The goal of GetPy is to provide the highest performance python dict/set that integrates into the python scientific ecosystem.
pip install getpy
Note only a linux build is currently distributed. If you would like to build the package from source you can clone the repo and run python setup.py install. Compilation will require 16gb of ram. I am working on getting that down.
GetPy is a thin binding to the Parallel Hashmap (https://github.com/greg7mdp/parallel-hashmap.git) which is the current state of the art unordered map/set with minimal memory overhead and fast runtime speed. The binding layer is supported by PyBind11 (https://github.com/pybind/pybind11.git) which is fast to compile and simple to extend.
The gp.Dict and gp.Set objects are designed to maintain a similar interface to the corresponding standard python objects. There are some key differences though, which are necessary for vectorization and other performance considerations.
-
gp.Dict.__init__has three argumentskey_type,value_type, anddefault_value. The type arguments are define which compiled data structure will be used under the hood, and the full list of preset combinations ofnp.dtypes is found withgp.dict_types. You can also specify adefault_valueat construction which must be castable to thevalue_type. This is the value returned by the dictionary if a key is not found. -
All of
getpy.Dictmethods support a vectorized interface. Therefore, methods likegp.Dict.__getitem__,gp.Dict.__setitem__, andgp.Dict.__delitem__can be performed with annp.ndarray. That allows the performance critical for-loop to happen within the compiled c++. Note that some dunder methods cannot be vectorized such as__contains__. Therefore, some keywords likeindo not behave as expected. Those methods are renamed without the double underscores to note their deviation from the standard interface. -
If a key does not exist,
gp.Dict.__getitem__will return thedefault_value. If you do not specify thedefault_value, it will default to the default constructor of your data type (all 0 bits). If you would like to know the difference between a key that does not exist and a key that returns the default value, you should first rungp.containson your key/array of keys, and then retrieve values corresponding to keys that exist. -
There is also a
gp.MultiDictobject. This object stores multiple unique values per key.
import numpy as np
import getpy as gp
key_type = np.dtype('u8')
value_type = np.dtype('u8')
keys = np.random.randint(1, 1000, size=10**2, dtype=key_type)
values = np.random.randint(1, 1000, size=10**2, dtype=value_type)
gp_dict = gp.Dict(key_type, value_type)
gp_dict[keys] = valuesimport numpy as np
import getpy as gp
key_type = np.dtype('u8')
value_type = np.dtype('u8')
keys = np.random.randint(1, 1000, size=10**2, dtype=key_type)
values = np.random.randint(1, 1000, size=10**2, dtype=value_type)
gp_dict = gp.Dict(key_type, value_type, default_value=42)
gp_dict[keys] = values
random_keys = np.random.randint(1, 1000, size=500, dtype=key_type)
random_values = gp_dict[random_keys]import numpy as np
import getpy as gp
key_type = np.dtype('S8')
value_type = np.dtype('S8')
keys = np.array([np.random.bytes(8) for i in range(10**2)], dtype=key_type)
values = np.array([np.random.bytes(8) for i in range(10**2)], dtype=value_type)
gp_dict = gp.Dict(key_type, value_type)
gp_dict[keys] = valuesimport numpy as np
import getpy as gp
key_type = np.dtype('u8')
value_type = np.dtype('u8')
keys = np.random.randint(1, 1000, size=10**2, dtype=key_type).reshape(10,10)
values = np.random.randint(1, 1000, size=10**2, dtype=value_type).reshape(10,10)
gp_dict = gp.Dict(key_type, value_type)
gp_dict[keys] = valuesimport numpy as np
import getpy as gp
key_type = np.dtype('u8')
value_type = np.dtype('u8')
keys = np.random.randint(1, 1000, size=10**2, dtype=np.dtype('u2')).reshape(25,4).view(key_type)
values = np.random.randint(1, 1000, size=(10**2)/2, dtype=np.dtype('u4')).reshape(25,2).view(value_type)
gp_dict = gp.Dict(key_type, value_type)
gp_dict[keys] = values
unpacked_values = gp_dict[keys].view(np.dtype('u4'))import numpy as np
import getpy as gp
key_type = np.dtype('u8')
value_type = np.dtype('u8')
keys = np.random.randint(1, 1000, size=10**1, dtype=key_type)
values = np.random.randint(1, 1000, size=10**1, dtype=value_type)
gp_dict_1 = gp.Dict(key_type, value_type)
gp_dict_1[keys] = values
gp_dict_1.dump('test/test.hashtable.bin')
gp_dict_2 = gp.Dict(key_type, value_type)
gp_dict_2.load('test/test.hashtable.bin')dict_types = {
(np.dtype('u4'), np.dtype('u1')) : _gp.Dict_u4_u1,
(np.dtype('u4'), np.dtype('u2')) : _gp.Dict_u4_u2,
(np.dtype('u4'), np.dtype('u4')) : _gp.Dict_u4_u4,
(np.dtype('u4'), np.dtype('u8')) : _gp.Dict_u4_u8,
(np.dtype('u4'), np.dtype('i1')) : _gp.Dict_u4_i1,
(np.dtype('u4'), np.dtype('i2')) : _gp.Dict_u4_i2,
(np.dtype('u4'), np.dtype('i4')) : _gp.Dict_u4_i4,
(np.dtype('u4'), np.dtype('i8')) : _gp.Dict_u4_i8,
(np.dtype('u4'), np.dtype('f4')) : _gp.Dict_u4_f4,
(np.dtype('u4'), np.dtype('f8')) : _gp.Dict_u4_f8,
(np.dtype('u4'), np.dtype('S8')) : _gp.Dict_u4_S8,
(np.dtype('u4'), np.dtype('S16')) : _gp.Dict_u4_S16,
(np.dtype('u8'), np.dtype('u1')) : _gp.Dict_u8_u1,
(np.dtype('u8'), np.dtype('u2')) : _gp.Dict_u8_u2,
(np.dtype('u8'), np.dtype('u4')) : _gp.Dict_u8_u4,
(np.dtype('u8'), np.dtype('u8')) : _gp.Dict_u8_u8,
(np.dtype('u8'), np.dtype('i1')) : _gp.Dict_u8_i1,
(np.dtype('u8'), np.dtype('i2')) : _gp.Dict_u8_i2,
(np.dtype('u8'), np.dtype('i4')) : _gp.Dict_u8_i4,
(np.dtype('u8'), np.dtype('i8')) : _gp.Dict_u8_i8,
(np.dtype('u8'), np.dtype('f4')) : _gp.Dict_u8_f4,
(np.dtype('u8'), np.dtype('f8')) : _gp.Dict_u8_f8,
(np.dtype('u8'), np.dtype('S8')) : _gp.Dict_u8_S8,
(np.dtype('u8'), np.dtype('S16')) : _gp.Dict_u8_S16,
(np.dtype('i4'), np.dtype('u1')) : _gp.Dict_i4_u1,
(np.dtype('i4'), np.dtype('u2')) : _gp.Dict_i4_u2,
(np.dtype('i4'), np.dtype('u4')) : _gp.Dict_i4_u4,
(np.dtype('i4'), np.dtype('u8')) : _gp.Dict_i4_u8,
(np.dtype('i4'), np.dtype('i1')) : _gp.Dict_i4_i1,
(np.dtype('i4'), np.dtype('i2')) : _gp.Dict_i4_i2,
(np.dtype('i4'), np.dtype('i4')) : _gp.Dict_i4_i4,
(np.dtype('i4'), np.dtype('i8')) : _gp.Dict_i4_i8,
(np.dtype('i4'), np.dtype('f4')) : _gp.Dict_i4_f4,
(np.dtype('i4'), np.dtype('f8')) : _gp.Dict_i4_f8,
(np.dtype('i4'), np.dtype('S8')) : _gp.Dict_i4_S8,
(np.dtype('i4'), np.dtype('S16')) : _gp.Dict_i4_S16,
(np.dtype('i8'), np.dtype('u1')) : _gp.Dict_i8_u1,
(np.dtype('i8'), np.dtype('u2')) : _gp.Dict_i8_u2,
(np.dtype('i8'), np.dtype('u4')) : _gp.Dict_i8_u4,
(np.dtype('i8'), np.dtype('u8')) : _gp.Dict_i8_u8,
(np.dtype('i8'), np.dtype('i1')) : _gp.Dict_i8_i1,
(np.dtype('i8'), np.dtype('i2')) : _gp.Dict_i8_i2,
(np.dtype('i8'), np.dtype('i4')) : _gp.Dict_i8_i4,
(np.dtype('i8'), np.dtype('i8')) : _gp.Dict_i8_i8,
(np.dtype('i8'), np.dtype('f4')) : _gp.Dict_i8_f4,
(np.dtype('i8'), np.dtype('f8')) : _gp.Dict_i8_f8,
(np.dtype('i8'), np.dtype('S8')) : _gp.Dict_i8_S8,
(np.dtype('i8'), np.dtype('S16')) : _gp.Dict_i8_S16,
(np.dtype('S8'), np.dtype('u1')) : _gp.Dict_S8_u1,
(np.dtype('S8'), np.dtype('u2')) : _gp.Dict_S8_u2,
(np.dtype('S8'), np.dtype('u4')) : _gp.Dict_S8_u4,
(np.dtype('S8'), np.dtype('u8')) : _gp.Dict_S8_u8,
(np.dtype('S8'), np.dtype('i1')) : _gp.Dict_S8_i1,
(np.dtype('S8'), np.dtype('i2')) : _gp.Dict_S8_i2,
(np.dtype('S8'), np.dtype('i4')) : _gp.Dict_S8_i4,
(np.dtype('S8'), np.dtype('i8')) : _gp.Dict_S8_i8,
(np.dtype('S8'), np.dtype('f4')) : _gp.Dict_S8_f4,
(np.dtype('S8'), np.dtype('f8')) : _gp.Dict_S8_f8,
(np.dtype('S8'), np.dtype('S8')) : _gp.Dict_S8_S8,
(np.dtype('S8'), np.dtype('S16')) : _gp.Dict_S8_S16,
(np.dtype('S16'), np.dtype('u1')) : _gp.Dict_S16_u1,
(np.dtype('S16'), np.dtype('u2')) : _gp.Dict_S16_u2,
(np.dtype('S16'), np.dtype('u4')) : _gp.Dict_S16_u4,
(np.dtype('S16'), np.dtype('u8')) : _gp.Dict_S16_u8,
(np.dtype('S16'), np.dtype('i1')) : _gp.Dict_S16_i1,
(np.dtype('S16'), np.dtype('i2')) : _gp.Dict_S16_i2,
(np.dtype('S16'), np.dtype('i4')) : _gp.Dict_S16_i4,
(np.dtype('S16'), np.dtype('i8')) : _gp.Dict_S16_i8,
(np.dtype('S16'), np.dtype('f4')) : _gp.Dict_S16_f4,
(np.dtype('S16'), np.dtype('f8')) : _gp.Dict_S16_f8,
(np.dtype('S16'), np.dtype('S8')) : _gp.Dict_S16_S8,
(np.dtype('S16'), np.dtype('S16')) : _gp.Dict_S16_S16,
}
set_types = {
np.dtype('u4') : _gp.Set_u4,
np.dtype('u8') : _gp.Set_u8,
np.dtype('i4') : _gp.Set_i4,
np.dtype('i8') : _gp.Set_i8,
np.dtype('S8') : _gp.Set_S8,
np.dtype('S16') : _gp.Set_S16,
}