[llvm-branch-commits] [llvm] [offload] Copy the kernel launch environment with a direct-write memcpy (PR #231923)
Alex Duran via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Oct 28 13:51:55 PDT 2026
================
@@ -772,12 +772,22 @@ getKernelLaunchEnvironment(DeviceTy &Device, uint32_t OmpABIVersion,
freeAfterSynchronization(Device, AsyncInfo, *ReductionAllocOrErr);
}
- if (Device.submitData(*AllocOrErr, &LocalKLE,
- sizeof(KernelLaunchEnvironmentTy), AsyncInfo,
- /*Entry=*/nullptr, /*HDTTMapPtr=*/nullptr,
- /*Loc=*/nullptr, /*Name=*/"KernelLaunchEnv"))
+ INFO(OMP_INFOTYPE_DATA_TRANSFER, Device.DeviceID,
+ "Copying data from host to device, HstPtr=" DPxMOD ", TgtPtr=" DPxMOD
+ ", Size=%" PRId64 ", Name=KernelLaunchEnv\n",
+ DPxPTR(&LocalKLE), DPxPTR(*AllocOrErr),
+ sizeof(KernelLaunchEnvironmentTy));
+
+ // The allocation is fresh, so no work pending on the queue accesses it and
+ // the copy can bypass the queue. Without a queue, the plugin may write the
+ // memory directly.
+ if (const auto *Res = olMemcpy(
----------------
adurang wrote:
Any reason to use olMemcpy directly here? I think you could use submitDataSync.
https://github.com/llvm/llvm-project/pull/231923
More information about the llvm-branch-commits
mailing list