diff --git a/include/communicator_quda.h b/include/communicator_quda.h
index 19bbf6e42c..0dc22fdfab 100644
--- a/include/communicator_quda.h
+++ b/include/communicator_quda.h
@@ -741,6 +741,8 @@ namespace quda
 
   void comm_allreduce_sum_array(double *data, size_t size);
 
+  void comm_allreduce_sum(size_t &a);
+
   void comm_allreduce_max_array(double *data, size_t size);
 
   void comm_allreduce_max_array(deviation_t<double> *data, size_t size);
diff --git a/lib/color_spinor_util.in.cu b/lib/color_spinor_util.in.cu
index 3681438c9f..c0baeb0ec5 100644
--- a/lib/color_spinor_util.in.cu
+++ b/lib/color_spinor_util.in.cu
@@ -191,7 +191,7 @@ namespace quda {
   int compareSpinor(const U &u, const V &v, const int tol)
   {
     int fail_check = 16*tol;
-    std::vector<int> fail(fail_check);
+    std::vector<size_t> fail(fail_check);
     for (int f=0; f<fail_check; f++) fail[f] = 0;
 
     int N = 2*u.Nspin()*u.Ncolor();
@@ -224,9 +224,9 @@ namespace quda {
 
     // reduce over all processes
     for (int i=0; i<N; i++) comm_allreduce_int(iter[i]);
-    for (int f=0; f<fail_check; f++) comm_allreduce_int(fail[f]);
+    for (int f = 0; f < fail_check; f++) comm_allreduce_sum(fail[f]);
 
-    for (int i=0; i<N; i++) printfQuda("%d fails = %d\n", i, iter[i]);
+    for (int i = 0; i < N; i++) printfQuda("%d fails = %d\n", i, iter[i]);
 
     int accuracy_level =0;
     for (int f=0; f<fail_check; f++) {
@@ -235,8 +235,8 @@ namespace quda {
 
     size_t total = u.Nparity()*u.VolumeCB()*N*comm_size();
     for (int f=0; f<fail_check; f++) {
-      printfQuda("%e Failures: %d / %lu  = %e\n", pow(10.0,-(f+1)/(double)tol),
-		 fail[f], total, fail[f] / (double)total);
+      printfQuda("%e Failures: %lu / %lu  = %e\n", pow(10.0, -(f + 1) / (double)tol), fail[f], total,
+                 fail[f] / (double)total);
     }
 
     return accuracy_level;
diff --git a/lib/communicator_mpi.cpp b/lib/communicator_mpi.cpp
index c42345ef77..16d6f46c17 100644
--- a/lib/communicator_mpi.cpp
+++ b/lib/communicator_mpi.cpp
@@ -291,6 +291,16 @@ namespace quda
     }
   }
 
+  void Communicator::comm_allreduce_sum(size_t &a)
+  {
+    if (sizeof(size_t) != sizeof(unsigned long)) {
+      errorQuda("sizeof(size_t) != sizeof(unsigned long): %lu != %lu\n", sizeof(size_t), sizeof(unsigned long));
+    }
+    size_t recv;
+    MPI_CHECK(MPI_Allreduce(&a, &recv, 1, MPI_UNSIGNED_LONG, MPI_SUM, MPI_COMM_HANDLE));
+    a = recv;
+  }
+
   void Communicator::comm_allreduce_max_array(deviation_t<double> *data, size_t size)
   {
     size_t n = comm_size();
diff --git a/lib/communicator_qmp.cpp b/lib/communicator_qmp.cpp
index 174c95fa9d..1a4a2a7878 100644
--- a/lib/communicator_qmp.cpp
+++ b/lib/communicator_qmp.cpp
@@ -318,6 +318,14 @@ void Communicator::comm_allreduce_sum_array(double *data, size_t size)
   }
 }
 
+void Communicator::comm_allreduce_sum(size_t &a)
+{
+  if (sizeof(size_t) != sizeof(uint64_t)) {
+    errorQuda("sizeof(size_t) != sizeof(uint64_t): %lu != %lu\n", sizeof(size_t), sizeof(uint64_t));
+  }
+  QMP_CHECK(QMP_comm_sum_uint64_t(QMP_COMM_HANDLE, reinterpret_cast<uint64_t *>(&a)));
+}
+
 void Communicator::comm_allreduce_max_array(deviation_t<double> *data, size_t size)
 {
   size_t n = comm_size();
diff --git a/lib/communicator_single.cpp b/lib/communicator_single.cpp
index e6757a1265..978fbcc069 100644
--- a/lib/communicator_single.cpp
+++ b/lib/communicator_single.cpp
@@ -83,6 +83,8 @@ namespace quda
 
   void Communicator::comm_allreduce_sum_array(double *, size_t) { }
 
+  void Communicator::comm_allreduce_sum(size_t &) { }
+
   void Communicator::comm_allreduce_max_array(deviation_t<double> *, size_t) { }
 
   void Communicator::comm_allreduce_max_array(double *, size_t) { }
diff --git a/lib/communicator_stack.cpp b/lib/communicator_stack.cpp
index 9d6f1eb7e3..fe9d1faaba 100644
--- a/lib/communicator_stack.cpp
+++ b/lib/communicator_stack.cpp
@@ -216,6 +216,8 @@ namespace quda
 
   template <> void comm_allreduce_sum<double>(double &a) { comm_allreduce_sum_array(&a, 1); }
 
+  template <> void comm_allreduce_sum<size_t>(size_t &a) { get_current_communicator().comm_allreduce_sum(a); }
+
   void comm_allreduce_max_array(double *data, size_t size)
   {
     get_current_communicator().comm_allreduce_max_array(data, size);
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 9ca912f8b3..5f4c9808ae 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -801,14 +801,24 @@ endif()
 
   if(QUDA_DIRAC_STAGGERED)
     set(DIRAC_NAME staggered)
-    add_test(NAME dslash_${DIRAC_NAME}_policy${pol2}
+    add_test(NAME dslash_${DIRAC_NAME}_matpc_policy${pol2}
              COMMAND ${QUDA_CTEST_LAUNCH} $<TARGET_FILE:staggered_dslash_ctest> ${MPIEXEC_POSTFLAGS}
                      --dslash-type ${DIRAC_NAME}
                      --test MatPC
                      --dim 2 4 6 8
-                     --gtest_output=xml:dslash_${DIRAC_NAME}_test_pol${pol2}.xml)
+                     --gtest_output=xml:dslash_${DIRAC_NAME}_matpc_test_pol${pol2}.xml)
     if(polenv)
-      set_tests_properties(dslash_${DIRAC_NAME}_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
+      set_tests_properties(dslash_${DIRAC_NAME}_matpc_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
+    endif()
+
+    add_test(NAME dslash_${DIRAC_NAME}_mat_policy${pol2}
+             COMMAND ${QUDA_CTEST_LAUNCH} $<TARGET_FILE:staggered_dslash_ctest> ${MPIEXEC_POSTFLAGS}
+                     --dslash-type ${DIRAC_NAME}
+                     --test Mat
+                     --dim 2 4 6 8
+                     --gtest_output=xml:dslash_${DIRAC_NAME}_mat_test_pol${pol2}.xml)
+    if(polenv)
+      set_tests_properties(dslash_${DIRAC_NAME}_mat_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
     endif()
 
     add_test(NAME benchmark_dslash_${DIRAC_NAME}_policy${pol2}
@@ -824,15 +834,26 @@ endif()
     endif()
 
     set(DIRAC_NAME asqtad)
-    add_test(NAME dslash_${DIRAC_NAME}_policy${pol2}
+    add_test(NAME dslash_${DIRAC_NAME}_matpc_policy${pol2}
              COMMAND ${QUDA_CTEST_LAUNCH} $<TARGET_FILE:staggered_dslash_ctest> ${MPIEXEC_POSTFLAGS}
                      --dslash-type ${DIRAC_NAME}
                      --all-partitions 1
                      --test MatPC
                      --dim 6 8 10 12
-                     --gtest_output=xml:dslash_${DIRAC_NAME}_test_pol${pol2}.xml)
+                     --gtest_output=xml:dslash_${DIRAC_NAME}_matpc_test_pol${pol2}.xml)
+    if(polenv)
+      set_tests_properties(dslash_${DIRAC_NAME}_matpc_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
+    endif()
+
+    add_test(NAME dslash_${DIRAC_NAME}_mat_policy${pol2}
+             COMMAND ${QUDA_CTEST_LAUNCH} $<TARGET_FILE:staggered_dslash_ctest> ${MPIEXEC_POSTFLAGS}
+                     --dslash-type ${DIRAC_NAME}
+                     --all-partitions 1
+                     --test Mat
+                     --dim 6 8 10 12
+                     --gtest_output=xml:dslash_${DIRAC_NAME}_mat_test_pol${pol2}.xml)
     if(polenv)
-      set_tests_properties(dslash_${DIRAC_NAME}_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
+      set_tests_properties(dslash_${DIRAC_NAME}_mat_policy${pol2} PROPERTIES ENVIRONMENT QUDA_ENABLE_DSLASH_POLICY=${pol2})
     endif()
 
     add_test(NAME benchmark_dslash_${DIRAC_NAME}_policy${pol2}
diff --git a/tests/host_reference/staggered_dslash_reference.cpp b/tests/host_reference/staggered_dslash_reference.cpp
index e95532590f..86ecd17464 100644
--- a/tests/host_reference/staggered_dslash_reference.cpp
+++ b/tests/host_reference/staggered_dslash_reference.cpp
@@ -116,10 +116,10 @@ void staggeredDslashReference(sFloat *res, gFloat **fatlink, gFloat **longlink,
           sub(&res[offset], &res[offset], gaugedSpinor, stag_spinor_site_size);
         }
       }
+    } // forward/backward in all four directions
 
-      if (daggerBit) negx(&res[offset], stag_spinor_site_size);
-    } // 4-d volume
-  }   // right-hand-side
+    if (daggerBit) negx(&res[offset], stag_spinor_site_size);
+  } // 4-d volume
 }
 
 void staggeredDslash(ColorSpinorField &out, void **fatlink, void **longlink, void **ghost_fatlink,