Psyco, the Python Specializing Compiler

Hello Tim,

On Mon, Jun 21, 2004 at 03:07:14PM -0700, Tim Hochberg wrote:
> Now, being a glutton for punishment, I'm wondering if there is a way to 
> virtualize access to the attributes of a complex number.  In principal, 
> it seems should be possible to implement z.real and z.imag faster than 
> going through all the python machinery, but I have no idea how to spell 
> that. I'd also be interested in figuring out how to speed z.conjugate(), 
> but that's a lower priority.

Psyco virtualizes the whole PyObject_GenericGetAttr(), including looking up
the 'real' and 'imag' properties of the complex class.  It then calls the
'tp_descr_get' of these properties.  This is 2.2 behavior:

>>> c = type(2j)
>>> c.real
<member 'real' of 'complex' objects>

Accelerating this is not obvious.  It involves first getting used to what this
strange object is (PyMemberDescr_Type in descrobject.c), and writing a virtual
implementation for its tp_descr_get (i.e. member_get()).  This virtual
implementation should have a way to associate meta-implementations with
individual members (a tp_meta_members table)...

Unfortunately, c.conjugate() is much easier to implement.  See e.g. how
list.append() is implemented.


A bientot,

Armin.

Hello Armin,

Armin Rigo wrote:

>Hello Tim,
>
>On Mon, Jun 21, 2004 at 03:07:14PM -0700, Tim Hochberg wrote:
>  
>
>>Now, being a glutton for punishment, I'm wondering if there is a way to 
>>virtualize access to the attributes of a complex number.  In principal, 
>>it seems should be possible to implement z.real and z.imag faster than 
>>going through all the python machinery, but I have no idea how to spell 
>>that. I'd also be interested in figuring out how to speed z.conjugate(), 
>>but that's a lower priority.
>>    
>>
>
>Psyco virtualizes the whole PyObject_GenericGetAttr(), including looking up
>the 'real' and 'imag' properties of the complex class.  It then calls the
>'tp_descr_get' of these properties.  This is 2.2 behavior:
>
>  
>
>>>>c = type(2j)
>>>>c.real
>>>>        
>>>>
><member 'real' of 'complex' objects>
>
>Accelerating this is not obvious.  It involves first getting used to what this
>strange object is (PyMemberDescr_Type in descrobject.c), and writing a virtual
>implementation for its tp_descr_get (i.e. member_get()).  This virtual
>implementation should have a way to associate meta-implementations with
>individual members (a tp_meta_members table)...
>  
>
Brrrr! Ok, I think I'll forgo attempting to do anything about this for now.


>Unfortunately, c.conjugate() is much easier to implement.  See e.g. how
>list.append() is implemented.
>  
>
Where would I find this? I couldn't find anything obvious in 
plistobject.[ch]. Is it lurking in one of the other subdirectories?

Thanks,

-tim


>
>A bientot,
>
>Armin.
>
>
>
>  
>

Hello Tim,

Hum, I appear to have forgotten a few things about Psyco...

On Tue, Jun 22, 2004 at 11:52:35AM -0700, Tim Hochberg wrote:
> >Accelerating this is not obvious.  It involves first getting used to w=
hat=20
> >this
> >strange object is (PyMemberDescr_Type in descrobject.c), and writing a=
=20
> >virtual
> >implementation for its tp_descr_get (i.e. member_get()).  This virtual
> >implementation should have a way to associate meta-implementations wit=
h
> >individual members (a tp_meta_members table)...
> >=20
> >
> Brrrr! Ok, I think I'll forgo attempting to do anything about this for =
now.

Actually, there is a pdescrobject.c doing some of it already.  It calls
PsycoMember_GetOne() in pstructmember.c.  This means that for the 'real' =
and
'imag' members of complex objects, they are read directly from the
PyComplexObject.  This is quite good, expect that I guess you're implemen=
ting
virtual complex objects, like pfloatobject.c has virtual float objects; t=
hen
this will force the PyComplexObject to be created whenever 'real' or 'ima=
g' is
read.

Well, I guess I have to think a bit more about it -- and also about the
conjugate() method:

> >Unfortunately, c.conjugate() is much easier to implement.  See e.g. ho=
w
> >list.append() is implemented.

That's a lie.  list.append() is not implemented in Psyco -- it doesn't re=
ally
need to be, because the built-in list_append() is about as fast as it can=
 be. =20
Actually, I now think that no built-in method whatsoever is implemented b=
y
Psyco, only built-in operations.  I'd have to think a bit about how to do=
 that
too.


A bient=F4t,

Armin