = complex_pupil_gpu.get()
complex_pupil_gpu.gpudata.free()
free, total = cuda.mem_get_info()
print("*Free: %.2f percent" % (free / total * 100))
complex_slicer_cpu = np.stack([complex_slicer_cpu] * self.N_slices)
complex_slicer_gpu = gpuarray.to_gpu(complex_slicer_cpu)
slicer_masks_gpu = gpuarray.to_gpu(self.slicer_masks_fftshift)
clinalg.multiply(slicer_masks_gpu, complex_slicer_gpu, overwrite=True)
slicer_masks_gpu.gpudata.free()
free, total = cuda.mem_get_info()
print("**Free: %.2f percent" % (free / total * 100))
# Slicer -> Pupil Mirror
plan = cu_fft.Plan((N, N), np.complex64, np.complex64, self.N_slices)
cu_fft.ifft(complex_slicer_gpu, complex_slicer_gpu, plan, scale=True)
mirror_mask_gpu = gpuarray.to_gpu(self.pupil_mirror_masks_fft)
clinalg.multiply(mirror_mask_gpu, complex_slicer_gpu, overwrite=True)
# Pupil Mirror -> Slits
cu_fft.fft(complex_slicer_gpu, complex_slicer_gpu, plan)
slits = complex_slicer_gpu.get()
complex_slicer_gpu.gpudata.free()
mirror_mask_gpu.gpudata.free()
slit = fftshift(np.sum((np.abs(slits))**2, axis=0))
free, total = cuda.mem_get_info()
print("***Free: %.2f percent" % (free / total * 100))
return slit
def __init__(self, nx, ny):
shapeX = [ny, nx]
shapeK = [ny, nx // 2 + 1]
self.shapeX = shapeX
self.shapeK = shapeK
self.fftplan = skfft.Plan(self.shapeX, np.float32, np.complex64)
self.ifftplan = skfft.Plan(self.shapeX, np.complex64, np.float32)
self.coef_norm = nx * ny
def __init__(self, mesh, context=None):
'''
Args:
mesh The mesh on which the solver will operate. The dimensionality
is deducted from mesh.dimension
'''
# create the mesh grid and compute the greens function on it
self.mesh = mesh
self._context = context
mesh_shape = self.mesh.shape # nz, ny, (nx)
mesh_shape2 = [2*n for n in mesh_shape] # 2*nz, 2*ny, (2*nx)
mesh_distances = list(reversed(self.mesh.distances)) #dz, dy, dx
self.fgreentr = gpuarray.empty(mesh_shape2,
dtype=np.complex128)
self.tmpspace = gpuarray.zeros_like(self.fgreentr)
sizeof_complex = np.dtype(np.complex128).itemsize
# dimensionality function dispatch
dim = self.mesh.dimension
self._fgreen = getattr(self, '_fgreen' + str(dim) + 'd')
self._mirror = getattr(self, '_mirror' + str(dim) + 'd')
copy_fn = {'3d' : get_Memcpy3D_d2d, '2d': get_Memcpy2D_d2d}
memcpy_nd = copy_fn[str(dim) + 'd']
dim_args = self.mesh.shape
self._cpyrho2tmp = memcpy_nd(
src=None, dst=self.tmpspace, # None because src(rho) not yet known
src_pitch=self.mesh.nx*sizeof_complex,
dst_pitch=2*self.mesh.nx*sizeof_complex,
dim_args=dim_args,
itemsize=np.dtype(np.complex128).itemsize,
src_height=self.mesh.ny,
dst_height=2*self.mesh.ny)
self._cpytmp2rho = memcpy_nd(
src=self.tmpspace, dst=None, # None because dst(rho) not yet know
src_pitch=2*self.mesh.nx*sizeof_complex,
dst_pitch=self.mesh.nx*sizeof_complex,
dim_args=dim_args,
itemsize=np.dtype(np.complex128).itemsize,
src_height=2*self.mesh.ny,
dst_height=self.mesh.ny)
mesh_arr = [-mesh_distances[i]/2 + np.arange(mesh_shape[i]+1)
* mesh_distances[i]
for i in xrange(self.mesh.dimension)
]
# mesh_arr is [mz, my, mx]
mesh_grids = np.meshgrid(*mesh_arr, indexing='ij')
fgreen = self._fgreen(*mesh_grids)
fgreen = self._mirror(fgreen)
self.plan_forward = cu_fft.Plan(self.tmpspace.shape, in_dtype=np.complex128,
out_dtype=np.complex128)
self.plan_backward = cu_fft.Plan(self.tmpspace.shape, in_dtype=np.complex128,
out_dtype=np.complex128)
cu_fft.fft(gpuarray.to_gpu(fgreen), self.fgreentr, plan=self.plan_forward)
def set_focal_series(self, fs, defoci=None):
"""
Setting the focal series also sets important parameters as the padding
and defocus values.
Parameters
----------
fs : Hyperspy Image
Mandatory! A focal series with all the appropriate parameters. Padding
is read from metadata.
defoci : array
Optional, use it to explicitly set defoci values. Must have a dimension
equal to the navigation axis of fs.
"""
# Real space parameters
Nz, Ny, Nx = fs.data.shape
if defoci is None:
defoci = fs.axes_manager.navigation_axes[0].axis
Nz = len(defoci)
self.zdim = defoci
self.k2 = fs.get_fourier_space()
self.Nz = Nz
# In the middle of the focal series is the reference plane
Nz_half = np.ceil(Nz / 2).astype('int') - 1
# Pad mask
Npy, Npx = fs.metadata.Signal.pad_tuple
mask = np.zeros((Nz, Ny, Nx), dtype=np.bool)
mask[:, Npy[0]:(Ny - Npy[1]), Npx[0]:(Nx - Npx[1])] = True
# Set some parameters
self.new_wave = fs._get_signal_signal()
self.new_wave.metadata = fs.metadata.deepcopy()
self.shape = (Nz, Ny, Nx)
# Allocate things ...
if self.using_gpu:
# ... in GPU!
self.Iexp = to_gpu_f(fs.data)
self.mask = to_gpu_b(mask)
# - The plans for FFT
self.pft3dcc = cu_fft.Plan((Ny, Nx), np.complex64, np.complex64,
Nz)
self.pft2dcc = cu_fft.Plan((Ny, Nx), np.complex64, np.complex64)
else:
# ... in CPU!
self.Iexp = fs.data.astype('complex64')
self.mask = mask
def filter(self):
import pycuda.gpuarray as gpuarray
import skcuda.fft as cu_fft
import skcuda.linalg as linalg
import pycuda.driver as cuda
from pycuda.tools import make_default_context
cuda.init()
context = make_default_context()
device = context.get_device()
signal = self.series[0]
window = self.series[1]
linalg.init()
nfft = determine_size(len(signal) + len(window) - 1)
# Move data to GPU
sig_zero_pad = np.zeros(nfft, dtype=self.precision['float'])
win_zero_pad = np.zeros(nfft, dtype=self.precision['float'])
sig_gpu = gpuarray.zeros(sig_zero_pad.shape,
dtype=self.precision['float'])
win_gpu = gpuarray.zeros(win_zero_pad.shape,
dtype=self.precision['float'])
sig_zero_pad[0:len(signal)] = signal
win_zero_pad[0:len(window)] = window
sig_gpu.set(sig_zero_pad)
win_gpu.set(win_zero_pad)
# Plan forwards
sig_fft_gpu = gpuarray.zeros(nfft, dtype=self.precision['complex'])
win_fft_gpu = gpuarray.zeros(nfft, dtype=self.precision['complex'])
sig_plan_forward = cu_fft.Plan(sig_fft_gpu.shape,
self.precision['float'],
self.precision['complex'])
win_plan_forward = cu_fft.Plan(win_fft_gpu.shape,
self.precision['float'],
self.precision['complex'])
cu_fft.fft(sig_gpu, sig_fft_gpu, sig_plan_forward)
cu_fft.fft(win_gpu, win_fft_gpu, win_plan_forward)
# Convolve
out_fft = linalg.multiply(sig_fft_gpu, win_fft_gpu, overwrite=True)
linalg.scale(2.0, out_fft)
# Plan inverse
out_gpu = gpuarray.zeros_like(out_fft)
plan_inverse = cu_fft.Plan(out_fft.shape, self.precision['complex'],
self.precision['complex'])
cu_fft.ifft(out_fft, out_gpu, plan_inverse, True)
out_np = np.zeros(len(out_gpu), self.precision['complex'])
out_gpu.get(out_np)
context.pop()
return out_np
def cuda_efftn(H, axes, forward):
hShape = H.shape
hDim = len(hShape)
fftDim = len(axes)
# Reshape 'axes' to be the array's end dimensions and ensure contiguity
H = np.ascontiguousarray(
np.moveaxis(H, axes, np.arange(hDim - fftDim, hDim, 1)))
# Calculate number of batches
batchSize = 1
for i in range(hDim - fftDim):
batchSize *= H.shape[i]
# Reshape to accomodate batching
H = np.reshape(
H,
(batchSize, H.shape[hDim - 3], H.shape[hDim - 2], H.shape[hDim - 1]))
# Pass array to the GPU and perform iFFT on each batch
H_gpu = gpuarray.to_gpu(H)
plan = skfft.Plan(H_gpu.shape[1:fftDim + 1], H.dtype, H.dtype,
H_gpu.shape[0])
if forward:
skfft.fft(H_gpu, H_gpu, plan)
else:
skfft.ifft(H_gpu, H_gpu, plan, True)
# Reshape to original dimensions
H = np.moveaxis(H_gpu.get(), 0, fftDim)
H = np.reshape(H, hShape)
return H
def ifft2_gpu(y, fftshift=False):
"""
C2C iFFT
do numpy.fft.ifft2
The input y is a 2D complex numpy array
"""
#get the shape of the initial numpy array
n1, n2 = y.shape
#from numpy array to GPUarray. Take the only first n2/2+1 non-redundant FFT coefficients when R2C.
# For C2C, the dimensions of input and output are the same.
#if fftshift is False:
# y2 = np.asarray(y[:,0:n2//2+1],np.complex64)
#else:
# y2 = np.asarray(np.fft.ifftshift(y)[:,0:n2//2+1],np.complex64)
if fftshift:
y2 = np.fft.ifftshift(y)
else:
y2 = y
ygpu = gpuarray.to_gpu(y2)
#Initialise empty output GPUarray
x = gpuarray.empty((n1, n2), np.complex128)
#inverse FFT
plan_backward = cu_fft.Plan((n1, n2), np.complex128, np.complex128)
cu_fft.ifft(ygpu, x, plan_backward)
#Must divide by the total number of pixels in the image to get the normalization right
xout = x.get() / n1 / n2
return xout
def ifft2_gpu(y, fftshift=False):
''' This function produce an output that is
completely compatible with numpy.fft.ifft2
The input y is a 2D complex numpy array'''
# Convert the input array to complex64
if y.dtype != 'complex64':
y = y.astype('complex64')
# Get the shape of the initial numpy array
n1, n2 = y.shape
# From numpy array to GPUarray. Take only the first n2/2+1 non redundant FFT coefficients
if fftshift is False:
y2 = np.asarray(y[:, 0:n2 // 2 + 1], np.complex64)
else:
y2 = np.asarray(np.fft.ifftshift(y)[:, :n2 // 2 + 1], np.complex64)
ygpu = gpuarray.to_gpu(y2)
# Initialise empty output GPUarray
x = gpuarray.empty((n1, n2), np.float32)
# Inverse FFT
plan_backward = cu_fft.Plan((n1, n2), np.complex64, np.float32)
cu_fft.ifft(ygpu, x, plan_backward)
# Must divide by the total number of pixels in the image to get the normalisation right
xout = x.get() / n1 / n2
return xout
def setup_mesh(self, mesh):
'''Create the meshgrid, compute and store integrated Green's
function from mesh distances.
Only accepts meshes with same shape as self.mesh .
'''
assert (mesh.shape == self.mesh.shape)
self.mesh = mesh
mesh_arr = [
-mesh.distances[i]/2 +
np.arange(mesh.shape_r[i] + 1.) * mesh.distances[i]
for i in [1, 0]
]
# mesh_arr is [my, mx]
mesh_grids = np.meshgrid(*mesh_arr, indexing='ij') #choose my, mx
fgreen2 = self._fgreen(*mesh_grids)
fgreen2 = self._mirror(fgreen2)
# tile in 3d dimension, yields to memerror, uses huuge amount of memory!
#fgreen = np.tile(fgreen, (mesh.nz, 2*mesh.ny, 2*mesh.nx))
if self.save_memory:
plan_2d = cu_fft.Plan([2*self.mesh.ny, 2*self.mesh.nx],
in_dtype=np.complex128, out_dtype=np.complex128)
cu_fft.fft(gpuarray.to_gpu(fgreen2), self.fgreentr, plan=plan_2d)
else:
fgreen = np.empty(shape=(mesh.nz, 2*mesh.ny, 2*mesh.nx),
dtype=np.complex128)
for nn in range(mesh.nz):
fgreen[nn,:,:] = fgreen2
cu_fft.fft(gpuarray.to_gpu(fgreen), self.fgreentr,
plan=self.plan_forward)
def do_ffts(img):
if img.dtype != 'float32':
img = img.astype('float32')
sx, sy = img.shape # or the convention (width x height) vs (rows x columns)
# is sure to bite you
## Prepare and run CUDA FFT on image
# See https://www.idtools.com.au/gpu-accelerated-fft-compatible-numpy/
time0 = time.time()
# Initialise CUDA input GPUArray
x_gpu = gpuarray.to_gpu(img)
# Initialise output GPUarray
# N/2+1 non-redundant coefficients of a length-N input signal
y_gpu = gpuarray.empty((sx, sy // 2 + 1), np.complex64)
# Plan and run Cuda fft
plan_fft = cu_fft.Plan((sx, sy), np.float32, np.complex64)
cu_fft.fft(x_gpu, y_gpu, plan_fft)
gpu_fft = y_gpu.get()
gpu_time = time.time() - time0
print(f'GPU FFT preparation, execution and retrieval in {gpu_time:6.4f} s')
## Run np fft
time0 = time.time()
cpu_fft = np.fft.fft2(img)
cpu_time = time.time() - time0
print(f'NumPY CPU FFT in {cpu_time:6.4f} s')
return (gpu_fft, cpu_fft, gpu_time, cpu_time)
def __init__(self, mesh, context=None, save_memory=True):
'''
Args:
mesh The mesh on which the solver will operate. The dimensionality
is deducted from mesh.dimension
save_memory: Decide whether to store all slices of the transformed
greens function (more memory but faster) or save 1 slice only
(saves memory but slower, default)
'''
# create the mesh grid and compute the greens function on it
if (mesh.dimension != 3):
print ('Error: Use a 3d mesh for the 2.5d algorithm!. Abort.')
return None
self.is_25D = True
self.save_memory = save_memory
self.mesh = mesh
self._context = context
nz, ny, nx = mesh.shape
mesh_shape2 = [2*n for n in mesh.shape] # 2*nz, 2*ny, (2*nx)
if save_memory:
self.fgreentr = gpuarray.empty((2*ny, 2*nx),
dtype=np.complex128)
self._solve_kernel = self._solve_kernel_slow
else:
self.fgreentr = gpuarray.empty((nz, 2*ny, 2*nx),
dtype=np.complex128)
self._solve_kernel = self._solve_kernel_fast
self.tmpspace = gpuarray.zeros((nz, 2*ny, 2*nx), dtype=np.complex128)
sizeof_complex = np.dtype(np.complex128).itemsize
# dimensionality function dispatch
self._fgreen = getattr(self, '_fgreen25d')
self._mirror = getattr(self, '_mirror2d')
#copy_fn = {'3d' : get_Memcpy3D_d2d, '2d': get_Memcpy2D_d2d}
memcpy_nd = get_Memcpy3D_d2d
#memcpy_nd = copy_fn[str(dim) + 'd']
dim_args = mesh.shape
self._cpyrho2tmp = memcpy_nd(
src=None, dst=self.tmpspace, # None because src(rho) not yet known
src_pitch=mesh.nx*sizeof_complex,
dst_pitch=2*mesh.nx*sizeof_complex,
dim_args=dim_args,
itemsize=np.dtype(np.complex128).itemsize,
src_height=mesh.ny,
dst_height=2*mesh.ny)
self._cpytmp2rho = memcpy_nd(
src=self.tmpspace, dst=None, # None because dst(rho) not yet known
src_pitch=2*mesh.nx*sizeof_complex,
dst_pitch=mesh.nx*sizeof_complex,
dim_args=dim_args,
itemsize=np.dtype(np.complex128).itemsize,
src_height=2*mesh.ny,
dst_height=mesh.ny)
self.plan_forward = cu_fft.Plan([2*mesh.ny, 2*mesh.nx],
in_dtype=np.complex128, out_dtype=np.complex128, batch=mesh.nz)
self.plan_backward = self.plan_forward
self.setup_mesh(mesh)
def test_fft_float32_to_complex64_1d(self):
x = np.asarray(np.random.rand(self.N), np.float32)
xf = np.fft.rfftn(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty(self.N // 2 + 1, np.complex64)
plan = fft.Plan(x.shape, np.float32, np.complex64)
fft.fft(x_gpu, xf_gpu, plan)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float32)
def test_multiple_streams(self):
x = np.asarray(np.random.rand(self.N), np.float32)
xf = np.fft.rfftn(x)
y = np.asarray(np.random.rand(self.N), np.float32)
yf = np.fft.rfftn(y)
x_gpu = gpuarray.to_gpu(x)
y_gpu = gpuarray.to_gpu(y)
xf_gpu = gpuarray.empty(self.N // 2 + 1, np.complex64)
yf_gpu = gpuarray.empty(self.N // 2 + 1, np.complex64)
stream0 = drv.Stream()
stream1 = drv.Stream()
plan1 = fft.Plan(x.shape, np.float32, np.complex64, stream=stream0)
plan2 = fft.Plan(y.shape, np.float32, np.complex64, stream=stream1)
fft.fft(x_gpu, xf_gpu, plan1)
fft.fft(y_gpu, yf_gpu, plan2)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float32)
assert np.allclose(yf, yf_gpu.get(), atol=atol_float32)
def test_ifft_complex128_to_float64_1d(self):
x = np.asarray(np.random.rand(self.N), np.float64)
xf = np.asarray(np.fft.rfftn(x), np.complex128)
xf_gpu = gpuarray.to_gpu(xf)
x_gpu = gpuarray.empty(self.N, np.float64)
plan = fft.Plan(x.shape, np.complex128, np.float64)
fft.ifft(xf_gpu, x_gpu, plan, True)
assert np.allclose(x, x_gpu.get(), atol=atol_float64)
def test_batch_fft_float32_to_complex64_2d(self):
x = np.asarray(np.random.rand(self.B, self.N, self.M), np.float32)
xf = np.fft.rfftn(x, axes=(1,2))
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((self.B, self.N, self.M//2+1), np.complex64)
plan = fft.Plan([self.N, self.M], np.float32, np.complex64, batch=self.B)
fft.fft(x_gpu, xf_gpu, plan)
np.testing.assert_allclose(xf, xf_gpu.get(), atol=atol_float32)
def _get_inv_plan(itype, otype, outlen, batch=1):
try:
theplan = _reverse_plans[(itype, otype, outlen, batch)]
except KeyError:
theplan = cu_fft.Plan((outlen, ), itype, otype, batch=batch)
_reverse_plans.update({(itype, otype, outlen): theplan})
return theplan
def _get_fwd_plan(itype, otype, inlen, batch=1):
try:
theplan = _forward_plans[(itype, otype, inlen, batch)]
except KeyError:
theplan = cu_fft.Plan((inlen, ), itype, otype, batch=batch)
_forward_plans.update({(itype, otype, inlen): theplan})
return theplan
def test_batch_fft_float64_to_complex128_1d(self):
x = np.asarray(np.random.rand(self.B, self.N), np.float64)
xf = np.fft.rfft(x, axis=1)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((self.B, self.N // 2 + 1), np.complex128)
plan = fft.Plan(x.shape[1], np.float64, np.complex128, batch=self.B)
fft.fft(x_gpu, xf_gpu, plan)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float64)
def filter_fft_cuda(signal: np.array, window: np.array, prec: dict):
"""
Computes the low_pass filter using the numpy pycuda method.
Also auto-inits the pycuda library
:param signal: The input series
:param window: The input window
:param prec: The precision entry
:return: The filtered signal
"""
import pycuda.autoinit # Here because it initialises a new cuda environment every trial.
import pycuda.gpuarray as gpuarray
import skcuda.fft as cu_fft
import skcuda.linalg as linalg
linalg.init()
nfft = determine_size(len(signal) + len(window) - 1)
# Move data to GPU
sig_zero_pad = np.zeros(nfft, dtype=prec['float'])
win_zero_pad = np.zeros(nfft, dtype=prec['float'])
sig_gpu = gpuarray.zeros(sig_zero_pad.shape, dtype=prec['float'])
win_gpu = gpuarray.zeros(win_zero_pad.shape, dtype=prec['float'])
sig_zero_pad[0:len(signal)] = signal
win_zero_pad[0:len(window)] = window
sig_gpu.set(sig_zero_pad)
win_gpu.set(win_zero_pad)
# Plan forwards
sig_fft_gpu = gpuarray.zeros(nfft, dtype=prec['complex'])
win_fft_gpu = gpuarray.zeros(nfft, dtype=prec['complex'])
sig_plan_forward = cu_fft.Plan(sig_fft_gpu.shape, prec['float'],
prec['complex'])
win_plan_forward = cu_fft.Plan(win_fft_gpu.shape, prec['float'],
prec['complex'])
cu_fft.fft(sig_gpu, sig_fft_gpu, sig_plan_forward)
cu_fft.fft(win_gpu, win_fft_gpu, win_plan_forward)
# Convolve
out_fft = linalg.multiply(sig_fft_gpu, win_fft_gpu, overwrite=True)
linalg.scale(2.0, out_fft)
# Plan inverse
out_gpu = gpuarray.zeros_like(out_fft)
plan_inverse = cu_fft.Plan(out_fft.shape, prec['complex'], prec['complex'])
cu_fft.ifft(out_fft, out_gpu, plan_inverse, True)
out_np = np.zeros(len(out_gpu), prec['complex'])
out_gpu.get(out_np)
return out_np
def test_fft_float64_to_complex128_2d(self):
x = np.asarray(np.random.rand(self.N, self.M), np.float64)
xf = np.fft.rfftn(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((self.N, self.M // 2 + 1), np.complex128)
plan = fft.Plan(x.shape, np.float64, np.complex128)
fft.fft(x_gpu, xf_gpu, plan)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float64)
def test_work_area(self):
x = np.asarray(np.random.rand(self.N), np.float32)
xf = np.fft.rfftn(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty(self.N // 2 + 1, np.complex64)
plan = fft.Plan(x.shape, np.float32, np.complex64, auto_allocate=False)
work_area = gpuarray.empty((plan.worksize, ), np.uint8)
plan.set_work_area(work_area)
fft.fft(x_gpu, xf_gpu, plan)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float32)
def test_batch_ifft_complex128_to_float64_1d(self):
# Note that since rfftn returns a Fortran-ordered array, it
# needs to be reformatted as a C-ordered array before being
# passed to gpuarray.to_gpu:
x = np.asarray(np.random.rand(self.B, self.N), np.float64)
xf = np.asarray(np.fft.rfft(x, axis=1), np.complex128)
xf_gpu = gpuarray.to_gpu(np.ascontiguousarray(xf))
x_gpu = gpuarray.empty((self.B, self.N), np.float64)
plan = fft.Plan(x.shape[1], np.complex128, np.float64, batch=self.B)
fft.ifft(xf_gpu, x_gpu, plan, True)
assert np.allclose(x, x_gpu.get(), atol=atol_float64)
def test_ifft_complex64_to_float32_2d(self):
# Note that since rfftn returns a Fortran-ordered array, it
# needs to be reformatted as a C-ordered array before being
# passed to gpuarray.to_gpu:
x = np.asarray(np.random.rand(self.N, self.M), np.float32)
xf = np.asarray(np.fft.rfftn(x), np.complex64)
xf_gpu = gpuarray.to_gpu(np.ascontiguousarray(xf))
x_gpu = gpuarray.empty((self.N, self.M), np.float32)
plan = fft.Plan(x.shape, np.complex64, np.float32)
fft.ifft(xf_gpu, x_gpu, plan, True)
assert np.allclose(x, x_gpu.get(), atol=atol_float32)
def allocate_grid(self, **kwargs):
self.nf = kwargs.get('nf', self.nf)
assert (self.nf is not None)
self.n = int(self.sigma * self.nf)
self.ghat_g = gpuarray.zeros(self.n, dtype=self.complex_type)
self.cu_plan = cufft.Plan(self.n,
self.complex_type,
self.complex_type,
stream=self.stream)
return self
def irfft(a, normalize=True, nthreads=0):
if is_memory_enough(a):
arg = gpuarray.to_gpu(a)
shape = [s for s in a.shape]
shape[-1] = (shape[-1]-1)*2
rtype = G_CTYPES[a.dtype.type]
afg = gpuarray.empty(shape, rtype)
plan = fft.Plan(shape, a.dtype.type, rtype)
fft.ifft(arg, afg, plan)
return afg.get()
else:
return _irfft(a)
def rfft(a, nthreads=0):
if is_memory_enough(a):
arg = gpuarray.to_gpu(a)
shape = [s for s in a.shape]
shape[-1] = shape[-1]//2 + 1
ctype = G_RTYPES[a.dtype.type]
afg = gpuarray.empty(shape, ctype)
plan = fft.Plan(shape, a.dtype.type, ctype)
print(shape, a.dtype.type, ctype)
fft.fft(arg, afg, plan)
return afg.get()
else:
return _rfft(a)
def __init__(self, Nx, xmax, **kwargs):
"""__init__(self, Nx, xmax, **kwargs) -> None
initialize this class.
Parameters
----------
Nx : int
the number of grid points
xmax : float
the maximum of space in the x direction.
The spatial resolution is defined as 2*xmax/Nx.
kwargs : options
fft_type : str (default : 'numpy')
'numpy' : use fft methods in the numpy module
'fftw' : use fft methods in the pyfftw module if supported
'cufft' : use fft methods in the cufft module if supported
others : See the following classes or methods.
space class
InitPhoton()
InitMSFT()
InitAxes()
InitDensity()
"""
space.__init__(self, Nx, xmax, **kwargs)
self.InitPhoton(**kwargs)
self.InitMSFT(**kwargs)
self.InitAxes(**kwargs)
self.InitDensity(**kwargs)
# Check the validity of `fft_type`
if kwargs.get('fft_type') is None:
self.__fft_type = 'cufft'
elif kwargs.get('fft_type') not in ['numpy', 'fftw', 'cufft']:
raise ValueError('Invalid value for the keyword "fft_type."')
else:
self.__fft_type = kwargs.get('fft_type')
# Check the validity of cufft
if self.__fft_type == 'cufft':
if found_cufft is True: # in case of the existence of CUDA
buff = self.mesh_space()[0].shape
self.__x_gpu = gpuarray.empty(buff, np.complex64)
self.__xf_gpu = gpuarray.empty(buff, np.complex64)
self.__plan = cu_fft.Plan(buff, np.complex64, np.complex64)
else: # change into fftw
self.__fft_type = 'fftw'
# Check the validity of fftw
if self.__fft_type == 'fftw':
if found_pyfftw is False: # change into numpy
self.__fft_type = 'numpy'
