Skip to content

Commit 5b5d978

Browse files
committed
GPU TPC: synchronize threads in warp between fit iterations to reduce warp divergence
1 parent ec7b77b commit 5b5d978

1 file changed

Lines changed: 1 addition & 0 deletions

File tree

‎GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -94,6 +94,7 @@ GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger& GPUrestrict() merger, int32_
9494
const int8_t adjacentLeg = 1 + (track.Leg() == 0);
9595

9696
for (int32_t iWay = rebuilt ? (wayRebuildStart + 2) : 0; iWay < nWays; iWay++) { // TODO DR: Unrolling has no performance improvement on GPU, why?
97+
GPUbarrierWarp(); // TODO: Validate on different GPUs, check if we should do similar things in other places
9798
int32_t nMissed = 0, nMissedHard = 0; // nMissed: cluster rejected, fit failure. nMissedHard: also propagation failure / high incl.
9899
float sumInvSqrtCharge = 0.f; // TODO: Compute in first iteration and store!
99100
int32_t nAvgCharge = 0;

0 commit comments

Comments
 (0)