def __init__(self, ring, header, gulp_nframe, buf_nframe):
SequenceBase.__init__(self, ring)
self._header = header
# This allows passing DataType instances instead of string types
header['_tensor']['dtype'] = str(header['_tensor']['dtype'])
header_str = json.dumps(header)
header_size = len(header_str)
tensor = self.tensor
# **TODO: Consider moving this into bfRingSequenceBegin
self.ring.resize(gulp_nframe * tensor['frame_nbyte'],
buf_nframe * tensor['frame_nbyte'],
tensor['nringlet'])
offset_from_head = 0
# TODO: How to allow time_tag to be optional? Probably need to plumb support through to backend.
self.obj = _bf.BFwsequence()
try:
hname = header['name'].encode()
hstr = header_str.encode()
except AttributeError:
# Python2 catch
hname = header['name']
hstr = header_str
_check(_bf.bfRingSequenceBegin(
self.obj,
ring.obj,
hname,
header['time_tag'],
header_size,
hstr,
tensor['nringlet'],
offset_from_head))
def __init__(self, ring, name="", time_tag=-1, header="", nringlet=1):
SequenceBase.__init__(self, ring)
# TODO: Allow header to be a string, buffer, or numpy array
header_size = len(header)
if isinstance(header, np.ndarray):
header = header.ctypes.data
elif isinstance(header, str):
try:
header = header.encode()
except AttributeError:
# Python2 catch
pass
#print("hdr:", header_size, type(header))
name = str(name)
offset_from_head = 0
self.obj = _bf.BFwsequence()
try:
name = name.encode()
except AttributeError:
# Python2 catch
pass
_check(
_bf.bfRingSequenceBegin(self.obj, ring.obj, name, time_tag,
header_size, header, nringlet,
offset_from_head))
def init(self, coeffs, decim=1, space='cuda'):
space = _string2space(space)
psize = None
_check(
_bf.bfFirInit(self.obj,
asarray(coeffs).as_BFarray(), decim, space, 0,
psize))
def set_devices_no_spin_cpu():
"""Sets a flag on all GPU devices that tells them not to spin the CPU when
synchronizing. This is useful for reducing CPU load in GPU pipelines.
This function must be called _before_ any GPU devices are
initialized (i.e., at the start of the process)."""
_check(_bf.bfDevicesSetNoSpinCPU())
def execute(self, idata, odata):
# TODO: Work out how to integrate CUDA stream
_check(
_bf.bfFirExecute(self.obj,
asarray(idata).as_BFarray(),
asarray(odata).as_BFarray()))
return odata
def matmul(self, alpha, a, b, beta, c):
"""Computes:
c = alpha*a.b + beta*c
or if b is None:
c = alpha*a.a^H + beta*c
or if a is None:
c = alpha*b^H.b + beta*c
where '.' is matrix product and '^H' is Hermitian transpose.
Multi-dimensional semantics are the same as numpy.matmul:
The last two dims represent the matrix, and all other dims are
used as batch dims to be matched or broadcast between a and b.
"""
if alpha is None:
alpha = 1.
if beta is None:
beta = 0.
beta = float(beta)
alpha = float(alpha)
a_array = asarray(a).as_BFarray() if a is not None else None
b_array = asarray(b).as_BFarray() if b is not None else None
c_array = asarray(c).as_BFarray()
_check(_bf.bfLinAlgMatMul(self.obj,
alpha,
a_array,
b_array,
beta,
c_array))
return c
def reduce(idata, odata, op='sum'):
if op not in REDUCE_MAP:
raise ValueError("Invalid reduce op: " + str(op))
op = REDUCE_MAP[op]
_check(_bf.bfReduce(asarray(idata).as_BFarray(),
asarray(odata).as_BFarray(),
op))
def transpose(dst, src, axes=None):
if axes is None:
axes = reversed(range(len(dst.shape)))
dst_bf = asarray(dst).as_BFarray()
src_bf = asarray(src).as_BFarray()
array_type = ctypes.c_int * src.ndim
axes_array = array_type(*axes)
_check(_bf.bfTranspose(src_bf, dst_bf, axes_array))
def memset2D(dst, val=0):
assert (len(dst.shape) == 2)
space = _string2space(_get_space(dst))
height, width = dst.shape
width_bytes = width * dst.dtype.itemsize
_check(
_bf.bfMemset2D(dst.ctypes.data, dst.strides[0], space, val,
width_bytes, height))
def resize(self,
contiguous_span,
total_span=None,
nringlet=1,
buffer_factor=4):
if total_span is None:
total_span = contiguous_span * buffer_factor
_check(
_bf.bfRingResize(self.obj, contiguous_span, total_span, nringlet))
def execute_workspace(self, iarray, oarray, workspace_ptr, workspace_size,
inverse=False):
_check(_bf.bfFftExecute(
self.obj,
asarray(iarray).as_BFarray(),
asarray(oarray).as_BFarray(),
inverse,
workspace_ptr, workspace_size))
return oarray
def set_openmp_cores(cores):
# PYCLIBRARY ISSUE
# TODO: Would be really nice to be able to directly pass
# a list here instead of needing to specify _array+type.
# Should be able to do it safely for any const* argument
# Note that the base type of the pointer type could be
# derived via a reverse lookup table.
# E.g., Inverse of POINTER(c_int)-->LP_c_int
_check(_bf.bfAffinitySetOpenMPCores(len(cores), _array(cores, 'int')))
def __init__(self, ring, sequence, nframe, nonblocking=False):
SpanBase.__init__(self, ring, sequence, writeable=True)
nbyte = nframe * self._sequence.tensor['frame_nbyte']
self.obj = _bf.BFwspan()
_check(_bf.bfRingSpanReserve(self.obj, ring.obj, nbyte, nonblocking))
self._set_base_obj(self.obj)
# Note: We default to 0 instead of nframe so that we don't accidentally
# commit bogus data if a block throws an exception.
self.commit_nframe = 0
def execute(self, idata, odata, negative_delays=False):
# TODO: Work out how to integrate CUDA stream
psize = None
_check(
_bf.bfFdmtExecute(self.obj,
asarray(idata).as_BFarray(),
asarray(odata).as_BFarray(), negative_delays,
None, psize))
return odata
def address(self):
buflen = 128
buf = ctypes.create_string_buffer(buflen)
_check(_bf.bfAddressGetString(self.obj, buflen, buf))
try:
value = buf.value.decode()
except AttributeError:
# Python2 catch
value = buf.value
return value
def memcpy(dst, src):
assert (dst.flags['C_CONTIGUOUS'])
assert (src.shape == dst.shape)
dst_space = _string2space(_get_space(dst))
src_space = _string2space(_get_space(src))
count = dst.nbytes
_check(
_bf.bfMemcpy(dst.ctypes.data, dst_space, src.ctypes.data, src_space,
count))
return dst
def memcpy2D(dst, src):
assert (len(dst.shape) == 2)
assert (src.shape == dst.shape)
dst_space = _string2space(_get_space(dst))
src_space = _string2space(_get_space(src))
height, width = dst.shape
width_bytes = width * dst.dtype.itemsize
_check(
_bf.bfMemcpy2D(dst.ctypes.data, dst.strides[0], dst_space,
src.ctypes.data, src.strides[0], src_space, width_bytes,
height))
def __init__(self, sequence, frame_offset, nframe):
SpanBase.__init__(self, sequence.ring, sequence, writeable=False)
tensor = sequence.tensor
self.obj = _bf.BFrspan()
_check(
_bf.bfRingSpanAcquire(self.obj, sequence.obj,
frame_offset * tensor['frame_nbyte'],
nframe * tensor['frame_nbyte']))
self._set_base_obj(self.obj)
self.nframe_skipped = min(self.frame_offset - frame_offset, nframe)
self.requested_frame_offset = frame_offset
def copy_array(dst, src):
dst_bf = asarray(dst)
src_bf = asarray(src)
if (space_accessible(dst_bf.bf.space, ['system'])
and space_accessible(src_bf.bf.space, ['system'])):
np.copyto(dst_bf, src_bf)
else:
_check(_bf.bfArrayCopy(dst_bf.as_BFarray(), src_bf.as_BFarray()))
if dst_bf.bf.space != src_bf.bf.space:
# TODO: Decide where/when these need to be called
device.stream_synchronize()
return dst
def on_data(self, ispan):
# TODO: Make this work in CUDA space
dada_blk = next(self.hdu.data_block)
nbyte = ispan.data.nbytes
_check(
_bf.bfMemcpy(dada_blk.ptr, _bf.BF_SPACE_SYSTEM,
ispan.data.ctypes.data, _bf.BF_SPACE_SYSTEM, nbyte))
#dada_blk.data[:] = ispan.data.view('u8')
dada_blk.close()
def execute_workspace(self,
idata,
odata,
workspace_ptr,
workspace_size,
negative_delays=False):
size = _bf.BFsize(workspace_size)
_check(
_bf.bfFdmtExecute(self.obj,
asarray(idata).as_BFarray(),
asarray(odata).as_BFarray(), negative_delays,
workspace_ptr, size))
return odata
def __init__(self,
ring,
which='specific',
name="",
time_tag=None,
other_obj=None,
guarantee=True):
SequenceBase.__init__(self, ring)
self._ring = ring
self.obj = _bf.BFrsequence()
if which == 'specific':
_check(_bf.bfRingSequenceOpen(self.obj, ring.obj, name, guarantee))
elif which == 'latest':
_check(_bf.bfRingSequenceOpenLatest(self.obj, ring.obj, guarantee))
elif which == 'earliest':
_check(
_bf.bfRingSequenceOpenEarliest(self.obj, ring.obj, guarantee))
elif which == 'at':
_check(
_bf.bfRingSequenceOpenAt(self.obj, ring.obj, time_tag,
guarantee))
#elif which == 'next':
# self._check( self.lib.bfRingSequenceOpenNext(pointer(self.obj), other_obj) )
else:
raise ValueError(
"Invalid 'which' parameter; must be one of: 'specific', 'latest', 'earliest'"
)
def __init__(self,
ring,
which='specific',
name="",
time_tag=None,
other_obj=None,
guarantee=True,
header_transform=None):
SequenceBase.__init__(self, ring)
self._ring = ring
# A function for transforming the header before it's read
self.header_transform = header_transform
self.obj = _bf.BFrsequence()
if which == 'specific':
_check(_bf.bfRingSequenceOpen(self.obj, ring.obj, name, guarantee))
elif which == 'at':
assert (time_tag is not None)
_check(
_bf.bfRingSequenceOpenAt(self.obj, ring.obj, time_tag,
guarantee))
elif which == 'latest':
_check(_bf.bfRingSequenceOpenLatest(self.obj, ring.obj, guarantee))
elif which == 'earliest':
_check(
_bf.bfRingSequenceOpenEarliest(self.obj, ring.obj, guarantee))
else:
raise ValueError(
"Invalid 'which' parameter; must be one of: 'specific', 'latest', 'earliest'"
)
def update(self, contents):
"""Updates (replaces) the contents of the log
contents: string or dict containing data to write to the log
"""
if contents is None:
raise ValueError("Contents cannot be None")
if isinstance(contents, dict):
contents = '\n'.join(
['%s : %s' % item for item in contents.items()])
try:
contents = contents.encode()
except AttributeError:
# Python2 catch
pass
_check(_bf.bfProcLogUpdate(self.obj, contents))
def copy(self, space=None, order='C'):
if order != 'C':
raise NotImplementedError('Only order="C" is supported')
if space is None:
space = self.bf.space
if not self.flags['C_CONTIGUOUS']:
# Deal with arrays that need to have their layouts changed
# TODO: Is there a better way to handle this?
if space_accessible(self.bf.space, ['system']):
## For arrays that can be accessed from the system space, use
## numpy.ndarray.copy() to do the heavy lifting
if space == 'cuda_managed':
## TODO: Decide where/when these need to be called
device.stream_synchronize()
## This actually makes two copies and throws one away
temp = ndarray(shape=self.shape, dtype=self.dtype, space=self.bf.space)
temp[...] = np.array(self).copy()
if self.bf.space != space:
return ndarray(temp, space=space)
return temp
else:
## For arrays that can be access from CUDA, use bifrost.transpose
## to do the heavy lifting
### Figure out the correct axis order for C
permute = np.argsort(self.strides)[::-1]
c_shape = [self.shape[p] for p in permute]
### Make a BFarray wrapper for self so we can reset shape/strides
### to what they should be for a C ordered array
self_corder = self.as_BFarray()
shape_type = ctypes.c_long*_bf.BF_MAX_DIMS
self_corder.shape = shape_type(*c_shape)
self_corder.strides = shape_type(*[self.strides[p] for p in permute])
### Make a temporary array with the right shape that will be C ordered
temp = ndarray(shape=self.shape, dtype=self.dtype, space=self.bf.space)
### Run the transpose using the BFarray wrapper and the temporary array
array_type = ctypes.c_int * self.ndim
axes_array = array_type(*permute)
_check(_bf.bfTranspose(self_corder, temp.as_BFarray(), axes_array))
if self.bf.space != space:
return ndarray(temp, space=space)
return temp
# Note: This makes an actual copy as long as space is not None
return ndarray(self, space=space)
def __init__(self, space='system', name=None, core=None):
if name is None:
name = str(uuid4())
name = _slugify(name)
try:
name = name.encode()
except AttributeError:
# Python2 catch
pass
space = _string2space(space)
#self.obj = None
#self.obj = _get(_bf.bfRingCreate(name=name, space=space), retarg=0)
BifrostObject.__init__(self, _bf.bfRingCreate, _bf.bfRingDestroy, name,
space)
if core is not None:
try:
_check(_bf.bfRingSetAffinity(self.obj, core))
except RuntimeError:
pass
def readinto(self, buf):
dst_space = Space(_get_space(buf)).as_BFspace()
byte0 = 0
nbyte = buf.nbytes
nbyte_copy = min(nbyte - byte0, self.nbyte - self.byte0)
while nbyte_copy:
_check(
_bf.bfMemcpy(buf.ctypes.data + byte0, dst_space,
self.block.ptr + self.byte0, _bf.BF_SPACE_SYSTEM,
nbyte_copy))
byte0 += nbyte_copy
self.byte0 += nbyte_copy
nbyte_copy = min(nbyte - byte0, self.nbyte - self.byte0)
if self.nbyte - self.byte0 == 0:
self.block.close()
try:
self._open_next_block()
except StopIteration:
break
return byte0
def __init__(self, space='system', name=None, owner=None, core=None):
# If this is non-None, then the object is wrapping a base Ring instance
self.base = None
self.space = space
if name is None:
name = 'ring_%i' % Ring.instance_count
Ring.instance_count += 1
name = _slugify(name)
try:
name = name.encode()
except AttributeError:
# Python2 catch
pass
BifrostObject.__init__(self, _bf.bfRingCreate, _bf.bfRingDestroy, name,
_string2space(self.space))
if core is not None:
try:
_check(_bf.bfRingSetAffinity(self.obj, core))
except RuntimeError:
pass
self.owner = owner
self.header_transform = None
def quantize(src, dst, scale=1.):
src_bf = asarray(src).as_BFarray()
dst_bf = asarray(dst).as_BFarray()
_check(_bf.bfQuantize(src_bf, dst_bf, scale))
return dst
def stream_synchronize():
_check(_bf.bfStreamSynchronize())
