nnicolosi/Raptor
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

MaterializeMergeSchedule.cpp fix for yolo11_depth_18

Browse Source
This commit is contained in:
NiccoloN
2026-05-24 11:54:00 +02:00
parent 48ca6bd28d
commit 3a7a832198
1 changed files with 17 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes/MaterializeMergeSchedule.cpp
+17 -8
View File
@@ -2350,20 +2350,29 @@ LogicalResult collectPackedRunsForWholeBatchInput(MaterializerState& state,
if (run.sourceOp != key.instance.op || run.resultIndex != key.resultIndex) if (run.sourceOp != key.instance.op || run.resultIndex != key.resultIndex)
continue; continue;
SmallVector<WholeBatchAssemblyRange, 8> runRanges; SmallVector<WholeBatchAssemblyRange, 16> runRanges;
runRanges.reserve(run.slots.size());
for (const PackedScalarRunSlot& slot : run.slots) { for (const PackedScalarRunSlot& slot : run.slots) {
std::optional<ProducerKey> slotKey = getContiguousProducerKeyForKeys(slot.keys); for (ProducerKey fragmentKey : slot.keys) {
if (!slotKey) if (fragmentKey.instance.op != key.instance.op || fragmentKey.resultIndex != key.resultIndex)
return failure(); return failure();
if (wholeBatchRangeOverlaps(plan.coveredRanges, slotKey->instance.laneStart, slotKey->instance.laneCount)) if (fragmentKey.instance.laneCount == 0)
return failure(); return failure();
runRanges.push_back({slotKey->instance.laneStart, slotKey->instance.laneCount}); if (wholeBatchRangeOverlaps(plan.coveredRanges, fragmentKey.instance.laneStart, fragmentKey.instance.laneCount))
return failure();
if (wholeBatchRangeOverlaps(runRanges, fragmentKey.instance.laneStart, fragmentKey.instance.laneCount))
return failure();
runRanges.push_back({fragmentKey.instance.laneStart, fragmentKey.instance.laneCount});
}
} }
if (runRanges.empty())
continue;
plan.packedRuns.push_back(&run); plan.packedRuns.push_back(&run);
for (WholeBatchAssemblyRange range : runRanges) for (WholeBatchAssemblyRange range : runRanges)