- The application needs to inform a new flag when registering types that require 16byte alignment, e.g. asOBJ_APP_ALIGN16 (asCScriptEngine::RegisterObjectType)
I didn't add a flag for 16-byte alignment (since we'll need more than just 16 byte aligned variables soon), I added a new optional parameter to RegisterObjectType:
https://bitbucket.org/sherief/angelscript/commits/a6dd89ef276d09825c15a929dac6c6d2055d0cd6
- I didn't add a flag for 16-byte alignment (since we'll need more than just 16 byte aligned variables soon), I added a new optional parameter to RegisterObjectType:
What other alignment requirements are you in need of?
I'll review all above patches as soon as I can and provide my feedback.
What other alignment requirements are you in need of?
32, for AVX. For future compatibility I want the changes to be able to support any alignment.
- The engine must make sure to always allocate the memory for the script objects on 16byte aligned boundaries (asCScriptEngine::CallAlloc)
CallAlloc() relies on useralloc - should the signature be changed to add an alignment parameter? I can do this, and I'd also add a simple wrapper for the default malloc() / free() use case.
I don't think AVX will require 32 byte alignment. Do you have any reference that says it will? In fact the wikipedia entry on avx says the memory alignment requirement of SIMD instructions may be relaxed, so it is quite possible that not even 16byte alignment will be needed. (unfortunately no reference for this was given).
But, feel free to make the code generic to support any alignment requirement.
As for CallAlloc(). Not all memory allocations needs to be aligned and you definitely do not want to force alignment of small allocations as it may waste a lot of memory. I think you need to add a secondary method, e.g. CallAllocAligned(). This new method can optionally take an argument with the desired alignment (though, for performance it may be better just to hardcode it to 16). The code that allocates memory must know if the memory needs to be aligned or not.
vmovaps requires 32-byte alignment. We also see a performance benefit from aligning matrices on 32 or 64 byte boundaries due to cache behavior.
For a reference see Intel® Advanced Vector Extensions Programming Reference:
In http://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-optimization-manual.pdf, we can read in section 3.6.4 that:
Misaligned data access can incur significant performance penalties. This is particularly true for cache line
splits. The size of a cache line is 64 bytes in the Pentium 4 and other recent Intel processors, including
processors based on Intel Core microarchitecture.
An access to data unaligned on 64-byte boundary leads to two memory accesses and requires several
µops to be executed (instead of one). Accesses that span 64-byte boundaries are likely to incur a large
performance penalty, the cost of each stall generally are greater on machines with longer pipelines.
:)
The changes you've done so far seems to be fine.
Are you testing against the test_feature app in the svn? It's the best way to make sure you don't accidentally break anything.
As for CallAlloc(). Not all memory allocations needs to be aligned and you definitely do not want to force alignment of small allocations as it may waste a lot of memory. I think you need to add a secondary method, e.g. CallAllocAligned(). This new method can optionally take an argument with the desired alignment (though, for performance it may be better just to hardcode it to 16). The code that allocates memory must know if the memory needs to be aligned or not.
Some types do need natural alignment though, and the default allocator (that wraps malloc()) is already aligning to the largest supported type alignment on the platform, since malloc() is guaranteed to do that. On OS X, my primary platform, malloc() is guaranteed to return 16 byte aligned data. On almost every platform with doubles aligned is required by the spec to return a pointer that is at least as aligned as a double (8 bytes).
Also, the patches add an alignment type to the type id info that defaults to 4, and that's what'll get passed to allocators when allocating memory for a type - so in existing code the behavior of the changes should be exactly identical.
