def tv2(data, weight, Niter=50):
"""
chambolles tv regularized denoising
weight should be around 2+1.5*noise_sigma
"""
prog = OCLProgram(abspath("kernels/tv2.cl"))
data_im = OCLImage.from_array(data.astype(np, float32, copy=False))
pImgs = [
dev.createImage(
data.shape[::-1], mem_flags=cl.mem_flags.READ_WRITE, dtype=np.float32, channel_order=cl.channel_order.RGBA
)
for i in range(2)
]
outImg = dev.createImage(data.shape[::-1], dtype=np.float32, mem_flags=cl.mem_flags.READ_WRITE)
dev.writeImage(inImg, data.astype(np.float32))
dev.writeImage(pImgs[0], np.zeros((4,) + data.shape, dtype=np.float32))
dev.writeImage(pImgs[1], np.zeros((4,) + data.shape, dtype=np.float32))
for i in range(Niter):
proc.runKernel("div_step", inImg.shape, None, inImg, pImgs[i % 2], outImg)
proc.runKernel("grad_step", inImg.shape, None, outImg, pImgs[i % 2], pImgs[1 - i % 2], np.float32(weight))
return dev.readImage(outImg, dtype=np.float32)
def bilateral2(data, fSize, sigma_p, sigma_x = 10.):
"""bilateral filter """
dtype = data.dtype.type
dtypes_kernels = {np.float32:"bilat2_float",
np.uint16:"bilat2_short"}
if not dtype in dtypes_kernels.keys():
logger.info("data type %s not supported yet (%s), casting to float:"%(dtype,dtypes_kernels.keys()))
data = data.astype(np.float32)
dtype = data.dtype.type
img = OCLImage.from_array(data)
res = OCLArray.empty_like(data)
prog = OCLProgram(abspath("kernels/bilateral2.cl"))
prog.run_kernel(dtypes_kernels[dtype],
img.shape,None,
img,res.data,
np.int32(img.shape[0]),np.int32(img.shape[1]),
np.int32(fSize),np.float32(sigma_x),np.float32(sigma_p))
return res.get()
def tv2(data, weight, Niter=50):
"""
chambolles tv regularized denoising
weight should be around 2+1.5*noise_sigma
"""
prog = OCLProgram(abspath("kernels/tv2.cl"))
data_im = OCLImage.from_array(data.astype(np, float32, copy=False))
pImgs = [
dev.createImage(data.shape[::-1],
mem_flags=cl.mem_flags.READ_WRITE,
dtype=np.float32,
channel_order=cl.channel_order.RGBA) for i in range(2)
]
outImg = dev.createImage(data.shape[::-1],
dtype=np.float32,
mem_flags=cl.mem_flags.READ_WRITE)
dev.writeImage(inImg, data.astype(np.float32))
dev.writeImage(pImgs[0], np.zeros((4, ) + data.shape, dtype=np.float32))
dev.writeImage(pImgs[1], np.zeros((4, ) + data.shape, dtype=np.float32))
for i in range(Niter):
proc.runKernel("div_step", inImg.shape, None, inImg, pImgs[i % 2],
outImg)
proc.runKernel("grad_step", inImg.shape, None, outImg, pImgs[i % 2],
pImgs[1 - i % 2], np.float32(weight))
return dev.readImage(outImg, dtype=np.float32)
def nlm3(data,sigma, size_filter = 2, size_search = 3):
"""for noise level of sigma_0, choose sigma = 1.5*sigma_0
"""
prog = OCLProgram(abspath("kernels/nlm3.cl"),
build_options="-D FS=%i -D BS=%i"%(size_filter,size_search))
data = data.astype(np.float32, copy = False)
img = OCLImage.from_array(data)
distImg = OCLImage.empty_like(data)
distImg = OCLImage.empty_like(data)
tmpImg = OCLImage.empty_like(data)
tmpImg2 = OCLImage.empty_like(data)
accBuf = OCLArray.zeros(data.shape,np.float32)
weightBuf = OCLArray.zeros(data.shape,np.float32)
for dx in range(size_search+1):
for dy in range(-size_search,size_search+1):
for dz in range(-size_search,size_search+1):
prog.run_kernel("dist",img.shape,None,
img,tmpImg,np.int32(dx),np.int32(dy),np.int32(dz))
prog.run_kernel("convolve",img.shape,None,
tmpImg,tmpImg2,np.int32(1))
prog.run_kernel("convolve",img.shape,None,
tmpImg2,tmpImg,np.int32(2))
prog.run_kernel("convolve",img.shape,None,
tmpImg,distImg,np.int32(4))
prog.run_kernel("computePlus",img.shape,None,
img,distImg,accBuf.data,weightBuf.data,
np.int32(img.shape[0]),
np.int32(img.shape[1]),
np.int32(img.shape[2]),
np.int32(dx),np.int32(dy),np.int32(dz),
np.float32(sigma))
if any([dx,dy,dz]):
prog.run_kernel("computeMinus",img.shape,None,
img,distImg,accBuf.data,weightBuf.data,
np.int32(img.shape[0]),
np.int32(img.shape[1]),
np.int32(img.shape[2]),
np.int32(dx),np.int32(dy),np.int32(dz),
np.float32(sigma))
acc = accBuf.get()
weights = weightBuf.get()
return acc/weights
def nlm3(data,sigma, size_filter = 2, size_search = 3):
"""for noise level of sigma_0, choose sigma = 1.5*sigma_0
"""
prog = OCLProgram(abspath("kernels/nlm3.cl"),
build_options="-D FS=%i -D BS=%i"%(size_filter,size_search))
img = OCLImage.from_array(data)
distImg = OCLImage.empty_like(data)
distImg = OCLImage.empty_like(data)
tmpImg = OCLImage.empty_like(data)
tmpImg2 = OCLImage.empty_like(data)
accBuf = OCLArray.zeros(data.shape,np.float32)
weightBuf = OCLArray.zeros(data.shape,np.float32)
for dx in range(size_search+1):
for dy in range(-size_search,size_search+1):
for dz in range(-size_search,size_search+1):
prog.run_kernel("dist",img.shape,None,
img,tmpImg,np.int32(dx),np.int32(dy),np.int32(dz))
prog.run_kernel("convolve",img.shape,None,
tmpImg,tmpImg2,np.int32(1))
prog.run_kernel("convolve",img.shape,None,
tmpImg2,tmpImg,np.int32(2))
prog.run_kernel("convolve",img.shape,None,
tmpImg,distImg,np.int32(4))
prog.run_kernel("computePlus",img.shape,None,
img,distImg,accBuf.data,weightBuf.data,
np.int32(img.shape[0]),
np.int32(img.shape[1]),
np.int32(img.shape[2]),
np.int32(dx),np.int32(dy),np.int32(dz),
np.float32(sigma))
if any([dx,dy,dz]):
prog.run_kernel("computeMinus",img.shape,None,
img,distImg,accBuf.data,weightBuf.data,
np.int32(img.shape[0]),
np.int32(img.shape[1]),
np.int32(img.shape[2]),
np.int32(dx),np.int32(dy),np.int32(dz),
np.float32(sigma))
acc = accBuf.get()
weights = weightBuf.get()
return acc/weights
def _convolve_sep2_gpu(data_g, hx_g, hy_g, res_g = None):
assert_bufs_type(np.float32,data_g,hx_g,hy_g)
prog = OCLProgram(abspath("kernels/convolve_sep.cl"))
Ny,Nx = hy_g.shape[0],hx_g.shape[0]
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("conv_sep2_x",data_g.shape[::-1],None,data_g.data,hx_g.data,tmp_g.data,np.int32(Nx))
prog.run_kernel("conv_sep2_y",data_g.shape[::-1],None,tmp_g.data,hy_g.data,res_g.data,np.int32(Ny))
return res_g
def _convolve_sep2_gpu(data_g, hx_g, hy_g, res_g=None, dev=None):
assert_bufs_type(np.float32, data_g, hx_g, hy_g)
prog = OCLProgram(abspath("kernels/convolve_sep.cl"))
Ny, Nx = hy_g.shape[0], hx_g.shape[0]
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("conv_sep2_x", data_g.shape[::-1], None, data_g.data,
hx_g.data, tmp_g.data, np.int32(Nx))
prog.run_kernel("conv_sep2_y", data_g.shape[::-1], None, tmp_g.data,
hy_g.data, res_g.data, np.int32(Ny))
return res_g
def _convolve_buf(data_g, h_g , res_g = None):
"""
buffer variant
"""
assert_bufs_type(np.float32,data_g,h_g)
prog = OCLProgram(abspath("kernels/convolve.cl"))
if res_g is None:
res_g = OCLArray.empty(data_g.shape,dtype=np.float32)
Nhs = [np.int32(n) for n in h_g.shape]
kernel_name = "convolve%sd_buf"%(len(data_g.shape))
prog.run_kernel(kernel_name,data_g.shape[::-1],None,
data_g.data,h_g.data,res_g.data,
*Nhs)
return res_g
def affine(data, mat = np.identity(4), mode ="linear"):
"""affine transform data with matrix mat
"""
bop = {"linear":"","nearest":"-D USENEAREST"}
if not mode in bop.keys():
raise KeyError("mode = '%s' not defined ,valid: %s"%(mode, bop.keys()))
d_im = OCLImage.from_array(data)
res_g = OCLArray.empty(data.shape,np.float32)
mat_g = OCLArray.from_array(np.linalg.inv(mat).astype(np.float32,copy=False))
prog = OCLProgram(abspath("kernels/transformations.cl")
, build_options=[bop[mode]])
prog.run_kernel("affine",
data.shape[::-1],None,
d_im,res_g.data,mat_g.data)
return res_g.get()
def affine(data, mat = np.identity(4), interp = "linear"):
"""affine transform data with matrix mat
"""
bop = {"linear":"","nearest":"-D USENEAREST"}
if not interp in bop.keys():
raise KeyError("interp = '%s' not defined ,valid: %s"%(interp,bop.keys()))
d_im = OCLImage.from_array(data)
res_g = OCLArray.empty(data.shape,np.float32)
mat_g = OCLArray.from_array(np.linalg.inv(mat).astype(np.float32,copy=False))
prog = OCLProgram(abspath("kernels/transformations.cl")
, build_options=[bop[interp]])
prog.run_kernel("affine",
data.shape[::-1],None,
d_im,res_g.data,mat_g.data)
return res_g.get()
def bilateral2(data, fSize, sigma_p, sigma_x=10.):
"""bilateral filter """
dtype = data.dtype.type
dtypes_kernels = {np.float32: "bilat2_float", np.uint16: "bilat2_short"}
if not dtype in dtypes_kernels.keys():
logger.info("data type %s not supported yet (%s), casting to float:" %
(dtype, dtypes_kernels.keys()))
data = data.astype(np.float32)
dtype = data.dtype.type
img = OCLImage.from_array(data)
res = OCLArray.empty_like(data)
prog = OCLProgram(abspath("kernels/bilateral2.cl"))
prog.run_kernel(dtypes_kernels[dtype], img.shape, None, img, res.data,
np.int32(img.shape[0]), np.int32(img.shape[1]),
np.int32(fSize), np.float32(sigma_x), np.float32(sigma_p))
return res.get()
def _fftshift_single(d_g, res_g, ax = 0):
"""
basic fftshift of an OCLArray
shape(d_g) = [N_0,N_1...., N, .... N_{k-1, N_k]
= [N1, N, N2]
the we can address each element in the flat buffer by
index = i + N2*j + N2*N*k
where i = 1 .. N2
j = 1 .. N
k = 1 .. N1
and the swap of elements is performed on the index j
"""
dtype_kernel_name = {np.float32:"fftshift_1_f",
np.complex64:"fftshift_1_c"
}
N = d_g.shape[ax]
N1 = 1 if ax==0 else np.prod(d_g.shape[:ax])
N2 = 1 if ax == len(d_g.shape)-1 else np.prod(d_g.shape[ax+1:])
dtype = d_g.dtype.type
prog = OCLProgram(abspath("kernels/fftshift.cl"))
prog.run_kernel(dtype_kernel_name[dtype],(N2,N/2,N1),None,
d_g.data, res_g.data,
np.int32(N),
np.int32(N2))
return res_g
def scale(data, scale = (1.,1.,1.), interp = "linear"):
"""returns a interpolated, scaled version of data
scale = (scale_z,scale_y,scale_x)
or
scale = scale_all
interp = "linear" | "nearest"
"""
bop = {"linear":[],"nearest":["-D","USENEAREST"]}
if not interp in bop.keys():
raise KeyError("interp = '%s' not defined ,valid: %s"%(interp,bop.keys()))
if not isinstance(scale,(tuple, list, np.ndarray)):
scale = (scale,)*3
if len(scale) != 3:
raise ValueError("scale = %s misformed"%scale)
d_im = OCLImage.from_array(data)
nshape = np.array(data.shape)*np.array(scale)
nshape = tuple(nshape.astype(np.int))
res_g = OCLArray.empty(nshape,np.float32)
prog = OCLProgram(abspath("kernels/scale.cl"), build_options=bop[interp])
prog.run_kernel("scale",
res_g.shape[::-1],None,
d_im,res_g.data)
return res_g.get()
def _convolve3_old(data,h, dev = None):
"""convolves 3d data with kernel h on the GPU Device dev
boundary conditions are clamping to edge.
h is converted to float32
if dev == None the default one is used
"""
if dev is None:
dev = get_device()
if dev is None:
raise ValueError("no OpenCLDevice found...")
dtype = data.dtype.type
dtypes_options = {np.float32:"",
np.uint16:"-D SHORTTYPE"}
if not dtype in dtypes_options.keys():
raise TypeError("data type %s not supported yet, please convert to:"%dtype,dtypes_options.keys())
prog = OCLProgram(abspath("kernels/convolve3.cl"),
build_options = dtypes_options[dtype])
hbuf = OCLArray.from_array(h.astype(np.float32))
img = OCLImage.from_array(data)
res = OCLArray.empty(data.shape,dtype=np.float32)
Ns = [np.int32(n) for n in data.shape+h.shape]
prog.run_kernel("convolve3d",img.shape,None,
img,hbuf.data,res.data,
*Ns)
return res.get()
def scale(data, scale = (1.,1.,1.), interp = "linear"):
"""returns a interpolated, scaled version of data
scale = (scale_z,scale_y,scale_x)
or
scale = scale_all
interp = "linear" | "nearest"
"""
bop = {"linear":"","nearest":"-D USENEAREST"}
if not interp in bop.keys():
raise KeyError("interp = '%s' not defined ,valid: %s"%(interp,bop.keys()))
if not isinstance(scale,(tuple, list, np.ndarray)):
scale = (scale,)*3
if len(scale) != 3:
raise ValueError("scale = %s misformed"%scale)
d_im = OCLImage.from_array(data)
nshape = np.array(data.shape)*np.array(scale)
nshape = tuple(nshape.astype(np.int))
res_g = OCLArray.empty(nshape,np.float32)
prog = OCLProgram(abspath("kernels/scale.cl"), build_options=[bop[interp]])
prog.run_kernel("scale",
res_g.shape[::-1],None,
d_im,res_g.data)
return res_g.get()
def convolve_spatial2(im, hs,
mode = "constant",
plan = None,
return_plan = False):
"""
spatial varying convolution of an 2d image with a 2d grid of psfs
shape(im_ = (Ny,Nx)
shape(hs) = (Gy,Gx, Hy,Hx)
the input image im is subdivided into (Gy,Gz) blocks
hs[j,i] is the psf at the center of each block (i,j)
as of now each image dimension has to be divisble by the grid dim, i.e.
Nx % Gx == 0
Ny % Gy == 0
mode can be:
"constant" - assumed values to be zero
"wrap" - periodic boundary condition
"""
if im.ndim !=2 or hs.ndim !=4:
raise ValueError("wrong dimensions of input!")
if not np.all([n%g==0 for n,g in zip(im.shape,hs.shape[:2])]):
raise NotImplementedError("shape of image has to be divisible by Gx Gy = %s shape mismatch"%(str(hs.shape[:2])))
mode_str = {"constant":"CLK_ADDRESS_CLAMP",
"wrap":"CLK_ADDRESS_REPEAT"}
Ny, Nx = im.shape
Gy, Gx = hs.shape[:2]
# the size of each block within the grid
Nblock_y, Nblock_x = Ny/Gy, Nx/Gx
# the size of the overlapping patches with safety padding
Npatch_x, Npatch_y = _next_power_of_2(3*Nblock_x), _next_power_of_2(3*Nblock_y)
#Npatch_x, Npatch_y = _next_power_of_2(2*Nblock_x), _next_power_of_2(2*Nblock_y)
print Nblock_x, Npatch_x
hs = np.fft.fftshift(pad_to_shape(hs,(Gy,Gx,Npatch_y,Npatch_x)),axes=(2,3))
prog = OCLProgram(abspath("kernels/conv_spatial.cl"),
build_options=["-D","ADDRESSMODE=%s"%mode_str[mode]])
if plan is None:
plan = fft_plan((Npatch_y,Npatch_x))
patches_g = OCLArray.empty((Gy,Gx,Npatch_y,Npatch_x),np.complex64)
h_g = OCLArray.from_array(hs.astype(np.complex64))
im_g = OCLImage.from_array(im.astype(np.float32,copy=False))
x0s = Nblock_x*np.arange(Gx)
y0s = Nblock_y*np.arange(Gy)
print x0s
for i,_x0 in enumerate(x0s):
for j,_y0 in enumerate(y0s):
prog.run_kernel("fill_patch2",(Npatch_x,Npatch_y),None,
im_g,
np.int32(_x0+Nblock_x/2-Npatch_x/2),
np.int32(_y0+Nblock_y/2-Npatch_y/2),
patches_g.data,
np.int32(i*Npatch_x*Npatch_y+j*Gx*Npatch_x*Npatch_y))
# convolution
fft(patches_g,inplace=True, batch = Gx*Gy, plan = plan)
fft(h_g,inplace=True, batch = Gx*Gy, plan = plan)
prog.run_kernel("mult_inplace",(Npatch_x*Npatch_y*Gx*Gy,),None,
patches_g.data, h_g.data)
fft(patches_g,inplace=True, inverse = True, batch = Gx*Gy, plan = plan)
#return patches_g.get()
#accumulate
res_g = OCLArray.empty(im.shape,np.float32)
for i in xrange(Gx+1):
for j in xrange(Gy+1):
prog.run_kernel("interpolate2",(Nblock_x,Nblock_y),None,
patches_g.data,res_g.data,
np.int32(i),np.int32(j),
np.int32(Gx),np.int32(Gy),
np.int32(Npatch_x),np.int32(Npatch_y))
res = res_g.get()
if return_plan:
return res, plan
else:
return res
def _convolve_spatial2(im, hs,
mode = "constant",
grid_dim = None,
pad_factor = 2,
plan = None,
return_plan = False):
"""
spatial varying convolution of an 2d image with a 2d grid of psfs
shape(im_ = (Ny,Nx)
shape(hs) = (Gy,Gx, Hy,Hx)
the input image im is subdivided into (Gy,Gx) blocks
hs[j,i] is the psf at the center of each block (i,j)
as of now each image dimension has to be divisible by the grid dim, i.e.
Nx % Gx == 0
Ny % Gy == 0
mode can be:
"constant" - assumed values to be zero
"wrap" - periodic boundary condition
"""
if grid_dim:
Gs = tuple(grid_dim)
else:
Gs = hs.shape[:2]
mode_str = {"constant":"CLK_ADDRESS_CLAMP",
"wrap":"CLK_ADDRESS_REPEAT"}
Ny, Nx = im.shape
Gy, Gx = Gs
# the size of each block within the grid
Nblock_y, Nblock_x = Ny/Gy, Nx/Gx
# the size of the overlapping patches with safety padding
Npatch_x, Npatch_y = _next_power_of_2(pad_factor*Nblock_x), _next_power_of_2(pad_factor*Nblock_y)
prog = OCLProgram(abspath("kernels/conv_spatial2.cl"),
build_options=["-D","ADDRESSMODE=%s"%mode_str[mode]])
if plan is None:
plan = fft_plan((Npatch_y,Npatch_x))
x0s = Nblock_x*np.arange(Gx)
y0s = Nblock_y*np.arange(Gy)
patches_g = OCLArray.empty((Gy,Gx,Npatch_y,Npatch_x),np.complex64)
#prepare psfs
if grid_dim:
h_g = OCLArray.zeros((Gy,Gx,Npatch_y,Npatch_x),np.complex64)
tmp_g = OCLArray.from_array(hs.astype(np.float32, copy = False))
for i,_x0 in enumerate(x0s):
for j,_y0 in enumerate(y0s):
prog.run_kernel("fill_psf_grid2",
(Nblock_x,Nblock_y),None,
tmp_g.data,
np.int32(Nx),
np.int32(i*Nblock_x),
np.int32(j*Nblock_y),
h_g.data,
np.int32(Npatch_x),
np.int32(Npatch_y),
np.int32(-Nblock_x/2+Npatch_x/2),
np.int32(-Nblock_y/2+Npatch_y/2),
np.int32(i*Npatch_x*Npatch_y+j*Gx*Npatch_x*Npatch_y)
)
else:
hs = np.fft.fftshift(pad_to_shape(hs,(Gy,Gx,Npatch_y,Npatch_x)),axes=(2,3))
h_g = OCLArray.from_array(hs.astype(np.complex64))
#prepare image
im_g = OCLImage.from_array(im.astype(np.float32,copy=False))
for i,_x0 in enumerate(x0s):
for j,_y0 in enumerate(y0s):
prog.run_kernel("fill_patch2",(Npatch_x,Npatch_y),None,
im_g,
np.int32(_x0+Nblock_x/2-Npatch_x/2),
np.int32(_y0+Nblock_y/2-Npatch_y/2),
patches_g.data,
np.int32(i*Npatch_x*Npatch_y+j*Gx*Npatch_x*Npatch_y))
#return np.abs(patches_g.get())
# convolution
fft(patches_g,inplace=True, batch = Gx*Gy, plan = plan)
fft(h_g,inplace=True, batch = Gx*Gy, plan = plan)
prog.run_kernel("mult_inplace",(Npatch_x*Npatch_y*Gx*Gy,),None,
patches_g.data, h_g.data)
fft(patches_g,inplace=True, inverse = True, batch = Gx*Gy, plan = plan)
print Nblock_x, Npatch_x
#return np.abs(patches_g.get())
#accumulate
res_g = OCLArray.empty(im.shape,np.float32)
for j in xrange(Gy+1):
for i in xrange(Gx+1):
prog.run_kernel("interpolate2",(Nblock_x,Nblock_y),None,
patches_g.data,res_g.data,
np.int32(i),np.int32(j),
np.int32(Gx),np.int32(Gy),
np.int32(Npatch_x),np.int32(Npatch_y))
res = res_g.get()
if return_plan:
return res, plan
else:
return res
def convolve_spatial3(im, hs,
mode = "constant",
plan = None,
return_plan = False,
pad_factor = 2):
"""
spatial varying convolution of an 3d image with a 3d grid of psfs
shape(im_ = (Nz,Ny,Nx)
shape(hs) = (Gz,Gy,Gx, Hz,Hy,Hx)
the input image im is subdivided into (Gx,Gy,Gz) blocks
hs[k,j,i] is the psf at the center of each block (i,j,k)
as of now each image dimension has to be divisble by the grid dim, i.e.
Nx % Gx == 0
Ny % Gy == 0
Nz % Gz == 0
mode can be:
"constant" - assumed values to be zero
"wrap" - periodic boundary condition
"""
if im.ndim !=3 or hs.ndim !=6:
raise ValueError("wrong dimensions of input!")
if not np.all([n%g==0 for n,g in zip(im.shape,hs.shape[:3])]):
raise NotImplementedError("shape of image has to be divisible by Gx Gy = %s !"%(str(hs.shape[:3])))
mode_str = {"constant":"CLK_ADDRESS_CLAMP",
"wrap":"CLK_ADDRESS_REPEAT"}
Ns = tuple(im.shape)
Gs = tuple(hs.shape[:3])
# the size of each block within the grid
Nblocks = [n/g for n,g in zip(Ns,Gs)]
# the size of the overlapping patches with safety padding
Npatchs = tuple([_next_power_of_2(pad_factor*nb) for nb in Nblocks])
print hs.shape
hs = np.fft.fftshift(pad_to_shape(hs,Gs+Npatchs),axes=(3,4,5))
prog = OCLProgram(abspath("kernels/conv_spatial.cl"),
build_options=["-D","ADDRESSMODE=%s"%mode_str[mode]])
if plan is None:
plan = fft_plan(Npatchs)
patches_g = OCLArray.empty(Gs+Npatchs,np.complex64)
h_g = OCLArray.from_array(hs.astype(np.complex64))
im_g = OCLImage.from_array(im.astype(np.float32,copy=False))
Xs = [nb*np.arange(g) for nb, g in zip(Nblocks,Gs)]
print Nblocks
# this loops over all i,j,k
for (k,_z0), (j,_y0),(i,_x0) in product(*[enumerate(X) for X in Xs]):
prog.run_kernel("fill_patch3",Npatchs[::-1],None,
im_g,
np.int32(_x0+Nblocks[2]/2-Npatchs[2]/2),
np.int32(_y0+Nblocks[1]/2-Npatchs[1]/2),
np.int32(_z0+Nblocks[0]/2-Npatchs[0]/2),
patches_g.data,
np.int32(i*np.prod(Npatchs)+
j*Gs[2]*np.prod(Npatchs)+
k*Gs[2]*Gs[1]*np.prod(Npatchs)))
print patches_g.shape, h_g.shape
# convolution
fft(patches_g,inplace=True, batch = np.prod(Gs), plan = plan)
fft(h_g,inplace=True, batch = np.prod(Gs), plan = plan)
prog.run_kernel("mult_inplace",(np.prod(Npatchs)*np.prod(Gs),),None,
patches_g.data, h_g.data)
fft(patches_g,
inplace=True,
inverse = True,
batch = np.prod(Gs),
plan = plan)
#return patches_g.get()
#accumulate
res_g = OCLArray.zeros(im.shape,np.float32)
for k, j, i in product(*[range(g+1) for g in Gs]):
prog.run_kernel("interpolate3",Nblocks[::-1],None,
patches_g.data,
res_g.data,
np.int32(i),np.int32(j),np.int32(k),
np.int32(Gs[2]),np.int32(Gs[1]),np.int32(Gs[0]),
np.int32(Npatchs[2]),np.int32(Npatchs[1]),np.int32(Npatchs[0]))
res = res_g.get()
if return_plan:
return res, plan
else:
return res
def _convolve_spatial3(im, hs,
mode = "constant",
grid_dim = None,
plan = None,
return_plan = False,
pad_factor = 2):
if im.ndim !=3:
raise ValueError("wrong dimensions of input!")
if not (hs.ndim==6 or (hs.ndim==3 and grid_dim)):
raise ValueError("wrong dimensions of psf grid!")
if grid_dim:
if hs.shape != im.shape:
raise ValueError("if grid_dim is set, then im.shape = hs.shape !")
Gs = tuple(grid_dim)
else:
if not hs.ndim==6:
raise ValueError("wrong dimensions of psf grid! (Gy,Gx,Ny,Nx)")
Gs = hs.shape[:3]
if not np.all([n%g==0 for n,g in zip(im.shape,Gs)]):
raise NotImplementedError("shape of image has to be divisible by Gx Gy = %s shape mismatch"%(str(hs.shape[:2])))
mode_str = {"constant":"CLK_ADDRESS_CLAMP",
"wrap":"CLK_ADDRESS_REPEAT"}
Ns = im.shape
# the size of each block within the grid
Nblocks = [n/g for n,g in zip(Ns,Gs)]
# the size of the overlapping patches with safety padding
Npatchs = tuple([_next_power_of_2(pad_factor*nb) for nb in Nblocks])
prog = OCLProgram(abspath("kernels/conv_spatial3.cl"),
build_options=["-D","ADDRESSMODE=%s"%mode_str[mode]])
if plan is None:
plan = fft_plan(Npatchs)
Xs = [nb*np.arange(g) for nb, g in zip(Nblocks,Gs)]
patches_g = OCLArray.empty(Gs+Npatchs,np.complex64)
#prepare psfs
if grid_dim:
h_g = OCLArray.zeros(Gs+Npatchs,np.complex64)
tmp_g = OCLArray.from_array(hs.astype(np.float32, copy = False))
for (k,_z0), (j,_y0),(i,_x0) in product(*[enumerate(X) for X in Xs]):
prog.run_kernel("fill_psf_grid3",
Nblocks[::-1],None,
tmp_g.data,
np.int32(im.shape[2]),
np.int32(im.shape[1]),
np.int32(i*Nblocks[2]),
np.int32(j*Nblocks[1]),
np.int32(k*Nblocks[0]),
h_g.data,
np.int32(Npatchs[2]),
np.int32(Npatchs[1]),
np.int32(Npatchs[0]),
np.int32(-Nblocks[2]/2+Npatchs[2]/2),
np.int32(-Nblocks[1]/2+Npatchs[1]/2),
np.int32(-Nblocks[0]/2+Npatchs[0]/2),
np.int32(i*np.prod(Npatchs)+
j*Gs[2]*np.prod(Npatchs)+
k*Gs[2]*Gs[1]*np.prod(Npatchs)))
else:
hs = np.fft.fftshift(pad_to_shape(hs,Gs+Npatchs),axes=(3,4,5))
h_g = OCLArray.from_array(hs.astype(np.complex64))
im_g = OCLImage.from_array(im.astype(np.float32,copy=False))
# this loops over all i,j,k
for (k,_z0), (j,_y0),(i,_x0) in product(*[enumerate(X) for X in Xs]):
prog.run_kernel("fill_patch3",Npatchs[::-1],None,
im_g,
np.int32(_x0+Nblocks[2]/2-Npatchs[2]/2),
np.int32(_y0+Nblocks[1]/2-Npatchs[1]/2),
np.int32(_z0+Nblocks[0]/2-Npatchs[0]/2),
patches_g.data,
np.int32(i*np.prod(Npatchs)+
j*Gs[2]*np.prod(Npatchs)+
k*Gs[2]*Gs[1]*np.prod(Npatchs)))
# convolution
fft(patches_g,inplace=True, batch = np.prod(Gs), plan = plan)
fft(h_g,inplace=True, batch = np.prod(Gs), plan = plan)
prog.run_kernel("mult_inplace",(np.prod(Npatchs)*np.prod(Gs),),None,
patches_g.data, h_g.data)
fft(patches_g,
inplace=True,
inverse = True,
batch = np.prod(Gs),
plan = plan)
#return patches_g.get()
#accumulate
res_g = OCLArray.zeros(im.shape,np.float32)
for k, j, i in product(*[range(g+1) for g in Gs]):
prog.run_kernel("interpolate3",Nblocks[::-1],None,
patches_g.data,
res_g.data,
np.int32(i),np.int32(j),np.int32(k),
np.int32(Gs[2]),np.int32(Gs[1]),np.int32(Gs[0]),
np.int32(Npatchs[2]),np.int32(Npatchs[1]),np.int32(Npatchs[0]))
res = res_g.get()
if return_plan:
return res, plan
else:
return res
