Add block_sync_lds between two consecutive blockwise reductions

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp

@@ -330,6 +330,9 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
         }
 
         static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
             BlockwiseReduce::Reduce(reduce_work_buf, dscale_thread_buf(I));
             block_sync_lds();
             BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp

@@ -495,6 +495,9 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
         };
 
         static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
             BlockwiseReduce::Reduce(reduce_work_buf, reduce_dscale_thread_buf(I));
             block_sync_lds();
             BlockwiseReduce::Reduce(reduce_work_buf, reduce_dbias_thread_buf(I));