Intel Speeds Up ISPC Via LLVM/Clang LTO

[phoronix]

Phoronix: Intel Speeds Up ISPC Via LLVM/Clang LTO

Intel has released a new version of ISPC, their Implicit SPMD Program Compiler. The ISPC 1.22 release is clocking in a few percent faster across the board as with their release builds they are now making use of LLVM Clang's link-time optimizations (LTO) for speeding up the binaries...

https://www.phoronix.com/news/Intel-ISPC-1.22

[ms178]

PGO is next, please?! See https://github.com/llvm/llvm-project/pull/71067 for the gains of the Windows release binary.

[zamazan4ik]

PGO is next, please?! See https://github.com/llvm/llvm-project/pull/71067 for the gains of the Windows release binary.

Unfortunately, right now for some reasons ISPC dev team has no plans for PGO: https://github.com/ispc/ispc/issues/2687

[ssokolow]

I'm always reminded of Matt Pharr's The story of ispc: origins (part 1), where he quoted Tim Foley's "auto-vectorization is not a programming model" when talking about reasons Larrabee failed.

[User42]

Agreed. That's why an approach like openmp makes sense (if only openmp had not grown organically, after the fact and only for 2 languages).

Languages ought to provide primitives to express parallelism and dependence. If vectorization can be inferred, good. But expressing dependencies explicitly is always better (if only for the compiler to yell at you if you break the contract instead of breaking the auto-vectorization randomly).

[ms178]

I'm always reminded of Matt Pharr's The story of ispc: origins (part 1), where he quoted Tim Foley's "auto-vectorization is not a programming model" when talking about reasons Larrabee failed.

Yeah, his blog series is legendary for all the insights behind ISPC and the related hardware features. And it is quite sad to see that AVX-512 became relevant only these days on consumer platforms.