It's Coming Down To The Wire For New AMD Zen Scheduler Model In LLVM 5.0
Tomorrow is the feature freeze / branching deadline for LLVM 5.0 and as of writing the new AMD Zen scheduler model hasn't been merged, but it looks like it could still happen in time.
Going back to March we've been looking forward to a new AMD Zen scheduler model for LLVM, as explained in LLVM 5.0 Should Offer Better AMD Ryzen Performance. The "znver1" (Zen 1st gen) code in LLVM has been relying upon the AMD Bobcat scheduler model up until now while the work is pending for introducing the proper Zen scheduler model. This has the potential to benefit the performance of resulting binaries targeting the znver1 architecture. Even right now if using the Haswell scheduler model, in some workloads can be a 5~10% boost of the compiled binaries.
We've been waiting for AMD to publish this new znver1 scheduler model and last week they finally did. Over the past few days, that new scheduler model has been under review. As of this morning, the AMD developer who posted the patch for review is requesting an LLVM developer with commit rights to commit the patch. Now hopefully that will happen prior to the LLVM 5.0 branching tomorrow.
The official release of LLVM 5.0.0 meanwhile is expected on 23 August following a few release candidates while LLVM 6.0.0 will be the next follow-on release.
Going back to March we've been looking forward to a new AMD Zen scheduler model for LLVM, as explained in LLVM 5.0 Should Offer Better AMD Ryzen Performance. The "znver1" (Zen 1st gen) code in LLVM has been relying upon the AMD Bobcat scheduler model up until now while the work is pending for introducing the proper Zen scheduler model. This has the potential to benefit the performance of resulting binaries targeting the znver1 architecture. Even right now if using the Haswell scheduler model, in some workloads can be a 5~10% boost of the compiled binaries.
We've been waiting for AMD to publish this new znver1 scheduler model and last week they finally did. Over the past few days, that new scheduler model has been under review. As of this morning, the AMD developer who posted the patch for review is requesting an LLVM developer with commit rights to commit the patch. Now hopefully that will happen prior to the LLVM 5.0 branching tomorrow.
The official release of LLVM 5.0.0 meanwhile is expected on 23 August following a few release candidates while LLVM 6.0.0 will be the next follow-on release.
1 Comment