diff --git a/nvbandwidth.cpp b/nvbandwidth.cpp
index 61a228f..488372a 100644
--- a/nvbandwidth.cpp
+++ b/nvbandwidth.cpp
@@ -29,6 +29,7 @@
 #include "kernels.cuh"
 #include "output.h"
 #include "testcase.h"
+#include "testcases_patched.h"
 #include "version.h"
 #include "inline_common.h"

@@ -73,8 +74,10 @@ std::vector<Testcase*> createTestcases() {
         new DeviceToDeviceBidirWriteCE(),
         new AllToHostCE(),
         new AllToHostBidirCE(),
+        new AllToHostBatchCE(),
         new HostToAllCE(),
         new HostToAllBidirCE(),
+        new HostToAllBatchCE(),
         new AllToOneWriteCE(),
         new AllToOneReadCE(),
         new OneToAllWriteCE(),
@@ -89,8 +92,10 @@ std::vector<Testcase*> createTestcases() {
         new DeviceToDeviceBidirWriteSM(),
         new AllToHostSM(),
         new AllToHostBidirSM(),
+        new AllToHostBatchSM(),
         new HostToAllSM(),
         new HostToAllBidirSM(),
+        new HostToAllBatchSM(),
         new AllToOneWriteSM(),
         new AllToOneReadSM(),
         new OneToAllWriteSM(),
diff --git a/testcase.h b/testcase.h
index c276850..f26e7d8 100644
--- a/testcase.h
+++ b/testcase.h
@@ -39,6 +39,7 @@ class Testcase {
     void oneToAllHelper(unsigned long long size, MemcpyOperation &memcpyInstance, PeerValueMatrix<double> &bandwidthValues, bool isRead);
     void allHostHelper(unsigned long long size, MemcpyOperation &memcpyInstance, PeerValueMatrix<double> &bandwidthValues, bool sourceIsHost);
     void allHostBidirHelper(unsigned long long size, MemcpyOperation &memcpyInstance, PeerValueMatrix<double> &bandwidthValues, bool sourceIsHost);
+    void allHostHelperBatch(unsigned long long size, MemcpyOperation &memcpyInstance, PeerValueMatrix<double> &bandwidthValues, bool sourceIsHost);
     void latencyHelper(const MemcpyBuffer &dataBuffer, bool measureDeviceToDeviceLatency);

  public: