def compute_model_cuda(image_size,texref,c,kernelIndex,extendedBasis,params):
# Import CUDA function to perform the convolution
cu_compute_model = cu_matrix_kernel.get_function('cu_compute_model')
# Create a numpy array for the model M
M = np.zeros(image_size).astype(np.float32).copy()
# Call the cuda function to perform the convolution
blockDim = (256,1,1)
gridDim = (image_size[1],image_size[0])+(1,)
k0 = kernelIndex[:,0].astype(np.int32).copy()
k1 = kernelIndex[:,1].astype(np.int32).copy()
cu_compute_model(np.int32(params.pdeg),np.int32(params.sdeg),
np.int32(params.bdeg),cuda.In(k0),
cuda.In(k1), cuda.In(extendedBasis),
np.int32(kernelIndex.shape[0]),
cuda.In(c),cuda.Out(M),
block=blockDim,grid=gridDim,texrefs=[texref])
return M
def compute_matrix_and_vector_cuda(R,RB,T,Vinv,mask,kernelIndex,extendedBasis,
kernelRadius,params,stamp_positions=None):
# Import CUDA function to compute the matrix
cu_compute_matrix = cu_matrix_kernel.get_function('cu_compute_matrix')
cu_compute_vector = cu_matrix_kernel.get_function('cu_compute_vector')
cu_compute_matrix_stamps = cu_matrix_kernel.get_function('cu_compute_matrix_stamps')
cu_compute_vector_stamps = cu_matrix_kernel.get_function('cu_compute_vector_stamps')
# Copy the reference, target and inverse variance images to
# GPU texture memory
RTV = np.array([R,RB,T,Vinv,mask]).astype(np.float32).copy()
RTV_cuda = numpy3d_to_array(RTV)
texref = cu_matrix_kernel.get_texref("tex")
texref.set_array(RTV_cuda)
texref.set_filter_mode(cuda.filter_mode.POINT)
# Create a numpy array for matrix H
dp = (params.pdeg+1)*(params.pdeg+2)/2
ds = (params.sdeg+1)*(params.sdeg+2)/2
db = (params.bdeg+1)*(params.bdeg+2)/2
hs = (kernelIndex.shape[0]-1)*ds+dp+db
H = np.zeros([hs,hs]).astype(np.float32).copy()
V = np.zeros(hs).astype(np.float32).copy()
# Fill the elements of H
print hs,' * ',hs,' elements'
blockDim = (256,1,1)
gridDim = (hs,hs,1)
k0 = kernelIndex[:,0].astype(np.int32).copy()
k1 = kernelIndex[:,1].astype(np.int32).copy()
if params.use_stamps:
posx = np.float32(stamp_positions[:params.nstamps,0].copy()-1.0)
posy = np.float32(stamp_positions[:params.nstamps,1].copy()-1.0)
cu_compute_matrix_stamps(np.int32(params.pdeg),
np.int32(params.sdeg),
np.int32(params.bdeg),
np.int32(R.shape[1]),
np.int32(R.shape[0]),
np.int32(params.nstamps),
np.int32(params.stamp_half_width),
cuda.In(posx),
cuda.In(posy),
cuda.In(k0),
cuda.In(k1),
cuda.In(extendedBasis),
np.int32(kernelIndex.shape[0]),
np.int32(kernelRadius),
cuda.Out(H),
block=blockDim,grid=gridDim,texrefs=[texref])
else:
cu_compute_matrix(np.int32(params.pdeg),np.int32(params.sdeg),
np.int32(params.bdeg),
np.int32(R.shape[1]),np.int32(R.shape[0]),
cuda.In(k0),
cuda.In(k1),
cuda.In(extendedBasis),
np.int32(kernelIndex.shape[0]),np.int32(kernelRadius),
cuda.Out(H),block=blockDim,grid=gridDim,
texrefs=[texref])
# Fill the elements of V
blockDim = (256,1,1)
gridDim = (hs,1,1)
if params.use_stamps:
cu_compute_vector_stamps(np.int32(params.pdeg),np.int32(params.sdeg),
np.int32(params.bdeg),
np.int32(R.shape[1]),np.int32(R.shape[0]),np.int32(params.nstamps),
np.int32(params.stamp_half_width),
cuda.In(posx),
cuda.In(posy),
cuda.In(k0),
cuda.In(k1),
cuda.In(extendedBasis),
np.int32(kernelIndex.shape[0]),np.int32(kernelRadius),
cuda.Out(V),block=blockDim,grid=gridDim,
texrefs=[texref])
else:
cu_compute_vector(np.int32(params.pdeg),np.int32(params.sdeg),
np.int32(params.bdeg),
np.int32(R.shape[1]),np.int32(R.shape[0]),
cuda.In(k0),
cuda.In(k1),
cuda.In(extendedBasis),
np.int32(kernelIndex.shape[0]),np.int32(kernelRadius),
cuda.Out(V),block=blockDim,grid=gridDim,
texrefs=[texref])
return H, V, texref
def convolve_image_with_psf(psf_image,image1,image2,c,kernelIndex,extendedBasis,
kernelRadius,params):
from astropy.io import fits
# Read the PSF
psf,psf_hdr = fits.getdata(psf_image,0,header='true')
psf_height = psf_hdr['PSFHEIGH']
psf_x = psf_hdr['PSFX']
psf_y = psf_hdr['PSFY']
psf_size = psf.shape[1]
psf_fit_rad = params.psf_fit_radius
if params.psf_profile_type == 'gaussian':
psf_sigma_x = psf_hdr['PAR1']*0.8493218
psf_sigma_y = psf_hdr['PAR2']*0.8493218
psf_parameters = np.array([psf_size,psf_height,psf_sigma_x,psf_sigma_y,psf_x,
psf_y,psf_fit_rad,params.gain]).astype(np.float32)
profile_type = 0
elif params.psf_profile_type == 'moffat25':
print 'params.psf_profile_type moffat25 not working yet. Exiting.'
sys.exit(0)
psf_sigma_x = psf_hdr['PAR1']
psf_sigma_y = psf_hdr['PAR2']
psf_sigma_xy = psf_hdr['PAR3']
psf_parameters = np.array([psf_size,psf_height,psf_sigma_x,psf_sigma_y,psf_x,
psf_y,
psf_fit_rad,params.gain,psf_sigma_xy]).astype(np.float32)
profile_type = 1
else:
print 'params.psf_profile_type undefined'
sys.exit(0)
# Copy the images into GPU texture memory
nx, ny = image1.shape
RR = np.array([image1,image2]).astype(np.float32).copy()
image_cuda = numpy3d_to_array(RR)
texref = cu_matrix_kernel.get_texref("tex")
texref.set_array(image_cuda)
texref.set_filter_mode(cuda.filter_mode.POINT)
# Call the CUDA function to perform the double convolution.
# Each block is one image section.
# Each thread is one pixel of the PSF, but 32 threads per warp
cu_convolve = cu_matrix_kernel.get_function('convolve_image_psf')
k0 = kernelIndex[:,0].astype(np.int32).copy()
k1 = kernelIndex[:,1].astype(np.int32).copy()
#psf_0 = convolve_undersample(psf[0]).astype(np.float32).copy()
#psf_xd = convolve_undersample(psf[1]).astype(np.float32).copy()*0.0
#psf_yd = convolve_undersample(psf[2]).astype(np.float32).copy()*0.0
psf_0 = psf.astype(np.float32).copy()
psf_xd = psf.astype(np.float32).copy()*0.0
psf_yd = psf.astype(np.float32).copy()*0.0
image_section_size = 32
convolved_image1 = (0.0*image1).astype(np.float32)
convolved_image2 = (0.0*image1).astype(np.float32)
gridDim = (int((nx-1)/image_section_size+1),int((ny-1)/image_section_size+1),1)
blockDim = (16,16,1)
cu_convolve(np.int32(profile_type),np.int32(nx), np.int32(ny),
np.int32(image_section_size),
np.int32(image_section_size), np.int32(params.pdeg),
np.int32(params.sdeg),np.int32(c.shape[0]),
np.int32(kernelIndex.shape[0]),np.int32(kernelRadius),cuda.In(k0),
cuda.In(k1),cuda.In(extendedBasis),cuda.In(psf_parameters),cuda.In(psf_0),
cuda.In(psf_xd),cuda.In(psf_yd),cuda.In(c),cuda.Out(convolved_image1),
cuda.Out(convolved_image2),block=blockDim,grid=gridDim,texrefs=[texref])
return convolved_image1, convolved_image2
def photom_all_stars(diff,inv_variance,positions,psf_image,c,kernelIndex,
extendedBasis,kernelRadius,params,
star_group_boundaries=None,
detector_mean_positions_x=None,detector_mean_positions_y=None):
from astropy.io import fits
# Read the PSF
psf,psf_hdr = fits.getdata(psf_image,0,header='true')
print 'CIF psf_shape',psf.shape
print 'CIF psf_sum = ',np.sum(psf)
psf_height = psf_hdr['PSFHEIGH']
psf_x = psf_hdr['PSFX']
psf_y = psf_hdr['PSFY']
psf_size = psf.shape[1]
psf_fit_rad = params.psf_fit_radius
if params.psf_profile_type == 'gaussian':
psf_sigma_x = psf_hdr['PAR1']*0.8493218
psf_sigma_y = psf_hdr['PAR2']*0.8493218
psf_parameters = np.array([psf_size,psf_height,psf_sigma_x,psf_sigma_y,psf_x,
psf_y,psf_fit_rad,params.gain]).astype(np.float32)
profile_type = 0
elif params.psf_profile_type == 'moffat25':
print 'params.psf_profile_type moffat25 not working yet. Exiting.'
sys.exit(0)
psf_sigma_x = psf_hdr['PAR1']
psf_sigma_y = psf_hdr['PAR2']
psf_sigma_xy = psf_hdr['PAR3']
psf_parameters = np.array([psf_size,psf_height,psf_sigma_x,psf_sigma_y,psf_x,
psf_y,
psf_fit_rad,params.gain,psf_sigma_xy]).astype(np.float32)
print 'psf_parameters',psf_parameters
profile_type = 1
else:
print 'params.psf_profile_type undefined'
sys.exit(0)
# Copy the difference and inverse variance images into GPU texture memory
RR = np.array([diff,inv_variance]).astype(np.float32).copy()
diff_cuda = numpy3d_to_array(RR)
texref = cu_matrix_kernel.get_texref("tex")
texref.set_array(diff_cuda)
texref.set_filter_mode(cuda.filter_mode.POINT)
# Call the CUDA function to perform the photometry.
# Each block is one star.
# Each thread is one column of the PSF, but 32 threads per warp
nstars = positions.shape[0]
gridDim = (int(nstars),1,1)
blockDim = (16,16,1)
k0 = kernelIndex[:,0].astype(np.int32).copy()
k1 = kernelIndex[:,1].astype(np.int32).copy()
positions = positions.reshape(-1,2)
if params.star_file_is_one_based:
posx = np.float32(positions[:,0].copy()-1.0)
posy = np.float32(positions[:,1].copy()-1.0)
else:
posx = np.float32(positions[:,0].copy())
posy = np.float32(positions[:,1].copy())
#psf_0 = convolve_undersample(psf[0]).astype(np.float32).copy()
#psf_xd = convolve_undersample(psf[1]).astype(np.float32).copy()*0.0
#psf_yd = convolve_undersample(psf[2]).astype(np.float32).copy()*0.0
#psf_0 = psf[0].astype(np.float32).copy()
#psf_xd = psf[1].astype(np.float32).copy()*0.0
#psf_yd = psf[2].astype(np.float32).copy()*0.0
psf_0 = psf.astype(np.float32).copy()
psf_xd = psf.astype(np.float32).copy()*0.0
psf_yd = psf.astype(np.float32).copy()*0.0
flux = np.float32(posy.copy() * 0.0);
dflux = np.float32(posy.copy() * 0.0);
cu_photom = cu_matrix_kernel.get_function('cu_photom')
try:
cu_photom(np.int32(profile_type),np.int32(diff.shape[0]), np.int32(diff.shape[1]),
np.int32(params.pdeg),
np.int32(params.sdeg),np.int32(c.shape[0]),np.int32(kernelIndex.shape[0]),
np.int32(kernelRadius),cuda.In(k0),
cuda.In(k1),cuda.In(extendedBasis),
cuda.In(psf_parameters),cuda.In(psf_0),cuda.In(psf_xd),cuda.In(psf_yd),
cuda.In(posx),cuda.In(posy),cuda.In(c),cuda.Out(flux),cuda.Out(dflux),
block=blockDim,grid=gridDim,
texrefs=[texref])
except:
print 'Call to cu_photom failed.'
print 'psf_parameters', psf_parameters
print 'size of posx, posy:', posx.shape, posy.shape
print 'Parameters:'
for par in dir(params):
print par, getattr(params, par)
print
return flux, dflux
