SIMD detection doesn't work on MSVC x64 targets

Question

SIMD detection doesn't work on MSVC x64 targets

cdwfs opened this issue 7 years ago · comments

The checks in mathfu/utilities.h to enable MATHFU_COMPILE_WITH_SIMD don't cover Visual Studio x64 targets. The code currently checks the existence (and value) of _M_IX86_FP, which is only defined for x86 (32-bit) targets. SSE and SSE2 support is implicit on x64 targets, which can be handled with the following addition (or similar):

#elif defined(_M_IX86_FP)
#if _M_IX86_FP >= 1        // SSE enabled
#define MATHFU_COMPILE_WITH_SIMD
#endif  // _M_IX86_FP >= 1
+ #elif (defined(_M_AMD64) || defined(_M_X64))
+ #define MATHFU_COMPILE_WITH_SIMD // MSVC targeting X64 implies SSE+SSE2
#endif
#endif  // !defined(MATHFU_COMPILE_WITHOUT_SIMD_SUPPORT)

Stewart Miles · Answer 1 · Thu Aug 17 2017 02:26:18 GMT+0800 (China Standard Time)

Wanna send us a pull request?

Cort · Answer 2 · Thu Aug 17 2017 02:28:05 GMT+0800 (China Standard Time)

Sure, will do

Loïc Baumann · Answer 3 · Sun Jan 14 2018 03:24:24 GMT+0800 (China Standard Time)

@cdwfs do you see performance improvement while using SIMD on MSVC target?
For me the Vector Benchmark is 50% slower using SIMD...

Cort · Answer 4 · Thu Jan 18 2018 05:28:20 GMT+0800 (China Standard Time)

I see the same results: a ~50% slowdown when enabling SIMD. See the comments on my pull request #26.

Loïc Baumann · Answer 5 · Thu Jan 18 2018 05:32:18 GMT+0800 (China Standard Time)

Just did, thank you. Indeed x64 is irrelevant to the issue you mentioned, I was running my bench in x86.
For me the way the lib was designed and coded make its hard to really benefit of SSE2 Instruction as there're still conversion, copy constructors, temp objects...