def load_to_buffer(self, file, buffer):
""" Decompress data from a file to a provided buffer. """
_ = dill.load(file)
compressed = file.read()
blosc.decompress_ptr(compressed, buffer.__array_interface__['data'][0])
return buffer
def decompress_ndarray(binary: bytes, output_array=None) -> numpy.ndarray:
"""
decompress_ndarray(binary[, output_array=None])
Decompress array from buffer.
Data will write in output_array if it is provided. Save time request memory spaces and avoid out of memory.
"""
assert type(binary) is bytes
assert type(output_array) in [numpy.ndarray, numpy.memmap, type(None)]
buffer, dtype, shape = pickle.loads(binary)
if output_array is None:
output_array = numpy.empty(shape, dtype)
else:
assert dtype == output_array.dtype
if type(output_array) is numpy.ndarray:
output_array.resize(shape, refcheck=False)
else:
output_array = numpy.memmap(output_array.filename, shape=shape, dtype=dtype)
assert output_array.shape == shape, "Array output's shape is't same as compressed array."
blosc.decompress_ptr(buffer, output_array.__array_interface__['data'][0])
return output_array
def decompress_pre(Path, arr):
f = open(Path, "rb")
shape, dtype = pickle.load(f)
c = f.read()
#array allocation takes most of the time
blosc.decompress_ptr(c, arr.__array_interface__['data'][0])
return arr
def test_decompress_ptr_input_types(self):
import numpy as np
# assume the expected answer was compressed from bytes
expected = b'0123456789'
out = np.zeros(len(expected), dtype=np.byte)
compressed = blosc.compress(expected, typesize=1)
# now for all the things that support the buffer interface
out[:] = 0 # reset the output array
nout = blosc.decompress_ptr(compressed, out.ctypes.data)
self.assertEqual(expected, out.tobytes())
self.assertEqual(len(expected),
nout) # check that we didn't write too many bytes
out[:] = 0
nout = blosc.decompress_ptr(memoryview(compressed), out.ctypes.data)
self.assertEqual(expected, out.tobytes())
self.assertEqual(len(expected), nout)
out[:] = 0
nout = blosc.decompress_ptr(bytearray(compressed), out.ctypes.data)
self.assertEqual(expected, out.tobytes())
self.assertEqual(len(expected), nout)
out[:] = 0
nout = blosc.decompress_ptr(np.frombuffer(compressed, dtype=np.byte),
out.ctypes.data)
self.assertEqual(expected, out.tobytes())
self.assertEqual(len(expected), nout)
def load(self, file):
""" Read the shape and dtype from a file, create a buffer, decompress data into it. """
shape, dtype = dill.load(file)
compressed = file.read()
array = np.empty(shape=shape, dtype=dtype)
blosc.decompress_ptr(compressed, array.__array_interface__['data'][0])
return array
def decompression(self):
""" data decompression
Return
-------------
uncompressed numpy array
"""
c = empty(self.size, dtype=self.dtype)
decompress_ptr(self.data, c.__array_interface__['data'][0])
return c
def decompression(self):
""" data decompression
Return
-------------
uncompressed numpy array
"""
c = empty(self.size, dtype=self.dtype)
decompress_ptr(self.data, c.__array_interface__['data'][0])
return c
def __getitem__(self, index):
fname = os.path.join(self.dataname, self.filenames[index])
if self.is_compressed:
with open(fname, 'rb') as f:
arr = numpy.empty(self.ndim, dtype=numpy.float64)
blosc.decompress_ptr(f.read(),
arr.__array_interface__['data'][0])
return torch.from_numpy(arr)
else:
return torch.from_numpy(numpy.load(fname))
def test_codec(chunk, codec, filter_name, clevel):
"""
Compress the chunk and return tested data.
Parameters
----------
chunk: bytes-like object (supporting the buffer interface)
The data to be compressed.
codec : string
The name of the compressor used internally in Blosc. It can be
any of the supported by Blosc ('blosclz', 'lz4', 'lz4hc',
'snappy', 'zlib', 'zstd' and maybe others too).
filter_name : int
The shuffle filter to be activated. Allowed values are
blosc.NOSHUFFLE, blosc.SHUFFLE and blosc.BITSHUFFLE.
clevel : int
The compression level from 0 (no compression) to 9
(maximum compression).
Returns
-------
out: tuple
The associated compression rate, compression speed and
decompression speed (in GB/s).
Raises
------
TypeError
If bytesobj doesn't support the buffer interface.
ValueError
If bytesobj is too long.
If typesize is not within the allowed range.
If clevel is not within the allowed range.
If cname is not a valid codec.
"""
t0 = time()
c = blosc.compress_ptr(chunk.__array_interface__['data'][0],
chunk.size,
chunk.dtype.itemsize,
clevel=clevel,
shuffle=filter_name,
cname=codec)
tc = time() - t0
out = np.empty(chunk.size, dtype=chunk.dtype)
times = []
for i in range(3):
t0 = time()
blosc.decompress_ptr(c, out.__array_interface__['data'][0])
times.append(time() - t0)
chunk_byte_size = chunk.size * chunk.dtype.itemsize
rate = chunk_byte_size / len(c)
c_speed = chunk_byte_size / tc / SPEED_UNIT
d_speed = chunk_byte_size / min(times) / SPEED_UNIT
# print(" *** %-8s, %-10s, CL%d *** %6.4f s / %5.4f s " %
# ( codec, blosc.filters[filter], clevel, tc, td), end='')
# print("\tCompr. ratio: %5.1fx" % rate)
return rate, c_speed, d_speed
def decompress_ndarray(binary: bytes, output_array=None) -> numpy.ndarray:
assert type(binary) is bytes
assert type(output_array) is numpy.ndarray
buffer, dtype, shape = pickle.loads(binary)
if not output_array:
output_array = numpy.empty(shape, dtype)
assert output_array.shape == shape, "Array output's shape is't same as compressed array."
blosc.decompress_ptr(buffer, output_array.__array_interface__['data'][0])
return output_array
def test_codec( chunk, codec, filter, clevel ):
"""
Compresses the array chunk with the given codec, filter and clevel
and return the compression time and rate.
Parameters
----------
chunk : bytes-like object (supporting the buffer interface)
The data to be compressed.
codec : string
The name of the compressor used internally in Blosc. It can be
any of the supported by Blosc ('blosclz', 'lz4', 'lz4hc',
'snappy', 'zlib', 'zstd' and maybe others too).
clevel : int
The compression level from 0 (no compression) to 9
(maximum compression).
shuffle : int
The shuffle filter to be activated. Allowed values are
blosc.NOSHUFFLE, blosc.SHUFFLE and blosc.BITSHUFFLE.
Returns
-------
out : tuple
The associated compression time, rate and decompression time.
Raises
------
TypeError
If bytesobj doesn't support the buffer interface.
ValueError
If bytesobj is too long.
If typesize is not within the allowed range.
If clevel is not within the allowed range.
If cname is not a valid codec.
"""
t0 = time()
c = blosc.compress_ptr(chunk.__array_interface__['data'][0],
chunk.size, chunk.dtype.itemsize,
clevel = clevel, shuffle = filter, cname = codec)
tc = time() - t0
out = np.empty(chunk.size, dtype = chunk.dtype)
t0 = time()
blosc.decompress_ptr(c, out.__array_interface__['data'][0])
td = time() - t0
rate = (chunk.size * chunk.dtype.itemsize / len(c))
assert ((chunk == out).all())
# print(" *** %-8s, %-10s, CL%d *** %6.4f s / %5.4f s " %
# ( codec, blosc.filters[filter], clevel, tc, td), end='')
# print("\tCompr. ratio: %5.1fx" % rate)
return (rate, tc, td)
def _expanded_copy(obj):
"""Expand arrays within dicts, tuples and lists, do not dig other objects for now
"""
if isinstance(obj, _SqueezedArray):
shape, dtype, comp = obj
array = np.empty(shape, dtype=dtype)
blosc.decompress_ptr(comp, array.__array_interface__['data'][0])
return array
tpe = type(obj)
if tpe is tuple or tpe is list:
return tpe(_expanded_copy(el) for el in obj)
if tpe is dict:
return tpe((k, _expanded_copy(v)) for k, v in obj.items())
return obj
def read_blosc(stream, out=None):
meta = read_json(stream)
shape = tuple(meta['shape'])
dtype = restore_dtype(meta['dtype'])
if out is None:
out = np.empty(shape, dtype)
elif not isinstance(out, np.ndarray):
raise TypeError('expected ndarray, got {}'.format(type(out).__name__))
elif out.shape != shape:
raise ValueError('incompatible shape: expected {}, got {}'.format(shape, out.shape))
elif out.dtype != dtype:
raise ValueError('incompatible dtype: expected {}, got {}'.format(dtype, out.dtype))
elif not out.flags.contiguous:
raise ValueError('expected contiguous array')
blosc.decompress_ptr(
stream.read(meta['length']),
out.__array_interface__['data'][0]
)
if out.dtype.type is np.record:
out = out.view(np.recarray)
return out
def receive_msg(self):
# receive (blocking) messages
self.log.debug("Waiting for multipart message")
frames = self.zmq_socket.recv_multipart(flags=0, copy=False, track=False)
self.log.debug(" Multipart message received. Length: %d", len(frames))
arr_desc = msgpack.unpackb(frames[0].bytes)
self.log.debug('Array description: %s', str(arr_desc))
self.log.debug("Unpacking numpy array from bytes")
# Create an empty numpy array placeholder to unpack the compressed array into
arr = np.empty(arr_desc['shape'], dtype=arr_desc['dtype'])
dest_arr_ptr = arr.__array_interface__['data'][0]
# Unfortunately the access to Frame.bytes makes a copy of the compressed data.
# As we only read the compressed data it is not strictly necessary to make
# a copy, however it seems impossible to get a string object out without making a copy...
# We would have to modify the blosc python bindings to add a decompress_ptr function which
# would work when given a python memoryview object (i.e. a pointer)
compressed_bytes = frames[1].bytes
blosc.decompress_ptr(compressed_bytes, dest_arr_ptr)
self.log.debug(" unpacked array: shape: %s", str(arr.shape))
attr = msgpack.unpackb(frames[2].bytes)
self.log.debug(" unpacked attributes: %s", str(attr))
return arr, attr
def decompress_ptr():
cx = blosc.compress_ptr(address, num_elements, typesize, clevel=0)
blosc.decompress_ptr(cx, address)
print("Using *** %s *** compressor::" % cname)
ctic = time.time()
c = blosc.pack_array(in_, clevel=clevel, shuffle=True, cname=cname)
ctoc = time.time()
dtic = time.time()
out = blosc.unpack_array(c)
dtoc = time.time()
assert ((in_ == out).all())
print(" Time for pack_array/unpack_array: %.3f/%.3f s." % \
(ctoc-ctic, dtoc-dtic), end='')
print("\tCompr ratio: %.2f" %
(in_.size * in_.dtype.itemsize * 1. / len(c)))
ctic = time.time()
c = blosc.compress_ptr(in_.__array_interface__['data'][0],
in_.size,
in_.dtype.itemsize,
clevel=clevel,
shuffle=True,
cname=cname)
ctoc = time.time()
out = np.empty(in_.size, dtype=in_.dtype)
dtic = time.time()
blosc.decompress_ptr(c, out.__array_interface__['data'][0])
dtoc = time.time()
assert ((in_ == out).all())
print(" Time for compress_ptr/decompress_ptr: %.3f/%.3f s." % \
(ctoc-ctic, dtoc-dtic), end='')
print("\tCompr ratio: %.2f" %
(in_.size * in_.dtype.itemsize * 1. / len(c)))
def decompress_ptr():
cx = blosc.compress_ptr(address, num_elements, typesize, clevel=0)
blosc.decompress_ptr(cx, address)
def put(self, compressed):
bwritten = blosc.decompress_ptr(compressed, self.ptr)
self.ptr += bwritten
return bwritten
def decompress_ndarray(binary, output_array=None) -> numpy.ndarray:
"""
decompress_ndarray(binary[, output_array=None])
Decompress array from buffer.
Data will write in output_array if it is provided. Save time request memory spaces and avoid out of memory.
Parameters
----------
binary: bytes
Numpy array bytes, which was compressed.
output_array: numpy.ndarray
Data after decompress will be write into output_array if give.
"""
if not isinstance(binary, bytes):
raise TypeError("Require byte type of input data.")
if output_array is not None and not isinstance(
output_array, (numpy.ndarray, numpy.memmap)):
raise TypeError("Require numpy.ndarray type of output array.")
cursor = 0
# get header_size
header_size = cvt_hex2dec(binary[cursor:cursor + 1])
cursor += 1
# get dtype_size
dtype_size = cvt_hex2dec(binary[cursor:cursor + 1])
cursor += 1
# get dtype
dtype = numpy.dtype(cvt_hex2str(binary[cursor:cursor + dtype_size]))
cursor += dtype_size
# get shape
shape = []
while 1:
if cursor >= header_size + 1:
break
shape.append(cvt_hex2dec(binary[cursor:cursor + 2]))
cursor += 2
if output_array is None:
output_array = numpy.empty(shape, dtype)
else:
if dtype != output_array.dtype:
raise TypeError("Type of output array and data aren't the same!")
if tuple(shape) != output_array.shape:
if isinstance(output_array, numpy.memmap):
output_array = numpy.memmap(output_array.filename,
shape=shape,
dtype=dtype)
else:
output_array.resize(shape, refcheck=False)
blosc.decompress_ptr(binary[cursor:],
output_array.__array_interface__['data'][0])
return output_array
def put(self, compressed):
bwritten = blosc.decompress_ptr(compressed, self.ptr)
self.ptr += bwritten
return bwritten
def compressed_bytes_to_arraydata(cbytes, size, dtype):
arraydata = np.empty(size, dtype=dtype)
blosc.decompress_ptr(cbytes, arraydata.__array_interface__['data'][0])
return arraydata
arrays = [None]*3
labels = [None]*3
arrays[0] = np.arange(N, dtype=np.int64)
labels[0] = "the arange linear distribution"
arrays[1] = np.linspace(0, 1000, N)
labels[1] = "the linspace linear distribution"
arrays[2] = np.random.random_integers(0, 1000, N)
labels[2] = "the random distribution"
tic = time.time()
out_ = np.copy(arrays[0])
toc = time.time()
print(" *** np.copy() **** Time for memcpy(): %.3f s" % (toc-tic,))
for (in_, label) in zip(arrays, labels):
print("\n*** %s ***" % label)
for cname in blosc.compressor_list():
ctic = time.time()
c = blosc.compress_ptr(in_.__array_interface__['data'][0],
in_.size, in_.dtype.itemsize,
clevel=clevel, shuffle=True, cname=cname)
ctoc = time.time()
out = np.empty(in_.size, dtype=in_.dtype)
dtic = time.time()
blosc.decompress_ptr(c, out.__array_interface__['data'][0])
dtoc = time.time()
assert((in_ == out).all())
print(" *** %-8s *** Time for comp/decomp: %.3f/%.3f s." % \
(cname, ctoc-ctic, dtoc-dtic), end='')
print("\tCompr ratio: %6.2f" % (in_.size*in_.dtype.itemsize*1. / len(c)))
def unpack(col):
a = np.empty(col['size'], dtype=col['dtype'])
blosc.decompress_ptr(bytes(col['data']), a.__array_interface__['data'][0])
return a
def decompress(size, dtype, data):
out = np.empty(size, dtype)
blosc.decompress_ptr(data, out.__array_interface__['data'][0])
return out
def decompress(self, blocks, out, **kwargs):
'''Useful decompression kwargs:
nthreads
'''
# TODO: controlled globally for now
#nthreads = kwargs.pop('nthreads',1)
#blosc.set_nthreads(nthreads)
_size = 0
_pos = 0
_buffer = None
_partial_len = b''
decompression_time = 0.
bytesout = 0
# Blosc code probably assumes contiguous buffer
if not out.contiguous:
raise ValueError(out.contiguous)
# get the out address
out = np.frombuffer(out, dtype=np.uint8).ctypes.data
for block in blocks:
block = memoryview(block).cast('c')
try:
block = block.toreadonly() # python>=3.8 only
except AttributeError:
pass
if not block.contiguous:
raise ValueError(block.contiguous)
while len(block):
if not _size:
# Don't know the (compressed) length of this block yet
if len(_partial_len) + len(block) < 4:
_partial_len += block
break # we've exhausted the data
if _partial_len:
# If we started to fill a len key, finish filling it
remaining = 4 - len(_partial_len)
if remaining:
_partial_len += block[:remaining]
block = block[remaining:]
_size = struct.unpack('!I', _partial_len)[0]
_partial_len = b''
else:
# Otherwise just read the len key directly
_size = struct.unpack('!I', block[:4])[0]
block = block[4:]
if len(block) < _size or _buffer is not None:
# If we have a partial block, or we're already filling a buffer, use the buffer
if _buffer is None:
_buffer = np.empty(
_size, dtype=np.byte
) # use numpy instead of bytearray so we can avoid zero initialization
_pos = 0
newbytes = min(
_size - _pos,
len(block)) # don't fill past the buffer len!
_buffer[_pos:_pos + newbytes] = np.frombuffer(
block[:newbytes], dtype=np.byte)
_pos += newbytes
block = block[newbytes:]
if _pos == _size:
start = time.perf_counter()
n_thisout = blosc.decompress_ptr(
memoryview(_buffer), out + bytesout, **kwargs)
decompression_time += time.perf_counter() - start
bytesout += n_thisout
_buffer = None
_size = 0
else:
# We have at least one full block
start = time.perf_counter()
n_thisout = blosc.decompress_ptr(memoryview(block[:_size]),
out + bytesout, **kwargs)
decompression_time += time.perf_counter() - start
bytesout += n_thisout
block = block[_size:]
_size = 0
return bytesout
def decompress(shape, dtype, compressed):
array = np.empty(shape, dtype=dtype)
blosc.decompress_ptr(compressed, array.__array_interface__['data'][0])
return BloscItem(array)
