There is no defensible reason why SIMD code should be harder to write than scalar code. GLSL has been around since 2004 and in GLSL it’s easier to write vector code than scalar code.
-
-
You may want to look at AVX512 which adds a lot of major functionality to "do vectorisation right".
-
And as others have said the best way to "look at AVX512" is to do so via ISPC rather than intrinsics. Writing ISPC->AVX512 is the HLSL/GLSL experience you're asking for.
- 4 more replies
New conversation -
-
-
It depends how you layout your data. Speed ups of 6-7x over scalar code are likely with 256bit SIMD.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I disagree with this. Intel's 'move' to 256-bit had a bunch of headwinds for many users: the fact that you'd still need a 128-bit path for Atom and older chips, the initial clock speed penalties for using AVX2 and most of all, the fact that AVX2 was really 2x128b not 256b.
-
Things like VPALIGNR and VPSHUFB really clearly illustrate the latter point. They also are exemplars of why vector coding is not as easy as scalar coding - a lot of vector coding is not simply 'doing stuff in-lane with an occasional masking-out', but using the vector ...
- 2 more replies
New conversation -
-
This Tweet is unavailable.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.