[OPTIMIZER] Fixed up order of shared layouts (#881)
This commit is contained in:
@@ -201,7 +201,9 @@ LogicalResult LoopPipeliner::initialize() {
|
||||
ty.getShape().end());
|
||||
bufferShape.insert(bufferShape.begin(), numStages);
|
||||
auto sharedEnc = ttg::SharedEncodingAttr::get(
|
||||
ty.getContext(), dotOpEnc, ty.getShape(), ty.getElementType());
|
||||
ty.getContext(), dotOpEnc, ty.getShape(),
|
||||
triton::gpu::getOrder(ty.getEncoding()),
|
||||
ty.getElementType());
|
||||
loadsBufferType[loadOp] = RankedTensorType::get(
|
||||
bufferShape, ty.getElementType(), sharedEnc);
|
||||
}
|
||||
|
Reference in New Issue
Block a user