diff --git a/ethosu/regor/compiler/high_level_command_stream_generator.cpp b/ethosu/regor/compiler/high_level_command_stream_generator.cpp
index c73a089186feb013d1a935f8a88cca0783560df9..4eb673c62d6c8ce84ee180ca883222e0d7f74086 100644
--- a/ethosu/regor/compiler/high_level_command_stream_generator.cpp
+++ b/ethosu/regor/compiler/high_level_command_stream_generator.cpp
@@ -671,7 +671,7 @@ void HLCStreamGenerator::GenerateHLCStripeCommands(SchedulerOperation *op, const
                     auto inputArea = TransformWithStridesAndSkirt(outputArea, &strides, ifmConn->stepXY, &skirt, ifmConn->shape,
                         opType, ofmConn->slice.offset, ifmConn->slice.offset, ifmConn->slice.shape, dilatedKernelHeight,
                         upscaling, hlcStripe->padding.top, hlcStripe->padding.bottom, ifmLimit, ofmConn->transpose, accIfm);
-                    if ( ofmConn->stepXY != Point2i{1, 1} || ifmConn->stepXY != Point2i{1, 1} )
+                    if ( !accIfm && (ofmConn->stepXY != Point2i{1, 1} || ifmConn->stepXY != Point2i{1, 1}) )
                     {
                         std::tie(inputArea, hlcStripe->padding) = TransformWithInputOutputSteps(inputArea,
                             ifmConn->stepXY, outputArea, ofmConn->stepXY, kernel, hlcStripe->padding, ifmConn->shape);
diff --git a/ethosu/regor/compiler/scheduler_decompose.cpp b/ethosu/regor/compiler/scheduler_decompose.cpp
index 947f7821ff9b6a2da35373dadc7436ba86f8048c..defa6679c8627c8cacbc02464d05288eb2929ca2 100644
--- a/ethosu/regor/compiler/scheduler_decompose.cpp
+++ b/ethosu/regor/compiler/scheduler_decompose.cpp
@@ -448,6 +448,14 @@ HandleDilation(Architecture *arch, std::unique_ptr<SchedulerOperation> op, const
                 auto *subOfmConn = subOp->Output(TensorUsage::OFM);
                 subOfmConn->slice = std::move(newOfmSlice);
                 subOfmConn->stepXY = ofmStrides;
+                if ( subOp->AccumulatorMode().source == AccumulatorSource::Ifm2 )
+                {
+                    auto *subIfm2Conn = subOp->Input(TensorUsage::IFM1);
+                    subIfm2Conn->slice.shape = subOfmConn->slice.shape;
+                    subIfm2Conn->slice.offset = subIfm2Conn->slice.shape.WithZeros().WithHW(
+                        subOfmConn->slice.offset.Height(), subOfmConn->slice.offset.Width());
+                    subIfm2Conn->stepXY = subOfmConn->stepXY;
+                }
                 auto subOps = doDecompose(arch, std::move(subOp));
                 result.insert(result.end(), std::make_move_iterator(subOps.begin()), std::make_move_iterator(subOps.end()));
             }