================
@@ -751,6 +764,15 @@ void CGOpenMPRuntimeGPU::emitKernelInit(const
OMPExecutableDirective &D,
: llvm::omp::OMPTgtExecModeFlags::OMP_TGT_EXEC_MODE_GENERIC;
computeMinAndMaxThreadsAndTeams(D, CGF, Attrs);
+ // Set MaxThreads to twice the default workgroup size for SPMD cross-team
+ // reductions. For these reductions, we want to have fewer, larger teams to
+ // make better use of intra-team reduction and not over-emphasize the
+ // inter-team part. Note that this only buys us more freedom. The actual grid
+ // size for the launch is going to be chosen later by the offload plugin.
+ if (IsSPMD && Attrs.MaxThreads.front() < 0 && hasTeamsReduction(D))
+ Attrs.MaxThreads.front() =
+ 2 * CGM.getTarget().getGridValue().GV_Default_WG_Size;
----------------
ro-i wrote:
I actually previously followed the strategy to aim for 50% occupancy (using
getHardwareParallelism()). But it seemed too random.
Being based on the default total number of threads means that the reductions
don't arbitrarily aim for a different total number of threads. They just
redistribute that total number by doubling the team size and cutting the number
of teams in half.
The reason why this makes sense is that we know the algorithm that is executed
by the cross-team reduction and that it benefits from a tendency towards larger
and fewer teams due to less inter-team synchronization and less work in the
final team.
This is far from perfect. It's just re-using the simple rule that we have for
ROCm in a more portable and less arbitrary way. See e.g.
https://github.com/ROCm/llvm-project/blob/7f90160e0c374525bbfd7d3485949f9130cb732b/llvm/include/llvm/Frontend/OpenMP/OMPConstants.h#L312.
Unfortunately, ROCm has 1000 aspects and tuning knobs that affect the actual
grid layout. As I mentioned, I'm trying to get started with some conservative
and simple rule.
https://github.com/llvm/llvm-project/pull/208253
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
