def prepare_gpu(self, type_t="double"):
from pycuda import gpuarray
nfuns, npoints = self.I.shape
nregions, nsubpoints = self.J.shape
ndim = self.v.shape[1]
if type_t == "double":
v, b = np.float64(self.v), np.float64(self.b)
else:
v, b = np.float32(self.v), np.float32(self.b)
constvars = {
'EPIGRAPH_PROJ': 1,
'ndim': ndim,
'nfuns': nfuns,
'nregions': nregions,
'npoints': npoints,
'nsubpoints': nsubpoints,
'I': self.I,
'J': self.J,
'A_STORE': v,
'B_STORE': b,
'term_maxiter': ndim * nsubpoints,
'term_tolerance': 1e-9,
'TYPE_T': type_t,
}
for f in ['fabs']:
constvars[f.upper()] = f if type_t == "double" else (f + "f")
files = [resource_stream('opymize.operators', 'proj.cu')]
templates = [("epigraphproj", "P", (nfuns, nregions, 1), (24, 12, 1))]
self._kernel = prepare_kernels(files, templates,
constvars)['epigraphproj']
def indexedmult_prepare_gpu(B, P, x, type_t="double"):
J = B.shape[0]
K = x[0]['shape'][1]
L = B.shape[2]
M = B.shape[1]
N = x[0]['shape'][0]
constvars = {
'J': J, 'K': K, 'L': L, 'M': M, 'N': N,
'P': P, 'B': B, 'TYPE_T': type_t
}
files = [resource_stream('opymize.linear', 'indexed.cu')]
templates = [
("indexedmult", "PP", (N, J, M), (32, 24, 1)),
("indexedmultadj", "PP", (N, 1, 1), (512, 1, 1))
]
return prepare_kernels(files, templates, constvars)
def prepare_gpu(self, type_t="double"):
constvars = {
'L1_NORMS_PROJ': 1,
'lbd': self.lbd,
'N': self.N,
'M1': self.M[0],
'M2': self.M[1],
'matrixnorm': self.matrixnorm[0].upper(),
'TYPE_T': type_t,
}
for f in ['fmin', 'fmax', 'sqrt', 'hypot']:
constvars[f.upper()] = f if type_t == "double" else (f + "f")
files = [resource_stream('opymize.operators', 'proj.cu')]
templates = [("l1normsproj", "P", (self.N, 1, 1), (200, 1, 1))]
self._kernel = prepare_kernels(files, templates,
constvars)['l1normsproj']
def prepare_gpu(self, type_t="double"):
if self._kernel is not None: return
N, K = self.x[0]['shape']
_, J = self.y[0]['shape']
constvars = {
'A': self.A,
'J': J,
'K': K,
'N': N,
'trans': 't' if self.trans else 'n',
'MATRIX_MULT_R': 1,
'TYPE_T': type_t
}
files = [resource_stream('opymize.linear', 'einsum.cu')]
templates = [("matrixmultr", "PP", (N, J, 1), (32, 24, 1))]
self._kernel = prepare_kernels(files, templates,
constvars)['matrixmultr']
self.adjoint.prepare_gpu(type_t=type_t)
def prepare_gpu(self, type_t="double"):
if self._kernel is not None: return
J, L, K = self.y[0]['shape']
M = self.x[0]['shape'][2]
constvars = {
'A': self.A,
'J': J,
'K': K,
'L': L,
'M': M,
'trans': 't' if self.trans else 'n',
'MATRIX_MULT_R_BATCHED': 1,
'TYPE_T': type_t
}
files = [resource_stream('opymize.linear', 'einsum.cu')]
templates = [("matrixmultrbatched", "PP", (J, L, K), (8, 16, 4))]
self._kernel = prepare_kernels(files, templates,
constvars)['matrixmultrbatched']
self.adjoint.prepare_gpu(type_t=type_t)
def prepare_gpu(self, type_t="double"):
constvars = {
'QUAD_EPI_PROJ': 1,
'alph': np.float64(self.alph),
'N': self.N,
'M': self.M,
'TYPE_T': type_t,
}
if self.lbd < np.inf:
constvars['lbd'] = np.float64(self.lbd)
constvars['USE_LBD'] = 1
if self.shift is not None:
constvars['shift'] = self.shift
constvars['USE_SHIFT'] = 1
for f in ['sqrt', 'cbrt', 'acos', 'cos', 'fabs']:
constvars[f.upper()] = f if type_t == "double" else (f + "f")
files = [resource_stream('opymize.operators', 'proj.cu')]
templates = [("quadepiproj", "P", (self.N, 1, 1), (200, 1, 1))]
self._kernel = prepare_kernels(files, templates,
constvars)['quadepiproj']
def prepare_gpu(self, type_t="double"):
if self._kernel is not None: return
npoints = np.prod(self.imagedims)
ndims = len(self.imagedims)
skips = imagedim_skips(self.imagedims)
constvars = {
'LAPLACIAN': 1,
'ADJOINT': 1 if self.bc[-4:] == "_adj" else 0,
'boundary_conditions': self.bc[0].upper(),
'N': npoints, 'D': ndims, 'C': self.nchannels,
'skips': np.array(skips, dtype=np.int64, order='C'),
'imagedims': np.array(self.imagedims, dtype=np.int64, order='C'),
'imageh': self.imageh,
'TYPE_T': type_t
}
files = [resource_stream('opymize.linear', 'diff.cu')]
templates = [
("laplacian", "PP", (npoints, self.nchannels, 1), (32, 24, 1)),
]
self._kernel = prepare_kernels(files, templates, constvars)['laplacian']
self.adjoint.prepare_gpu(type_t=type_t)
def diff_prepare_gpu(scheme, imagedims, imageh, nchannels, weights, type_t):
npoints = np.prod(imagedims)
ndims = len(imagedims)
skips = imagedim_skips(imagedims)
constvars = {
'SCHEME_%s'%scheme.upper(): 1,
'GRAD_DIV': 1,
'N': npoints, 'D': ndims, 'C': nchannels, 'dc_skip': ndims*nchannels,
'skips': np.array(skips, dtype=np.int64, order='C'),
'imagedims': np.array(imagedims, dtype=np.int64, order='C'),
'avgskips': staggered_diff_avgskips(imagedims),
'navgskips': 1 << (ndims - 1),
'weights': weights,
'imageh': imageh,
'TYPE_T': type_t
}
files = [resource_stream('opymize.linear', 'diff.cu')]
templates = [
("gradient", "PP", (npoints, nchannels, ndims), (24, 16, 2)),
("divergence", "PP", (npoints, nchannels, 1), (32, 24, 1)),
]
return prepare_kernels(files, templates, constvars)