Commit 95449c41 authored by Pierre Augier's avatar Pierre Augier
Browse files

List some Cython issues

parent d75a2c0a057d
......@@ -31,4 +31,6 @@ doc/ipynb/*.rst
**/__pythran__/*
**/__cython__/*
**/__numba__/*
**/__python__/*
\ No newline at end of file
**/__python__/*
doc/for_dev/**/*.c
......@@ -44,6 +44,12 @@ Less bugs and more Cython features...
- "exterior" functions + understand the tests
- void type
- nogil for function definition
- nogil context manager
setup.py & more than one backend at runtime
-------------------------------------------
......
......@@ -7,36 +7,33 @@ However, a descent set of Cython can be supported. We need to find Python
syntaxes for the most useful Cython special syntaxes.
Note that some Cython features are useless in Pythran (for example cdef of
locals variable, or `nogil`).
local variables, or `nogil`).
Note that ideally, we want to write Cython code that can be executed without
the Cython package and without Cython compilation. It is possible with a subset
of the "pure Python mode" of Cython for which the Python file does not depend
on Cython, i.e. all the supplementary information needed by Cython has to be
written in the .pxd file.
the Cython package and without Cython compilation. It is possible with the
["pure Python mode" of
Cython](https://cython.readthedocs.io/en/latest/src/tutorial/pure.html).
Therefore, we first need examples of Cython code written in this mode.
## Cython syntaxes already supported
Note however, that this mode is currently still experimental and that we hit
simple Cython bugs which limit a lot what can be done in practice with the
Cython backend. For example:
### Simple fused types, memory views
- Pure-Python mode and fused types <https://github.com/cython/cython/issues/3142>
- `cython.locals(arr=np.ndarray[...])` https://github.com/cython/cython/issues/3129
We already have fused types in Transonic. With Transonic, we can already do (no
notion of C order):
More generally, there are many known bugs in Cython which do not help! For example:
```python
import numpy as np
from transonic import Transonic, Type, NDim, Array
- `ctypedef` and buffer <https://github.com/cython/cython/issues/754>
- <https://stackoverflow.com/questions/57887972/defining-a-fused-type-using-a-fused-type>
np_floats = Type(np.float32, np.float64)
N = NDim(2, 3, 4)
A = Array[np_floats, N]
A1 = Array[np_floats, N + 1]
# or simply
A3d = Array[np_floats, "3d"]
```
I think at least some of these bugs have to be solved upstream...
## Cython syntaxes already supported
### `cpdef` signature with simple (basic and array) types for arguments
### cdef for type declaration of local variables: `cython.locals`
### `cdef` for type declaration of local variables: `cython.locals`
In "pure Python mode", one can write
......@@ -45,10 +42,9 @@ In "pure Python mode", one can write
cpdef mysum(np.float64_t[:] arr_input)
```
With variables annotations (which have to be removed for Pythran / Numba):
With variables annotations (which are removed for Pythran / Numba):
```python
from transonic import boost
@boost
......@@ -59,7 +55,6 @@ def mysum(arr_input: "float[]"):
for i in range(n):
result += arr_input[i]
return result
```
### Cython decorators
......@@ -71,7 +66,7 @@ def mysum(arr_input: "float[]"):
Currently only for simple functions (no methods).
## Cython syntaxes nearly supported
## Cython syntaxes partly supported
### Function definition (cdef, cpdef, inline, nogil, return type)
......@@ -86,45 +81,141 @@ def func(a: "float[]", n: int) -> "void":
which would translate in Cython as something like:
```cython
cdef inline void func(np.float_t[:] a, cython.int n) nogil
cpdef inline void func(np.ndarray[np.float_t, ndim=1] a, cython.int n) nogil
```
`boost(inline=True)` is supported for functions, see [this
example](https://transonic.readthedocs.io/en/latest/examples/inlined/txt.html).
- all function signatures use `cpdef`
## Cython syntaxes that can be supported quite easily
- `boost(inline=True)` is supported for functions, see [this
example](https://transonic.readthedocs.io/en/latest/examples/inlined/txt.html).
- Return type is supported but there is no void type.
### Fused types
We already have fused types in Transonic. With Transonic, we can already do:
```python
import numpy as np
from transonic import Array, Type, NDim
np_floats = Type(np.float32, np.float64)
N = NDim(2, 3, 4)
A = Array[np_floats, N]
A1 = Array[np_floats, N + 1]
# or simply
A3d = Array[np_floats, "3d"]
```
### ctypedef, fused types and contiguous arrays
However, Cython Fused types are currently very limited.
Even with something as simple as that
```python
from transonic import Array, Type
A = Array[Type(np.float64, np.complex128), "1d"]
def mysum(arr: A):
result: A.dtype = arr.dtype.type(0.)
i: int
for i in range(arr.shape[0]):
result += arr[i]
return result
```
should be translated to this (not supported, see
<https://github.com/cython/cython/issues/754>) Cython code:
```cython
cimport numpy as cnp
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
ctypedef np.ndarray[T0, ndim=1] A
def mysum(A arr):
cdef T0 ret = arr.dtype.type(0.)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
```
Note that it works with a memoryview... (but not in pure-Python mode!)
ctypedef fused np_floats:
cnp.float32_t
cnp.float64_t
Note that another working alternative is:
cdef np_floats[:, :, ::1] myvar
```
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
def mysum(np.ndarray[T0, ndim=1] arr):
cdef T0 ret = arr.dtype.type(0.)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
```
A nice Python syntax has to be found for `np_floats[:, :, ::1]`
But the corresponding pure-Python version does not work!
## Cython syntaxes that can be supported quite easily
### More array types (contiguous arrays, C or F order, memoryviews)
It could just be
I think we should support:
```python
A2 = Array[np_floats, "[:, :, ::1]"]
Array[int, NDim(3), "order=C"]
Array[int, "3d", "order=C"]
Array["int[:, :, ::1]"]
Array[int, "[:, :, ::1]"]
transonic.str2type("int[:, :, ::1]")
transonic.typeof(np.empty((2, 2, 2)))
```
and maybe also:
```python
transonic.int64[:, :, ::1]
```
I tend to think that the default (`"int[:,:]"`) should correspond to
`"order=C"`. "Fortran" order and "any" order (contiguous C or F) could be
obtained with `"order=F"` and `"order=any"`.
Strided arrays could be obtained with `Array[int, NDim(3), "strided"]` or
`str2type("int[::, ::, ::]")`.
Note that for Pythran, we could also support:
```python
A_fixed_dim = Array[np_floats, "[:, :, 3]"]
A_fixed_dim = Array[Type(np.float32, float), "[:, :, 3]"]
```
For Cython, we need to be able to specify if an array is a `np.ndarray` or a
`memoryview`. By default, we will use `np.ndarray` and `memoryview` could be
obtained with:
```python
Array[int, "[:, :, ::1]", "memview"]
```
### Special C types
For example `Py_ssize_t`
For example `Py_ssize_t` and `void`
```python
from ctypes import c_ssize_t as Py_ssize_t
......@@ -176,7 +267,7 @@ to suppress the `with nogil()`.
cdef Py_ssize_t *p_indexer
```
### Definition `struc`, `enum`, `class`
### Definition `struct`, `enum`, `class`
```cython
cdef struct Heap:
......
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
ctypedef np.ndarray[T0, ndim=1] A
def mysum(A arr):
cdef T0 ret = arr.dtype.type(0)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
\ No newline at end of file
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
ctypedef T0[:] A
def mysum(A arr):
cdef T0 ret = arr.dtype.type(0)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
\ No newline at end of file
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
@cython.locals(ret=T0, i=cython.int)
cpdef mysum(np.ndarray[T0, ndim=1] arr)
def mysum(arr):
ret = arr.dtype.type(0)
for i in range(arr.shape[0]):
ret += arr[i]
return ret
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
ctypedef T0[:] A
@cython.locals(ret=A, i=cython.int)
cpdef mysum(A arr)
def mysum(arr):
ret = arr.dtype.type(0)
for i in range(arr.shape[0]):
ret += arr[i]
return ret
\ No newline at end of file
import cython
import numpy as np
cimport numpy as np
# not supported, see: https://github.com/cython/cython/issues/754
ctypedef np.ndarray[np.float64_t, ndim=1] A
def mysum(A arr):
cdef np.float64_t ret = arr.dtype.type(0)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
\ No newline at end of file
import cython
import numpy as np
cimport numpy as np
ctypedef fused T0:
np.complex128_t
np.float64_t
def mysum(np.ndarray[T0, ndim=1] arr):
cdef T0 ret = arr.dtype.type(0)
cdef cython.int i
for i in range(arr.shape[0]):
ret += arr[i]
return ret
\ No newline at end of file
import cython
import numpy as np
cimport numpy as np
ctypedef fused A:
np.int_t[:]
np.float_t[:]
@cython.locals(arr=A)
cpdef func(A arg)
import numpy as np
def func(arg):
arr = np.empty_like(arg)
return arr
import cython
import numpy as np
cimport numpy as np
ctypedef fused A:
np.int_t[:]
np.float_t[:]
cpdef func(A arg):
cdef A arr = np.empty_like(arg)
return arr
......@@ -112,18 +112,31 @@ def f(x, y):
## Transonic
To define multisignatures from fused types, we can do:
To define multi-signatures from fused types, we can do:
```python
from transonic import Array, Type, NDim, boost
A = Array[Type(int, float), NDim(1, 2), "memview"]
A = Array[Type(int, float), NDim(1, 2)]
@boost
def func(a: A, b: A):
pass
```
Note that this other code should give a different result in term of signatures:
```python
from transonic import Array, Type, NDim, boost
A0 = Array[Type(int, float), NDim(1, 2)]
A1 = Array[Type(int, float), NDim(1, 2)]
@boost
def func(a: A0, b: A1):
pass
```
## Issue about fused types
This works with Pythran and it should also work with other backends.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment