Annotation aware numba njit.
Project description
numbakit-anjit: leveraging numba to speed up ODE integration
numbakit-anjit (nbkanjit) to assist Numba intensive project by providing anjit, an annotation aware numba jit decorator and manager object to handle Jit configuration.
It runs in Python 3.7+ depending on Numba. It is licensed under BSD.
It is extremely easy and natural to use:
>>> from numba import types as nt
>>> import nbkanjit
>>> @nbkanjit.anjit
... def func(x: nt.float64, y: nt.float64) -> nt.float64:
... return x + y
You can also use Python types:
>>> @nbkanjit.anjit
... def func(x: float, y: float) -> float:
... return x + y
which are mapped to numba types.
You can use:
>>> from nbkanjit import Function as F_
>>> @nbkanjit.anjit
... def func1(x: int, y: float) -> float:
... return x + y
>>> def func2(x: int, y: F_(funct1)._return) -> float:
... return x + y
You can also use the annotation of any argument. For example, F_(func).x in this case is equivalent to int. Or even the full function F_(func) that will return FunctionType(float64(int, float64))
It also provides a manager to encapsulate (and reuse different parameters)
>>> import nbkanjit
>>> jm = nbkanjit.JitManager(cache=True)
>>> @jm.anjit
... def func(x: float, y:float) -> nt.float64:
... return x + y
even to be applied in to the standard numba njit.
>>> jm = nbkanjit.JitManager(cache=True)
>>> @jm.njit
... def func(x, y):
... return x + y
And you can teach the manager new tricks:
>>> jm.mapping["array1d"] = nt.float64[:]
by mapping any python object into a numba type.
And a way to register a signature as a template (tmpl):
>>> import nbkanjit
>>> jm = nbkanjit.JitManager()
>>> jm.register("nice", nt.float64((nt.float64, nt.float64)))
and then use it for non-annotated function by explicitly name:
>>> @jm.njit_tmpl("nice")
... def other_func(x, y):
... return x + y
or using the name of the function:
>>> @jm.njit_tmpl
... def nice(x, y):
... return x + y
You can register directly from a function:
>>> @jm.register("nice")
... def _(x: float, y:float) -> nt.float64:
... pass
or again taking the function name:
>>> @jm.register
... def nice(x: float, y:float) -> nt.float64:
... pass
Quick Installation
To install numbakit-anjit, simply (soon):
$ pip install numbakit-anjit
or utilizing conda, with the conda-forge channel (soon):
$ conda install -c conda-forge numbakit-anjit
and then simply enjoy it!
Why
Numba njit is awesome. Simple to use, produces the appropriate machine code once that the function is called. As the Numba docs says:
in [Lazy mode], compilation will be deferred until the first function execution. Numba will infer the argument types at call time, and generate optimized code based on this information. Numba will also be able to compile separate specializations depending on the input types.
But numba also has an eager mode:
In which you can also tell Numba the function signature you are expecting. [..] In this case, the corresponding specialization will be compiled by the decorator, and no other specialization will be allowed. This is useful if you want fine-grained control over types chosen by the compiler (for example, to use single-precision floats).
This can produce slightly faster code as the compiler does not need to infer the types. It also provides type check at definition time ensuring correctness. In numba intensive projects, this can be an useful trait. Finally, eager compilation is currently required to have two functions with the same signature to be arguments of a third one, without needing to recompile this last one in each case.
Another think we like about njit is that is highly configurable using keyword arguments and even some configurations could be applied globally using env variables.
While developing numbakit-ode I was missing two things:
That eager compilation make use of function annotations
A way to manipulate njit options in a centralized but granular manner
So, numbakit-anjit was born.
numbakit-anjit is maintained by a community. See AUTHORS for a complete list.
To review an ordered list of notable changes for each version of a project, see CHANGES
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file numbakit-anjit-0.2.tar.gz
.
File metadata
- Download URL: numbakit-anjit-0.2.tar.gz
- Upload date:
- Size: 14.2 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/1.12.1 pkginfo/1.5.0.1 requests/2.23.0 setuptools/46.1.1.post20200323 requests-toolbelt/0.8.0 tqdm/4.43.0 CPython/3.6.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | ed0b392bc986f267d21ecbf88d313f233c257361a974d0578faa3040cf60a311 |
|
MD5 | 5931080623e2929f22489f902ab9b5cb |
|
BLAKE2b-256 | b8a81f9fb02bea0aedfdf1aeacf3731a5e3c9c5510912a0adcf7e4afffb42ebd |
File details
Details for the file numbakit_anjit-0.2-py2.py3-none-any.whl
.
File metadata
- Download URL: numbakit_anjit-0.2-py2.py3-none-any.whl
- Upload date:
- Size: 12.8 kB
- Tags: Python 2, Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/1.12.1 pkginfo/1.5.0.1 requests/2.23.0 setuptools/46.1.1.post20200323 requests-toolbelt/0.8.0 tqdm/4.43.0 CPython/3.6.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 3f59995d4949ea7f5f8d9d99aacb4dfb8b106f4cec1a6b5754f0272adf256681 |
|
MD5 | 2ee40eda6da8f74fc8763b11b56fbcdd |
|
BLAKE2b-256 | a26afee624312a405ee63061bb67a40c47145f14d416f270e8a59c40710353bb |