diff --git a/examples/tests/spmv_bcrs_laplace_test.hpp b/examples/tests/spmv_bcrs_laplace_test.hpp
index 9b932ed..badaeea 100644
--- a/examples/tests/spmv_bcrs_laplace_test.hpp
+++ b/examples/tests/spmv_bcrs_laplace_test.hpp
@@ -185,7 +185,7 @@ REGISTER_TEST(bspmv_laplace_test) {
     A_crs->val[A_crs->nnz - 2] = double(-1);
     A_crs->val[A_crs->nnz - 1] = double(2);
 
-    // test_blocked_conversion(A_crs, true, print_dense_matrix);
+    test_blocked_conversion(A_crs, true, print_dense_matrix);
     test_blocked_conversion(A_crs, false, print_dense_matrix);
 
     delete A_crs;
diff --git a/src/kernels/spmv/spmv_cuda/spmv_cuda_bcrs_impl.cu b/src/kernels/spmv/spmv_cuda/spmv_cuda_bcrs_impl.cu
index 9d6e698..b5dfa9e 100644
--- a/src/kernels/spmv/spmv_cuda/spmv_cuda_bcrs_impl.cu
+++ b/src/kernels/spmv/spmv_cuda/spmv_cuda_bcrs_impl.cu
@@ -161,7 +161,7 @@ __global__ void naive_bcrs_spmv_cuda_warp_per_row_by_shffl(
         for (ULL cur_cutoff = power_hint; cur_cutoff >= ULL(1);
              cur_cutoff /= ULL(2)) {
             loc_sum +=
-                __shfl_down_sync(0xffff, loc_sum, int(cur_cutoff * b_height));
+                __shfl_down_sync(0xffffffff, loc_sum, int(cur_cutoff * b_height));
         }
         // now correct values are inside the first b_height threads
         if (thread_idx < b_height)