def test_get_cbuffer_sizes(self):
s = b'0123456789' * 100000
blosc.set_blocksize(2**16)
c = blosc.compress(s, typesize=1)
t = blosc.get_cbuffer_sizes(c)
self.assertEqual(t[0], 1000000)
# One cannot be sure of the exact compressed bytes, so round to KB
self.assertEqual(t[1] // 2**10, 4354 // 2**10)
self.assertEqual(t[2], 2**16)
def test_get_cbuffer_sizes(self):
s = b'0123456789' * 100000
blosc.set_blocksize(2**16)
c = blosc.compress(s, typesize=1)
t = blosc.get_cbuffer_sizes(c)
self.assertEqual(t[0], 1000000)
# One cannot be sure of the exact compressed bytes, so round to KB
self.assertEqual(t[1] // 2**10, 4354 // 2**10)
self.assertEqual(t[2], 2**16)
def compress(self, data, **kwargs):
'''Useful compression kwargs:
nthreads
compression_block_size
blosc_block_size
shuffle
typesize
cname
clevel
'''
# Blosc code probably assumes contiguous buffer
assert data.contiguous
nthreads = kwargs.pop('nthreads', 1)
compression_block_size = kwargs.pop('compression_block_size', 1 << 22)
blosc_block_size = kwargs.pop('blosc_block_size', 512 * 1024)
typesize = kwargs.pop('typesize',
'auto') # dtype size in bytes, e.g. 8 for int64
clevel = kwargs.pop(
'clevel',
1) # compression level, usually only need lowest for zstd
cname = kwargs.pop(
'cname', 'zstd'
) # compressor name, default zstd, good performance/compression tradeoff
shuffle = kwargs.pop('shuffle', 'shuffle')
if shuffle == 'shuffle':
shuffle = blosc.SHUFFLE
elif shuffle == 'bitshuffle':
shuffle = blosc.BITSHUFFLE
elif shuffle == None:
shuffle = blosc.NOSHUFFLE
else:
raise ValueError(shuffle)
blosc.set_nthreads(nthreads)
blosc.set_blocksize(blosc_block_size)
if typesize == 'auto':
this_typesize = data.itemsize
else:
this_typesize = typesize
#assert this_typesize != 1
nelem = compression_block_size // data.itemsize
for i in range(0, len(data), nelem):
compressed = blosc.compress(data[i:i + nelem],
typesize=this_typesize,
clevel=clevel,
shuffle=shuffle,
cname=cname,
**kwargs)
header = struct.pack('!I', len(compressed))
# TODO: this probably triggers a data copy, feels inefficient. Probably have to add output array arg to blosc to fix
yield header + compressed
def doCompression(dataStack,
compressor='zstd',
blocksize=2**20,
n_threads=16,
shuffle=blosc.BITSHUFFLE,
clevel=5):
blosc.set_blocksize(blocksize)
blosc.set_nthreads(n_threads)
typeSize = dataStack.dtype.itemsize
packedDataList = [None] * dataStack.shape[0]
for J in np.arange(dataStack.shape[0]):
packedDataList[J] = blosc.compress(dataStack[J, :, :],
typesize=typeSize,
clevel=clevel,
shuffle=shuffle,
cname=compressor)
return packedDataList
def test_get_blocksize(self):
s = b'0123456789' * 1000
blosc.set_blocksize(2**14)
blosc.compress(s, typesize=1)
d = blosc.get_blocksize()
self.assertEqual(d, 2**14)
def test_get_blocksize(self):
s = b'0123456789' * 1000
blosc.set_blocksize(2**14)
blosc.compress(s, typesize=1)
d = blosc.get_blocksize()
self.assertEqual(d, 2**14)
def __MRCExport(input_image, header, MRCfilename, slices,
endchar='<', offset=0, idxnewfile=True):
'''
MRCExport private interface with a dictionary rather than a mess of function
arguments.
'''
if idxnewfile: # If forcing a new file we truncate it even if it already exists:
fmode = 'wb'
else: # Otherwise we'll just update its header and append images as required:
fmode = 'rb+'
with open(MRCfilename, fmode, buffering=BUFFERSIZE) as f:
extendedBytes = writeMRCHeader(f, header, slices, endchar=endchar)
f.seek(DEFAULT_HEADER_LEN + extendedBytes + offset)
dtype = header['dtype']
if ('compressor' in header) \
and (header['compressor'] in REVERSE_COMPRESSOR_ENUM) \
and (REVERSE_COMPRESSOR_ENUM[header['compressor']]) > 0:
# compressed MRCZ
logger.debug('Compressing %s with compressor %s%d' %
(MRCfilename, header['compressor'], header['clevel']))
applyCast = False
if slices > 0:
chunkSize = input_image[0].size
typeSize = input_image[0].dtype.itemsize
if dtype != 'uint4' and input_image[0].dtype != dtype:
applyCast = True
else:
chunkSize = input_image[0,:,:].size
typeSize = input_image.dtype.itemsize
if dtype != 'uint4' and input_image.dtype != dtype:
applyCast = True
blosc.set_nthreads(header['n_threads'])
# for small image dimensions we need to scale blocksize appropriately
# so we use the available cores
block_size = np.minimum(BLOSC_BLOCK, chunkSize//header['n_threads'])
blosc.set_blocksize(block_size)
header['packedBytes'] = 0
clevel = header['clevel']
cname = header['compressor']
# For 3D frames in lists, we need to further sub-divide each frame
# into slices so that each channel is compressed seperately by
# blosc.
if slices > 1:
deep_image = input_image # grab a reference
input_image = []
for frame in deep_image:
for I in range(slices):
input_image.append(frame[I,:,:])
for J, frame in enumerate(input_image):
if applyCast:
frame = frame.astype(dtype)
if frame.flags['C_CONTIGUOUS'] and frame.flags['ALIGNED']:
# Use pointer
compressedData = blosc.compress_ptr(frame.__array_interface__['data'][0],
frame.size,
typeSize,
clevel=header['clevel'],
shuffle=blosc.BITSHUFFLE,
cname=header['compressor'])
else:
# Use tobytes, which is slower in benchmarking
compressedData = blosc.compress(frame.tobytes(),
typeSize,
clevel=clevel,
shuffle=blosc.BITSHUFFLE,
cname=cname)
f.write(compressedData)
header['packedBytes'] += len(compressedData)
# Rewind and write out the total compressed size
f.seek(144)
np.int64(header['packedBytes']).astype(endchar + 'i8').tofile(f)
else: # vanilla MRC
if slices > 0:
if dtype != 'uint4' and dtype != input_image[0].dtype:
for z_slice in input_image:
z_slice.astype(dtype).tofile(f)
else:
for z_slice in input_image:
z_slice.tofile(f)
else:
if dtype != 'uint4' and dtype != input_image.dtype:
input_image = input_image.astype(dtype)
input_image.tofile(f)
return
tPool.join()
def decompressStack(imageShape,
imageDtype,
blosc_threads=1,
pool_threads=maxThreads):
blosc.set_nthreads(blosc_threads)
tPool = ThreadPool(pool_threads)
num_slices = imageShape[0]
imageStack = np.full(imageShape, fill_value=0)
blosc.print_versions()
blosc.set_blocksize(BLOCKSIZE)
print("Creating NumPy stack with %d float32 elements:" % (m * N * N))
stack = np.zeros([m, N, N], dtype=dtype)
xmesh, ymesh = np.meshgrid(np.arange(-N / 2, N / 2), np.arange(-N / 2, N / 2))
compress_mesh = (np.cos(xmesh) + np.exp(-ymesh**2 / N)).astype(dtype)
for J in np.arange(m):
stack[J, :, :] = compress_mesh
### Determine arrangement of pool threads and blosc threads
testCases = int(np.floor(np.log2(maxThreads)) + 1)
powProduct = 2**np.arange(0, testCases)
poolThreads = np.hstack([1, powProduct])
bloscThreads = np.hstack([1, powProduct[::-1]])
# Let's try instead just pool threads...
#poolThreads = np.arange( 1, maxThreads+1 )
from effects import SnpEff
# native Python imports
import os.path
import time
import sys
import sqlite3
import itertools as it
import toml # toml.py
# third-party imports
import cyvcf2 as vcf
import blosc
blosc.set_nthreads(1)
blosc.set_blocksize(8192)
import zlib
import cPickle
def opack_blob(obj, _none=buffer(zlib.compress(cPickle.dumps(None, cPickle.HIGHEST_PROTOCOL)))):
if obj is None: return _none
return buffer(zlib.compress(cPickle.dumps(obj, cPickle.HIGHEST_PROTOCOL), 1))
def pack_blob(obj):
if obj is None: return ''
return buffer(blosc.compress(obj.tostring(), obj.dtype.itemsize, clevel=5, shuffle=True))
#return buffer(blosc.pack_array(obj))
def is_number(op, field):
return field.endswith("_float") or op in ("mean", "median", "min", "max")
def __MRCExport( input_image, header, MRCfilename, endchar = '<' ):
"""
MRCExport private interface with a dictionary rather than a mess of function
arguments.
"""
with open( MRCfilename, 'wb', buffering=BUFFERSIZE ) as f:
writeMRCHeader( f, header, endchar )
f.seek(1024)
if ('compressor' in header) \
and (header['compressor'] in REVERSE_COMPRESSOR_ENUM) \
and (REVERSE_COMPRESSOR_ENUM[header['compressor']]) > 0:
# compressed MRCZ
print( "Compressing %s with compressor %s%d" %
(MRCfilename, header['compressor'], header['clevel'] ) )
if header['dtype'] != 'uint4' and input_image.dtype != header['dtype']:
# This correctly works for text to dtype comparison
input_image = input_image.astype(header['dtype'])
if input_image.ndim == 3:
chunkSize = input_image[0,:,:].size
else:
chunkSize = input_image.size
input_image = np.reshape( input_image, [1,input_image.shape[0],input_image.shape[1] ])
blosc.set_nthreads( header['n_threads'] )
blosc.set_blocksize( 65536 )
header['packedBytes'] = 0
typeSize = input_image.dtype.itemsize
print( input_image.shape )
for J in np.arange( input_image.shape[0] ):
# print( "Slice %d: Compressing address at: %d of %d:" % (J, int(J*typeSize*blockSize), input_image.nbytes) )
# Looks like I have problem for typesize > 1?
if int(J*typeSize*chunkSize) >= input_image.nbytes:
raise MemoryError( "MRCExport: Tried to reference past end of ndarray %d > %d" % (int(J*typeSize*chunkSize), input_image.nbytes ) )
compressedData = blosc.compress( input_image[J,:,:].tobytes(),
typeSize,
clevel=header['clevel'],
shuffle=blosc.BITSHUFFLE,
cname=header['compressor'] )
f.write( compressedData )
header['packedBytes'] += len(compressedData)
# print( "packedBytes = %d" % header['packedBytes'] )
# print( "Finished writing out compressedData" )
# Rewind and write out the total compressed size
f.seek(144)
np.int64( header['packedBytes'] ).astype( endchar + "i8" ).tofile(f)
else: # vanilla MRC
if header['dtype'] != 'uint4' and input_image.dtype != header['dtype']:
input_image = input_image.astype( header['dtype'] )
input_image.tofile(f)
return
def test_get_cbuffer_sizes(self):
s = b'0123456789' * 100000
blosc.set_blocksize(2**16)
c = blosc.compress(s, typesize=1)
t = blosc.get_cbuffer_sizes(c)
self.assertEqual(t, (1000000, 4354, 2**16))
def __MRCExport(input_image,
header,
MRCfilename,
endchar='<',
offset=0,
idxnewfile=True):
'''
MRCExport private interface with a dictionary rather than a mess of function
arguments.
'''
if idxnewfile:
# If forcing a new file we truncate it even if it already exists:
fmode = 'wb'
else:
# Otherwise we'll just update its header and append images as required:
fmode = 'rb+'
with open(MRCfilename, fmode, buffering=BUFFERSIZE) as f:
extendedBytes = writeMRCHeader(f, header, endchar)
f.seek(DEFAULT_HEADER_LEN + extendedBytes + offset)
if ('compressor' in header) \
and (header['compressor'] in REVERSE_COMPRESSOR_ENUM) \
and (REVERSE_COMPRESSOR_ENUM[header['compressor']]) > 0:
# compressed MRCZ
logger.info('Compressing %s with compressor %s%d' %
(MRCfilename, header['compressor'], header['clevel']))
if header['dtype'] != 'uint4' and input_image.dtype != header[
'dtype']:
# This correctly works for text to dtype comparison
input_image = input_image.astype(header['dtype'])
if input_image.ndim == 3:
chunkSize = input_image[0, :, :].size
else:
chunkSize = input_image.size
input_image = np.reshape(
input_image,
[1, input_image.shape[0], input_image.shape[1]])
blosc.set_nthreads(header['n_threads'])
blosc.set_blocksize(BLOSC_BLOCK)
header['packedBytes'] = 0
typeSize = input_image.dtype.itemsize
for J in np.arange(input_image.shape[0]):
# print( 'Slice %d: Compressing address at: %d of %d:' % (J, int(J*typeSize*blockSize), input_image.nbytes) )
# Looks like I have problem for typesize > 1?
if int(J * typeSize * chunkSize) >= input_image.nbytes:
raise MemoryError(
'MRCExport: Tried to reference past end of ndarray %d > %d'
% (int(J * typeSize * chunkSize), input_image.nbytes))
compressedData = blosc.compress(input_image[J, :, :].tobytes(),
typeSize,
clevel=header['clevel'],
shuffle=blosc.BITSHUFFLE,
cname=header['compressor'])
f.write(compressedData)
header['packedBytes'] += len(compressedData)
# Rewind and write out the total compressed size
f.seek(144)
np.int64(header['packedBytes']).astype(endchar + 'i8').tofile(f)
else: # vanilla MRC
if header['dtype'] != 'uint4' and input_image.dtype != header[
'dtype']:
input_image = input_image.astype(header['dtype'])
input_image.tofile(f)
return
# All operations are done 'in-place'
tPool.map( compressSlice, tArgs )
tPool.close()
tPool.join()
def decompressStack( imageShape, imageDtype, blosc_threads = 1, pool_threads=maxThreads ):
blosc.set_nthreads( blosc_threads )
tPool = ThreadPool( pool_threads )
num_slices = imageShape[0]
imageStack = np.empty( imageShape )
blosc.print_versions()
blosc.set_blocksize( BLOCKSIZE )
print("Creating NumPy stack with %d float32 elements:" %(m*N*N) )
stack = np.zeros( [m,N,N], dtype=dtype )
xmesh, ymesh = np.meshgrid( np.arange(-N/2,N/2), np.arange(-N/2,N/2) )
compress_mesh = (np.cos( xmesh ) + np.exp( -ymesh**2 / N )).astype(dtype)
for J in np.arange(m):
stack[J,:,:] = compress_mesh
### Determine arrangement of pool threads and blosc threads
testCases = int( np.floor( np.log2( maxThreads )) + 1 )
powProduct = 2**np.arange(0,testCases)
poolThreads = np.hstack( [1, powProduct] )
bloscThreads = np.hstack( [1, powProduct[::-1]] )
# Let's try instead just pool threads...
if buffer.dtype.kind in ('S', 'U'):
is_string = True
filters = (blosc.NOSHUFFLE, )
else:
is_string = False
filters = (blosc.NOSHUFFLE, blosc.SHUFFLE, blosc.BITSHUFFLE)
for i, chunk in enumerate(chunk_generator(buffer)):
if is_string:
chunk_features = (0, 0, 0, 0, 0, 0, 0, 0, 0, 0)
else:
chunk_features = extract_chunk_features(chunk)
chunk_features += (calculate_streaks(chunk,
chunk_features[1]), )
df = pd.DataFrame()
for block_size in BLOCK_SIZES:
blosc.set_blocksize(block_size)
for codec in blosc.compressor_list():
for filter in filters:
for clevel in C_LEVELS:
row_data = (filename, path, table, d_type, i + 1,
chunk.size * chunk.dtype.itemsize / MB) \
+ chunk_features \
+ (block_size / 2**10, codec, blosc.filters[filter], clevel) \
+ test_codec(chunk, codec,
filter, clevel)
df = df.append(dict(zip(COLS, row_data)),
ignore_index=True)
print("%5.2f%% %-s %-s t%-s chunk %d completed" %
((i + 1) / n_chunks * 100, filename, path, table, (i + 1)))
with open('blosc_test_data.csv', 'a') as f:
df = df[COLS]
