[MLIR][AMDGPU] Switch to code object version 5

[saiislam]

As AMDGPU backend has moved to cov5 as default, mlir should also switch to it.

[llvmbot]

@llvm/pr-subscribers-mlir-llvm

@llvm/pr-subscribers-mlir

Author: Saiyedul Islam (saiislam)

Changes

As AMDGPU backend has moved to cov5 as default, mlir should also switch to it.

---

Full diff: https://github.com/llvm/llvm-project/pull/79144.diff

1 Files Affected:

• (modified) mlir/lib/Target/LLVM/ROCDL/Target.cpp (+1-1)

diff --git a/mlir/lib/Target/LLVM/ROCDL/Target.cpp b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
index a589e5cbbb6198..cdcef1d6b459c6 100644
--- a/mlir/lib/Target/LLVM/ROCDL/Target.cpp
+++ b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
@@ -238,7 +238,7 @@ void SerializeGPUModuleBase::addControlVariables(
   addControlVariable("__oclc_wavefrontsize64", wave64);
 
   llvm::Type *i32Ty = llvm::Type::getInt32Ty(module.getContext());
-  int abi = 400;
+  int abi = 500;
   abiVer.getAsInteger(0, abi);
   llvm::GlobalVariable *abiVersion = new llvm::GlobalVariable(
       module, i32Ty, true, llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,

[arsenm]

I would have hoped leaving it at 4 would have worked, but maybe this is missing something else from the controls. Is it not emitting the module flag?

[fabianmcg]

LGTM!

I would have hoped leaving it at 4 would have worked, but maybe this is missing something else from the controls. Is it not emitting the module flag?

I haven't tested ROCM 6.0. As far as remember, it was still working with ROCM 5.6 and 5.7. However, I'll test it and see if there's a missing control variable.

[krzysz00]

@arsenm We haven't historically been setting that module flag

[krzysz00]

(also, does this commit address SerializeToHsaco as well?)

[saiislam]

(also, does this commit address SerializeToHsaco as well?)

If we mean this change, then yes it has been done.
addControlConstant("__oclc_ABI_version", 500, 32);

[arsenm]

@arsenm We haven't historically been setting that module flag

So yes. Without the flag you are subject to the default of the backend which is now mismatched with the device lib link. You should just set both for next time

[krzysz00]

IIRC, when the code for this was initially written (and so when it was ported into its current form) there wasn't such a flag to set or no one involved knew about it.

I can get this fixed in a followup patch