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 test_fft_float32_to_complex64_2d(self):
x = np.asarray(np.random.rand(self.N, self.M), np.float32)
xf = np.fft.rfftn(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((self.N, self.M//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_batch_fft_float64_to_complex128_2d(self):
x = np.asarray(np.random.rand(self.B, self.N, self.M), np.float64)
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.complex128)
plan = fft.Plan([self.N, self.M], 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 test_fft_float64_to_complex128_1d(self):
x = np.asarray(np.random.rand(self.N), np.float64)
xf = np.fft.rfftn(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty(self.N//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_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)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float32)
def test_batch_fft_float32_to_complex64_1d(self):
x = np.asarray(np.random.rand(self.B, self.N), np.float32)
xf = np.fft.rfft(x, axis=1)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((self.B, self.N//2+1), np.complex64)
plan = fft.Plan(x.shape[1], np.float32, np.complex64, batch=self.B)
fft.fft(x_gpu, xf_gpu, plan)
assert np.allclose(xf, xf_gpu.get(), atol=atol_float32)
def test1():
N = 128
x = np.asarray(np.random.rand(N), np.complex64)
xf = np.fft.fft(x)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty(N, np.complex64)
plan = Plan(x.shape, np.complex64, np.complex64)
fft(x_gpu, xf_gpu, plan)
print(np.allclose(xf[0:N], xf_gpu.get(), atol=1e-3))
def test3():
N = 128
x = np.asarray(np.random.rand(N, N, N), np.complex64)
xf = np.fft.fftn(x, s=None, axes=(0, 1, 2))
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((N, N, N), np.complex64)
plan = Plan(x.shape, np.complex64, np.complex64)
fft(x_gpu, xf_gpu, plan)
print(np.allclose(xf[0:N, 0:N, 0:N], xf_gpu.get(), atol=1e-2))
def _solve_kernel_slow(self):
''' Slow version, use when save_memory is True: Stores only 1 slice
of the fgreentr function and loops over all slices
'''
cu_fft.fft(self.tmpspace, self.tmpspace, plan=self.plan_forward)
for i in xrange(self.mesh.nz):
self.tmpspace[i,:,:] = self.tmpspace[i,:,:] * self.fgreentr
cu_fft.ifft(self.tmpspace, self.tmpspace,
plan=self.plan_backward)
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_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 inplaceFractShift(img, dx, dy, PhaseShiftFunc, bInverse=False):
if dx == 0 and dy == 0:
return
global plan
global FT
Cache(img.shape)
cu_fft.fft(img, FT, plan)
PhaseShiftFunc(FT, kxx, kyy, np.float32(dx), np.float32(dy))
cu_fft.ifft(FT, img, plan, True)
def process_video_cuda(data):
global cs, cs_first
# fft_overlap(data, FiltV_GPU)
if cs_first == True:
prepare_video_filters(SysParams)
prepare_video_cuda()
cs_first = False
fdata = np.float32(data)
gpudata = gpuarray.to_gpu(fdata)
# first fft->ifft cycle applies pre-decoding filtering (low pass filters, CAV/CLV emphasis)
# and very importantly, performs the Hilbert transform
fft.fft(gpudata, cs['fft1_out'], cs['plan1'])
if Inner:
cs['fft1_out'] *= cs['filt_video_inner']
else:
cs['fft1_out'] *= cs['filt_video']
fft.ifft(cs['fft1_out'], cs['filtered1'], cs['plan1i'], True)
cs['doanglediff'](cs['fm_demod'],
cs['filtered1'],
block=(1024, 1, 1),
grid=(blocklenk, 1))
# post-processing: output low-pass filtering and deemphasis
fft.fft(cs['fm_demod'], cs['fft2_out'], cs['plan2'])
cs['fft2_out'] *= cs['filt_post']
fft.ifft(cs['fft2_out'], cs['postlpf'], cs['plan2i'], True)
cs['doclamp16'](cs['clipped_gpu'],
cs['postlpf'],
np.float32(-SysParams['output_minfreq']),
np.float32(SysParams['output_scale']),
block=(1024, 1, 1),
grid=(blocklenk, 1))
output_16 = cs['clipped_gpu'].get()
chop = 512
return output_16[chop:len(output_16) - chop]
# graph for debug
# plt.plot(cs['postlpf'].get()[5000:7500])
plt.plot(output_16[5000:7000])
# plt.plot(range(0, len(output_16)), output_16)
# plt.plot(range(0, len(doutput)), doutput)
# plt.plot(range(0, len(output_prefilt)), output_prefilt)
plt.show()
exit()
def process_video_cuda(data):
global cs, cs_first
# fft_overlap(data, FiltV_GPU)
if cs_first == True:
prepare_video_filters()
prepare_video_cuda()
cs_first = False
fdata = np.float32(data)
gpudata = gpuarray.to_gpu(fdata)
# first fft->ifft cycle applies pre-decoding filtering (low pass filters, CAV/CLV emphasis)
# and very importantly, performs the Hilbert transform
fft.fft(gpudata, cs["fft1_out"], cs["plan1"])
if Inner:
cs["fft1_out"] *= cs["filt_video_inner"]
else:
cs["fft1_out"] *= cs["filt_video"]
fft.ifft(cs["fft1_out"], cs["filtered1"], cs["plan1i"], True)
cs["doanglediff"](cs["fm_demod"], cs["filtered1"], block=(1024, 1, 1), grid=(blocklenk, 1))
# post-processing: output low-pass filtering and deemphasis
fft.fft(cs["fm_demod"], cs["fft2_out"], cs["plan2"])
cs["fft2_out"] *= cs["filt_post"]
fft.ifft(cs["fft2_out"], cs["postlpf"], cs["plan2i"], True)
cs["doclamp16"](
cs["clipped_gpu"],
cs["postlpf"],
np.float32(-SP["output_minfreq"]),
np.float32(SP["output_scale"]),
block=(1024, 1, 1),
grid=(blocklenk, 1),
)
output_16 = cs["clipped_gpu"].get()
chop = 512
return output_16[chop : len(output_16) - chop]
# graph for debug
# output = (sps.lfilter(f_deemp_b, f_deemp_a, output)[128:len(output)]) / deemp_corr
# plt.plot(cs['postlpf'].get()[5000:7500])
plt.plot(output_16[5000:7000])
# plt.plot(range(0, len(output_16)), output_16)
# plt.plot(range(0, len(doutput)), doutput)
# plt.plot(range(0, len(output_prefilt)), output_prefilt)
plt.show()
exit()
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 cross_correlate(plan, normalize=True):
#norm_template = sum(plan.mask.d_data)
fft(plan.volume, plan.volume_fft, plan.fwd_plan)
fft(plan.templatePadded, plan.template_fft, plan.fwd_plan)
conj(plan.template_fft, overwrite=True)
volume_fft = plan.volume_fft * plan.template_fft
ifft(volume_fft, plan.ccc_map, plan.inv_plan, scale=True)
plan.ccc_map /= np.float32(plan.p.get()) * plan.stdV
def FractShift(src, dest, dx, dy, PhaseShiftFunc):
if dx == 0 and dy == 0:
return
global plan
global FT
Cache(src.shape)
cu_fft.fft(src, FT, plan, PhaseShiftFunc)
PhaseShift(FT, dx, dy)
cu_fft.ifft(FT, dest, plan, True)
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 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 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 RunCorrection(neib,ROI,DifPad,rspace,kspace,exitWave,buffer_exitWave,finalObj,offsetx,offsety,objsizex,roisizex,CopyFromROI,ExitwaveAndBuffer,ApplyDifPad,cufftplan,aperture,fcachevector): Fs = [] for jpos in range(-neib,neib+1): for ipos in range(-neib,neib+1): CopyFromROI(rspace, finalObj, np.int32(offsety+jpos), np.int32(offsetx+ipos), roisizex, objsizex) ExitwaveAndBuffer(exitWave, buffer_exitWave, aperture, rspace) # Compute exitwaves cu_fft.fft(exitWave,kspace,cufftplan) # kspace = wave at detector ApplyDifPad(kspace,DifPad,fcachevector) # replace amplitudes. cu_fft.ifft(kspace,exitWave,cufftplan,True) # new exitwave errori = np.sum(((exitWave-buffer_exitWave).__abs__()**2).get()) Fs.append(errori+0) return GetMin(Fs,neib)
def thunk():
input_shape = inputs[0][0].shape
s = inputs[1][0]
# Since padding is not supported, assert s matches input shape.
assert (input_shape[1:] == s).all()
# construct output shape
output_shape = [input_shape[0]] + list(s)
# DFT of real input is symmetric, no need to store
# redundant coefficients
output_shape[-1] = output_shape[-1] // 2 + 1
# extra dimension with length 2 for real/imag
output_shape += [2]
output_shape = tuple(output_shape)
z = outputs[0]
# only allocate if there is no previous allocation of the
# right size.
if z[0] is None or z[0].shape != output_shape:
z[0] = pygpu.zeros(output_shape,
context=inputs[0][0].context,
dtype="float32")
input_pycuda = inputs[0][0]
# I thought we'd need to change the type on output_pycuda
# so it is complex64, but as it turns out skcuda.fft
# doesn't really care either way and treats the array as
# if it is complex64 anyway.
output_pycuda = z[0]
with input_pycuda.context:
# only initialise plan if necessary
if plan[0] is None or plan_input_shape[0] != input_shape:
plan_input_shape[0] = input_shape
plan[0] = fft.Plan(s,
np.float32,
np.complex64,
batch=input_shape[0])
# Sync GPU variables before computation
input_pycuda.sync()
output_pycuda.sync()
fft.fft(input_pycuda, output_pycuda, plan[0])
# Sync results to ensure output contains completed computation
pycuda.driver.Context.synchronize()
def __init__(self, volume, template, gpu):
self.gpu = gpu
volume_gpu = gu.to_gpu(volume)
self.fwd_plan = Plan(volume.shape, volume.dtype, np.complex64)
self.volume_fft = gu.zeros_like(volume_gpu, dtype=np.complex64)
fft(volume_gpu, self.volume_fft, self.fwd_plan)
self.template_fft = gu.zeros_like(volume_gpu, dtype=np.complex64)
self.ccc_map = gu.zeros_like(volume_gpu, dtype=np.float32)
self.norm_volume = gu.prod(volume_gpu.shape)
#self.scores = gu.zeros_like(volume_gpu, dtype=np.float32)
#self.angles = gu.zeros_like(volume_gpu, dtype=np.float32)
self.padded_volume = gu.zeros_like(volume_gpu, dtype=np.float32)
del volume_gpu
self.inv_plan = Plan(volume.shape, np.complex64, volume.dtype)
self.template = Volume(template)
def fft_2d(x, N, M, batch_size):
# print('Testing in-place fft..')
# for i in range(batch_size):
# x[i, :, :] = np.asarray(np.random.rand(N, M), np.complex64)
x_gpu = gpuarray.to_gpu(x)
# start = timer()
plan = cu_fft.Plan((N, M), np.complex128, np.complex128, batch_size)
cu_fft.fft(x_gpu, x_gpu, plan)
# timeit2=timer()-start
x_gpu1 = x_gpu.get()
# print ('take time:',timeit2)
return x_gpu1
def fft2c2c_cuda(x, axes=(0, 1)):
rank = len(axes)
x = np.array(x).astype(np.complex64)
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty(x.shape, np.complex64)
if len(x.shape) > rank:
batch = np.prod(x.shape[rank:len(x.shape)])
plan = Plan(x.shape[0:rank], np.complex64, np.complex64, batch, None, 1, \
np.array(x.shape[0:rank]).astype(np.int32), np.prod(x.shape[rank:len(x.shape)]), 1, \
np.array(x.shape[0:rank]).astype(np.int32), np.prod(x.shape[rank:len(x.shape)]), 1 )
else:
batch = 1
plan = Plan(x.shape[0:rank], np.complex64, np.complex64)
fft(x_gpu, xf_gpu, plan)
xf = xf_gpu.get()
return xf
def fft2_gpu(x, fftshift=False):
"""
R2C FFT
This function produce an output that is compatible with numpy.fft.fft2.
The input x is a 2D numpy array
"""
#converting the input array to single precision float
if x.dtype != "float64":
x = x.astype(np.float64)
#get the shape of the initial numpy array
n1, n2 = x.shape
# from numpy array to GPUarray
xgpu = gpuarray.to_gpu(x)
#initialize output GPUarray
# For real to complex transformations, the fft function computes
# N/2+1 non-redundant coefficients of a length-N input signal
ysize = n2 // 2 + 1
y = gpuarray.empty((n1, ysize), np.complex128)
#forward FFT
plan_forward = cu_fft.Plan((n1, n2), np.float64, np.complex128)
cu_fft.fft(xgpu, y, plan_forward)
left = y.get()
#to make the output array compatible with the numpy output
# we need to stack horizontally the y.get() array and its flipped version
# we must take care of handling even or odd sized array to get the correct size of the final array
if n2 // 2 == n2 / 2:
#even
right = np.roll(np.fliplr(np.flipud(left))[:, 1:-1], 1, axis=0)
else:
#odd
right = np.roll(np.fliplr(np.flipud(left))[:, :-1], 1, axis=0)
print(right.shape)
print(left.shape)
#get a numpy array back to compatible with np.fft
if fftshift is False:
yout = np.hstack((left, right))
else:
yout = np.fft.fftshift(np.hstack((left, right)))
return yout
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_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 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 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 range(mesh.dimension)[::-1]
]
# mesh_arr is [mz, my, mx]
mesh_grids = np.meshgrid(*mesh_arr, indexing='ij')
fgreen = self._fgreen(*mesh_grids)
fgreen = self._mirror(fgreen)
cu_fft.fft(gpuarray.to_gpu(fgreen), self.fgreentr,
plan=self.plan_forward)
def fft2_gpu(x, fftshift=False):
#code taken verbatim from https://www.idtools.com.au/gpu-accelerated-fft-compatible-numpy/
''' This function produce an output that is
compatible with numpy.fft.fft2
The input x is a 2D numpy array'''
# Convert the input array to single precision float
if x.dtype != 'float32':
x = x.astype('float32')
# Get the shape of the initial numpy array
n1, n2 = x.shape
# From numpy array to GPUarray
xgpu = gpuarray.to_gpu(x)
# Initialise output GPUarray
# For real to complex transformations, the fft function computes
# N/2+1 non-redundant coefficients of a length-N input signal.
y = gpuarray.empty((n1, n2 // 2 + 1), np.complex64)
# Forward FFT
plan_forward = cu_fft.Plan((n1, n2), np.float32, np.complex64)
cu_fft.fft(xgpu, y, plan_forward)
left = y.get()
# To make the output array compatible with the numpy output
# we need to stack horizontally the y.get() array and its flipped version
# We must take care of handling even or odd sized array to get the correct
# size of the final array
if n2 // 2 == n2 / 2:
right = np.roll(np.fliplr(np.flipud(y.get()))[:, 1:-1], 1, axis=0)
else:
right = np.roll(np.fliplr(np.flipud(y.get()))[:, :-1], 1, axis=0)
# Get a numpy array back compatible with np.fft
if fftshift is False:
yout = np.hstack((left, right))
else:
yout = np.fft.fftshift(np.hstack((left, right)))
return yout.astype('complex128')
def Simulate(self, period, width):
period=np.float64(period)
width=np.float64(width)
for i in range(len(self.DatFiles)):
self.DatFiles[i].gpu_pulsar_signal = self.DatFiles[i].gpu_time - 0*period
self.MakeSignal(self.DatFiles[i].gpu_pulsar_signal, period, ((period*width)**2), grid=(self.DatFiles[i].Tblocks,1), block=(self.DatFiles[i].block_size,1,1))
s = self.DatFiles[i].gpu_pulsar_signal.get()
np.savetxt("realsig.dat", zip(np.arange(0,10000),s[:10000]))
fft.fft(self.DatFiles[i].gpu_pulsar_signal, self.DatFiles[i].gpu_pulsar_fft, self.DatFiles[i].Plan)
ranPhases = np.random.uniform(0,1, len(self.DatFiles[i].gpu_pulsar_fft))
CompRan = np.cos(2*np.pi*ranPhases) + 1j*np.sin(2*np.pi*ranPhases)
CompRan[0] = 1 + 0j
OComp = self.DatFiles[i].gpu_pulsar_fft.get()
NComp = OComp*CompRan
s = np.fft.irfft(NComp)
np.savetxt("ransig.dat", zip(np.arange(0,10000),s[:10000]))
def fft2_gpu_c2c(x, fftshift=True):
"""
C2C FFT
This function produce an output that is compatible with numpy.fft.fft2.
The input x is a 2D numpy array
"""
if x.dtype != np.complex128:
x = x.astype(np.complex128)
#get the shape of the initial numpy array
n1, n2 = x.shape
xgpu = gpuarray.to_gpu(x)
#Initialise empty output GPUarray
y = gpuarray.empty((n1, n2), np.complex128)
#FFT
plan_forward = cu_fft.Plan((n1, n2), np.complex128, np.complex128)
cu_fft.fft(xgpu, y, plan_forward)
#Must divide by the total number of pixels in the image to get the normalization right
yout = y.get() / n1 / n2
if fftshift:
yout = np.fft.fftshift(yout)
return yout
def poisson_solve(self, rho):
''' Solve the poisson equation with the given charge distribution
Args:
rho: Charge distribution (same dimensions as mesh)
Returns:
Phi (same dimensions as rho)
'''
rho = rho.astype(np.complex128)
self._cpyrho2tmp.set_src_device(rho.gpudata)
self._cpytmp2rho.set_dst_device(rho.gpudata)
# set to 0 since it might be filled with the old potential
self.tmpspace.fill(0)
self._cpyrho2tmp()
cu_fft.fft(self.tmpspace, self.tmpspace, plan=self.plan_forward)
cu_fft.ifft(self.tmpspace * self.fgreentr, self.tmpspace,
plan=self.plan_backward)
# store the result in the rho gpuarray to save space
self._cpytmp2rho()
# scale (cuFFT is unscaled)
phi = rho.real/(2**self.mesh.dimension * self.mesh.n_nodes)
phi *= self.mesh.volume_elem/(2**(self.mesh.dimension-1)*np.pi*epsilon_0)
return phi
def cu_lpf(stimulus, dt, freq):
"""
CUDA implementation of low-pass-filter.
stimulus: ndarray
The input to be filtered.
dt: float
The sampling interval of the input.
freq: float
The cut-off frequency of the low pass filter.
"""
num = len(stimulus)
num_fft = int(num / 2 + 1)
idtype = stimulus.dtype
odtype = np.complex128 if idtype == np.float64 else np.complex64
if not isinstance(stimulus, gpuarray.GPUArray):
d_stimulus = gpuarray.to_gpu(stimulus)
else:
d_stimulus = stimulus
plan = Plan(stimulus.shape, idtype, odtype)
d_fstimulus = gpuarray.empty(num_fft, odtype)
fft(d_stimulus, d_fstimulus, plan)
df = 1.0 / dt / num
idx = int(freq // df)
unit = int(d_fstimulus.dtype.itemsize / 4)
offset = int(d_fstimulus.gpudata) + d_fstimulus.dtype.itemsize * idx
cuda.memset_d32(offset, 0, unit * (num_fft - idx))
plan = Plan(stimulus.shape, odtype, idtype)
d_lpf_stimulus = gpuarray.empty(num, idtype)
ifft(d_fstimulus, d_lpf_stimulus, plan, False)
return d_lpf_stimulus.get()
def funcfftw(F, *args, **kwargs):
"""funcfftw(F, *args, **kwargs) -> numpy.2darray
apply 2D Fourier transform
Parameters
----------
F : numpy.2darray
args : options
kwargs : options
"""
if found_pyfftw is True and kwargs.get('fft_type') == 'fftw':
pyfftw.forget_wisdom()
func = pyfftw.builders.fft2(F,
overwrite_input=True,
planner_effort='FFTW_ESTIMATE',
threads=CPU_COUNT)
return func()
elif found_cufft is True and kwargs.get('fft_type') == 'cufft':
x_gpu = gpuarray.to_gpu(F.astype(np.complex64))
xf_gpu = gpuarray.empty(F.shape, np.complex64)
cu_fft.fft(x_gpu, xf_gpu, args[0])
return xf_gpu.get()
else:
return fft2(F)
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.mesh = mesh
self._context = context
mesh_shape = self.mesh.shape # nz, ny, (nx)
nz, ny, nx = mesh_shape
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
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 = 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 [1,2]
]
# mesh_arr is [mz, my, mx]
mesh_grids = np.meshgrid(*mesh_arr, indexing='ij') #choose my, mx
fgreen2 = self._fgreen(*mesh_grids)
fgreen2 = self._mirror(fgreen2)
fgreen = np.empty(shape=(mesh.nz, 2*mesh.ny, 2*mesh.nx),
dtype=np.complex128)
for nn in xrange(mesh.nz):
fgreen[nn,:,:] = 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))
self.plan_forward = cu_fft.Plan([2*self.mesh.ny, 2*self.mesh.nx],
in_dtype=np.complex128, out_dtype=np.complex128, batch=self.mesh.nz)
self.plan_backward = cu_fft.Plan([2*self.mesh.ny, 2*self.mesh.nx],
in_dtype=np.complex128, out_dtype=np.complex128, batch=self.mesh.nz)
if 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:
cu_fft.fft(gpuarray.to_gpu(fgreen), self.fgreentr,
plan=self.plan_forward)
import numpy as np
import skcuda.fft as cu_fft
print('Testing fft/ifft..')
N = 1024
M = N//2
x = np.asarray(np.random.rand(N, M), np.float32)
xf = np.fft.fft2(x)
y = np.real(np.fft.ifft2(xf))
x_gpu = gpuarray.to_gpu(x)
xf_gpu = gpuarray.empty((x.shape[0], x.shape[1]//2+1), np.complex64)
plan_forward = cu_fft.Plan(x_gpu.shape, np.float32, np.complex64)
cu_fft.fft(x_gpu, xf_gpu, plan_forward)
y_gpu = gpuarray.empty_like(x_gpu)
plan_inverse = cu_fft.Plan(x_gpu.shape, np.complex64, np.float32)
cu_fft.ifft(xf_gpu, y_gpu, plan_inverse, True)
print('Success status: %r' % np.allclose(y, y_gpu.get(), atol=1e-6))
print('Testing in-place fft..')
x = np.asarray(np.random.rand(N, M) + 1j * np.random.rand(N, M), np.complex64)
x_gpu = gpuarray.to_gpu(x)
plan = cu_fft.Plan(x_gpu.shape, np.complex64, np.complex64)
cu_fft.fft(x_gpu, x_gpu, plan)
cu_fft.ifft(x_gpu, x_gpu, plan, True)
def process_audio_cuda(data):
global cs, csa, csa_first
if csa_first == True:
prepare_audio_filters()
prepare_audio_cuda()
csa_first = False
fdata = np.float32(data)
gpudata = gpuarray.to_gpu(fdata)
fft.fft(gpudata, cs["fft1_out"], cs["plan1"])
cs["left_fft1"] = (cs["fft1_out"] * cs["filt_audio_left"])[
0 : (ablocklen // 2) + 1
] # [0:blocklen])[0:(ablocklen//2)+1]
cs["right_fft1"] = (cs["fft1_out"] * cs["filt_audio_right"])[0 : (ablocklen // 2) + 1]
fft.ifft(cs["left_fft1"], cs["fm_left"], cs["plan1i"], True)
fft.ifft(cs["right_fft1"], cs["fm_right"], cs["plan1i"], True)
cs["doanglediff_mac"](
cs["left_clipped"],
cs["fm_left"],
np.float32((afreq_hz / 1.0 / np.pi)),
np.float32(-SysParams["audio_lfreq"]),
block=(1024, 1, 1),
grid=(ablocklenk, 1),
)
cs["doanglediff_mac"](
cs["right_clipped"],
cs["fm_right"],
np.float32((afreq_hz / 1.0 / np.pi)),
np.float32(-SysParams["audio_rfreq"]),
block=(1024, 1, 1),
grid=(ablocklenk, 1),
)
fft.fft(cs["left_clipped"], cs["left_fft2"], cs["plan2"])
fft.fft(cs["right_clipped"], cs["right_fft2"], cs["plan2"])
cs["left_fft2"] *= cs["filt_audiolpf"]
cs["right_fft2"] *= cs["filt_audiolpf"]
fft.ifft(cs["left_fft2"], cs["left_out"], cs["plan2i"], True)
fft.ifft(cs["right_fft2"], cs["right_out"], cs["plan2i"], True)
aclip = 256
outlen = ablocklen
cs["doaudioscale"](
cs["scaledout"],
cs["left_out"],
cs["right_out"],
np.float32(20),
np.float32(0),
block=(32, 1, 1),
grid=(outlen // 32, 1),
)
output = cs["scaledout"].get()[aclip:-aclip]
return output, len(output) * 80 / 2
plt.plot(cs["scaledout"].get())
# plt.plot(cs['right_clipped'].get()[768:-768])
# plt.plot(cs['right_out'].get()[768:-768] + 100000)
plt.show()
exit()
def _solve_kernel_fast(self):
'''Fast kernel, use when save_memory is False
'''
cu_fft.fft(self.tmpspace, self.tmpspace, plan=self.plan_forward)
cu_fft.ifft(self.tmpspace * self.fgreentr, self.tmpspace,
plan=self.plan_backward)
def execute(self):
cu_fft.fft(self.invec, self.outvec, self.plan)
modSquared = mod.get_function("modSquared")
psiNonlinear = mod2.get_function("test")
modSquared.prepare(["P", "P", "I"])
psiNonlinear.prepare("FFFPPPI")
block = (16, 16, 1)
grid = (64, 64)
for n in np.arange(N_RUNS):
start = time.time()
for step in xrange(N_TIMESTEPS):
# print step
# Implementing split-step method
# Update wavefunction and resovoir, record density
cu_fft.fft(psi_gpu, psi_gpu, plan_forward)
psi_gpu *= kineticFactorHalf_gpu
cu_fft.ifft(psi_gpu, psi_gpu, plan_inverse, scale=True)
# currentDensity_gpu = abs(psi_gpu) ** 2
# currentDensity_gpu = psi_gpu.real **2 + psi_gpu.imag ** 2
currentDensity_gpu = (psi_gpu * psi_gpu.conj()).real
# modSquared.prepared_call(grid, block, psi_gpu.gpudata,
# currentDensity_gpu.gpudata, 1024)
# n_gpu *= cumath.exp(-gammaRdt_gpu + Rdt_gpu * currentDensity_gpu)
n_gpu *= cumath.exp(misc.add(- gammaRdt_gpu,
- misc.multiply(Rdt_gpu, currentDensity_gpu)))
n_gpu += Pdt_gpu
psi_gpu *= cumath.exp(
misc.add(
misc.add(misc.multiply(expFactorPolFirst_gpu, n_gpu),
def fft(invec, outvec, prec, itype, otype):
cuplan = _get_fwd_plan(invec.dtype, outvec.dtype, len(invec))
cu_fft.fft(invec.data, outvec.data, cuplan)
