This document describes features we want in numba, but do not have yet. We will first list what we want in upcoming versions, and then what features we want in general. Those features can always be added to the roadmap for upcoming versions if someone is interested in implementing them.
What we want for 1.0 is:
We also like some minimal Cython support, in addition to the longer term goals of SEP 200. One idea from Zaur Shibzukhov is to provide support for Cython pxd overlays:
# foo.py
def my_function(a):
b = 2
return a ** b
Such a module can be overlain with a Cython pxd file, e.g.
# foo.pxd
cimport cython
@cython.locals(b=double)
cpdef my_function(double a)
For some inspiration of what we can do with pxd overlays, see also: https://github.com/cython/cython/blob/master/Cython/Compiler/FlowControl.pxd
We can now compile foo.py with Cython. We should be able to similarly compile foo.py with numba, using pycc as well as at runtime to produce a new module with annotated functions compiled in the right order.
We will order these from less involved to more involved, to provide different entry points to numba development.
Here as some less intricate topics, providing easier starting points for new contributors:
We currently have (naive implementations of):
But we want many more! Some ideas:
typeddict
typedset
typedchannel
- one thread-safe (nogil) and one requiring the GIL
Perhaps also the ordered variants of typeddict and typedset.
Support for LLVM intrinsics (we only have instructions at the moment):
E.g.:
intrin = numba.declare_intrinsic(int64(), "llvm.readcyclecounter")
print intrin()
Analogous to cython --annotate, a tool that annotates numba source code and finds and highlights which parts contain object calls. Ideally, this would also include, for each source line (expand on click?):
The final (unoptimized) LLVM bitcode
- And optionally the optimized code and/or assembly
Code from intermediate numba representations
- After we start implementing several layers of IR, see http://numba.pydata.org/numba-doc/dev/doc/ir.html
The type of each sub-expression and variable (on hover?)
Allow two forms of code caching:
- For distribution (portable IR)
- Locally on disk (unportable compiled binaries)
The first bullet will allow library writers to distribute numba code while not being tied to numba versions that users have installed. This would be similar to distribution of C code compiled from Cython source:
$ numba --compile foo.py
Writing foo.numba
We can now distribute foo.numba. Load code explicitly:
from numba import loader
foo = loader.load("foo.numba")
foo.func()
... or use an import hook:
from numba import loader
loader.install_hook()
import foo
foo.func()
... or compile to extension modules during setup:
from numba.loader import NumbaExtension
setup(
...,
ext_modules=[
NumbaExtension("foo.bar",
sources=["foo/bar.numba"]),
],
)
Or perhaps more conveniently, implement find_numba_modules() to find all *.numba source files and return a list of NumbaExtension.
This also plays into the IR discussion found here: http://numba.pydata.org/numba-doc/dev/doc/ir.html
Jit operations that result in calls to special methods like __len__, __getitem__, etc. This requires some careful thought as to the stage where this transformation should take place.
Array Expression support in Numba, including scans, reductions, etc. Or maybe we should make Blaze a hard dependency for that?
More intricate topics, in no particular order:
@Any(int_, Any)
def my_method(self, a, b):
...
Infer the return type and specialize on parameter type b, but fix parameter type a.
What we also need is native dispatch of foreign callables, in a sustainable way: SEP 200 and SEP 201
- https://github.com/numfocus/sep/
- Widen support in scientific community
Support recursion for autojit functions and methods:
Construct call graph
Build condensation graph and resolve
- similar to cycles in SSA
Support for zero-cost exceptions: support in the runtime libraries for all models:
True zero-cost exceptions
- Stack trace through libunwind/apple backtrace/LLVM info based on instruction pointer
- http://llvm.org/docs/LangRef.html#invoke-instruction
- http://llvm.org/docs/ExceptionHandling.html
Setjmp/longjmp
- Optionally with exception analysis to allow cheap cleanup for the simpler cases
Costful exceptions
- “return -1”
- Implement fast NumbaErr_Occurred() or change calling convention for native or void returns
We also need to allow users to take the pointer to a numba jit function:
numba.addressof(my_numba_function)
We can allow specifying an exception model:
propagate=False: This does not propagate, but uses PyErr_WriteUnraisable
propagate=True: Implies write_unraisable=False. Callers check with NumbaErr_Occurred() (or for NULL if object return). Maybe also specify a range of badvals:
- int -> 0xdeadbeef (ret == 0xdeadbeef && NumbaErr_Occurred())
- float -> float(‘nan’) (ret != ret && NumbaErr_Occurred())
Note
We have numba.addressof(), but we don’t have NumbaErr_Occurred() yet.
Use cheap heap allocated objects + garbage collection?
- or atomic reference counts?
Use stack-allocation + escape analysis?
Blaze support:
- compile abstract blaze expressions into kernels
- generate native call to blaze kernel
Support for generators based on green threading support:
- Write typed channels as autojit class
- Support green thread context switching
- Rewrite iteration over generators
def g(N):
for i in range(N):
yield f(i) # write to channel (triggering a context switch)
def consume():
gen = g(100) # create task with bound parameter N and channel C
for i in gen: # read from C until exhaustion
use(i)
See also https://groups.google.com/a/continuum.io/forum/#!searchin/numba-users/generators/numba-users/gaVgArRrXqw/HTyTzaXsW_EJ for how this compares to generators based on closures.
We support Python 3.3, but we can additionally support type-annotations:
def func(a: int_, b: float_) -> double:
...
Maybe this can work with numba.automodule(my_numba_module) as well as with jit and autojit methods.
- SPIR support (OpenCL)
Vector-types in Numba
- What does this look like?