DirectXMath: F16C and FMA

re: DirectXMath: F16C and FMA

Chuck Walbourn - MSFT — Thu, 15 Nov 2012 08:15:03 GMT

Any use of intrinsics assumes the programmer knows what they are doing. It's basically like inline assembly.

re: DirectXMath: F16C and FMA

Alecazam — Wed, 14 Nov 2012 23:26:18 GMT

What is the recommended safe method of working with intrin.h? This can cause a crash if I don't have the appropriate SSE processor for an intrinsic, but they're all defined by that header (AVX, etc). In VS2012, intrin.h is also included by <string> now since it wants the InterlockedIncrement/Decrement calls in there. I much preferred explicitly including the SSE headers, but I see no way to mask out those in intrin.h.

re: DirectXMath: F16C and FMA

Chuck Walbourn - MSFT — Mon, 17 Sep 2012 05:49:05 GMT

Many of them are Intel codenames. 'pmmintrin.h' which houses SSE3 was originally codenamed "Prescott New Instructions (PNI)", hence the 'p'. As such the early ones tend to be a little unintuitive.

All the more recent stuff (AVX, F16C, FMA3) is in 'immintrin.h' which is Intel or (FMA4, XOP, etc.) in 'ammintrin.h' which is AMD, although that's of course not indicating which ones are vendor-specific and which have been adopted by the other vendor.

In theory most intrinsics should all be in intrin.h.

Header	Description
`intrin.h`	General intrinsics, notably `__cpuid` and various intrinsics forms of the CRT routines
`ammintrin.h`	SSE5, FMA4, and XOP instrinsics
`xmmintrin.h`	SSE intrinsics and the `__m128` type (single-precision float SIMD)
`emmintrin.h`	SSE2 intrinsics and the `__m128i/__m128d` types (double-precision float and integer SIMD)
`pmmintrin.h`	SSE3 intrinsics (horizontal adds and subtracts float/double operations, specific ‘dup’ operations)
`tmmintrin.h`	SSSE3 intrinsics (more horizontal ops, integer abs, ‘byte’ shuffle to augment SSE2)
`smmintrin.h`	SSE4.1 intrinsics (dot-product, rounding, augmented min/max support for SSE2)
`nmmintrin.h`	SSE4.2 intrinsics
`immintrin.h`	AVX, FMA3, F16C/CVT16, and AVX2 intrinsics
`wmmintrin.h`	AES intrinsics

re: DirectXMath: F16C and FMA

clayman — Mon, 17 Sep 2012 02:38:52 GMT

Nice series !

re: DirectXMath: F16C and FMA

barbie — Sat, 15 Sep 2012 07:25:57 GMT

Thanks for this series Chuck. I was wondering, do you know if there is any method to the names of the intrinsic headers? They seem to be using a bit of a random distribution, making remembering them ... hard!