fixing the broken build and optimizing the engine by 2.3 sec by using unused avx function namely computeBoundsOverlapMatrixAVX

[MajidAbdelilah]

as simple as that

[VonTum]

So, the thing is, SIMDHelper is used as a kind of grouping type for methods that can be well optimized with SIMD. It is meant to be filled out by the template using the function. This way we could use dynamic dispatch for the SIMD implementation on a higher level, reducing selection overhead. Similar to how Polyhedron's SSE and AVX variants are structured.

TrunkSIMDHelperFallback for this is one implementation for SIMDHelper that should not use SIMD. So we can use it on all platforms. Currently as it stands, you made the default code path require AVX, breaking the code for SSE-only systems.

The missing header is good though, It seems the CI build doesn't require it, but other systems may.

[MajidAbdelilah]

but the SIMDHelper didnt select the avx function for my system which is confermed with linux perf it was using the scalar function

[VonTum]

Yes, dynamic dispatch at the tree's root node was never finished.

Perhaps for the time being it would be easiest to just add an 'if' in computeBoundsOverlapMatrix that checks CPUID for AVX, and executes computeBoundsOverlapMatrixAVX instead if that's the case. It would be a minor overhead for a easy implementation.

[MajidAbdelilah]

and how to do this CPUID check

[VonTum]

See Physics3D/misc/cpuid.h

if hasTechnology(AVX) {...}

[MajidAbdelilah]

thanks

[MajidAbdelilah]

hey are you working today