[macstl-dev] Comments, help for macstlizer

[Glen Low]

Hi All,

Folks, I'd like some comments and help on macstlizer, which is a  
feature of the soon-to-be-released macstl 0.3 -- I'm currently in the  
midst of documenting it, it should be out on Monday.

macstlizer is a sed script and C++ header file designed to ease an  
Altivec programmer's transition to SSE. At the end of the day, your  
single SIMD source should compile and run in both Intel and PowerPC  
worlds. This is not a panacea like Rosetta claims to be -- I of all  
people know there are Altivec opcodes that simply have no good  
equivalent SSE equivalents -- rather, it's a plan that maintains your  
source code quality and compilability on PowerPC, and uses the  
compiler to guide your transition to the brave new world of Mactel.

1.    The sed script, altivec2macstl.sed, translates Altivec C code  
into the equivalent macstl C++ code. After running the script and  
some code cleanup, your translated source should still compile and  
run correctly in PowerPC.
2.    The header altivec2mmx.h provides inline versions of some  
Altivec intrinsics as SSE. Compiling the code on Intel should  
highlight only the opcodes that need to be ported, and the templated  
inline functions assure you of full performance otherwise. This  
should allow you to focus on the smaller part of your Altivec code  
that needs work.
3.    When you've successfully made the final code cross-platform,  
you can then run Shark etc. on the result to see if you need to  
revisit any of the inline functions.

I don't think a pure C inline function and/or macro-based solution  
will work, because:
1.    Altivec has a stronger type system than SSE. Converting all  
your __vector ints to __m128i, you then have to be careful to always  
treat it as a vector of 4 ints and not a vector of 8 shorts for example.
2.    Altivec has a distinct vector initialization syntax. This will  
bite especially for Codewarrior users who have to use the Motorola- 
style ( ) syntax, whereas gcc users can still use { } on both PowerPC  
and Intel and get the same syntax -- although the weak type system in  
SSE will hit you again, since you will have to cast back to the  
public type of __m128i to use the intrinsics at all.
3.    Altivec has overloaded intrinsics, SSE doesn't. Either you have  
to manually and carefully translate each vec_add into a _mm_add_ps or  
_mm_add_epi16, or you have to use a language with overloading like C++.
4.    Both Altivec and SSE intrinsics have literal (constant)  
parameters e.g. vec_sld. These cannot be modelled using C (__inline)  
functions at all, since the parameter must be a compiler constant.  
Using C macros has it own dangers e.g. repeated evaluation of macro  
parameters.

The macstlizer solution addresses all of these points -- comments,  
suggestions are welcome.

=== HELP WANTED ===

To test out the sed script and general scheme, I need suitable chunks  
of Altivec C source code to test against. Please email me off-list  
with a .c file with Altivec codes in it, preferably something in  
production already, don't worry I'm not about to steal your code :-).

Cheers, Glen Low


---
pixelglow software | simply brilliant stuff
www.pixelglow.com
aim: pixglen

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.pixelglow.com/lists/archive/macstl-dev/attachments/20050623/d56e1a4d/attachment.html