================
@@ -422,6 +427,19 @@ uint32_t GenericKernelTy::getEffectiveNumBlocks(
// will execute one iteration of the loop; rounded up to the nearest
// integer. However, if that results in too few blocks, we artificially
// reduce the thread count per block to increase the outer parallelism.
+
+ // Favor fewer, larger teams for cross-team reductions. As a heuristic,
we
+ // aim for double the default number of threads and half the default
+ // number of blocks.
+ if (isCrossTeamReduction()) {
+ // Never launch more teams than the trip count needs.
+ uint64_t NumBlocks =
----------------
dhruvachak wrote:
Your change halves the number of blocks even if the blocksize could not be
increased for some reason. To guard against that problem, guard it like below.
Additionally, instead of a hardcoded 2, use a scale-factor.
```
if (isCrossTeamReduction() && EffectiveNumThreads > PreferredNumThreads) {
uint64_t Scaled = std::max<uint64_t>(
(uint64_t(DefaultNumBlocks) * PreferredNumThreads) / EffectiveNumThreads,
1);
uint64_t NumBlocks = std::min<uint64_t>(
Scaled, ((LoopTripCount - 1) / EffectiveNumThreads) + 1);
return std::min(NumBlocks,
uint64_t(GenericDevice.getBlockLimit(EffectiveNumThreads)));
}
```
https://github.com/llvm/llvm-project/pull/208253
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
