diff --git a/test/host/hls/test.cpp b/test/host/hls/test.cpp
index 12b2e35a..1da8bb48 100644
--- a/test/host/hls/test.cpp
+++ b/test/host/hls/test.cpp
@@ -143,52 +143,6 @@ std::unique_ptr<ACCL::ACCL> test_vadd_put(options_t options) {
     return accl;
 }
 
-void test_copy(ACCL::ACCL& accl, options_t options) {
-    //run test here:
-    //initialize a CCLO BFM and streams as needed
-    hlslib::Stream<command_word> callreq, callack;
-    hlslib::Stream<stream_word, 512> data_cclo2krnl("cclo2krnl"), data_krnl2cclo("krnl2cclo");
-
-    std::vector<unsigned int> dest = {0};
-    std::unique_ptr<CCLO_BFM> cclo;
-
-    if (!options.hardware) {
-        cclo = std::make_unique<CCLO_BFM>(options.start_port, rank, size, dest, callreq, callack, data_cclo2krnl, data_krnl2cclo);
-        cclo->run();
-        std::cout << "CCLO BFM started" << std::endl;
-    }
-    MPI_Barrier(MPI_COMM_WORLD);
-
-    //allocate float arrays for the HLS function to use
-    auto src_buffer = accl.create_buffer<int>(options.count, ACCL::dataType::int32, 0);
-    auto dst_buffer = accl.create_buffer<int>(options.count, ACCL::dataType::int32, 0);
-    for(int i=0; i<options.count; i++){
-        src_buffer->buffer()[i] = rank;
-        dst_buffer->buffer()[i] = 0;
-    }
-
-    accl.copy_to_stream(*src_buffer, options.count, false);
-
-    //loop back data (divide count by 16 and round up to get number of stream words)
-    for (int i=0; i < (options.count+15)/16; i++) {
-        data_krnl2cclo.write(data_cclo2krnl.read());
-    }
-
-    accl.copy_from_stream(*dst_buffer, options.count, false);
-
-    //check HLS function outputs
-    unsigned int err_count = 0;
-    for(int i=0; i<options.count; i++){
-        err_count += (dst_buffer->buffer()[i] != rank);
-    }
-
-    std::cout << "Test finished with " << err_count << " errors" << std::endl;
-    if (!options.hardware) {
-        //clean up
-        cclo->stop();
-    }
-}
-
 void test_loopback_local_res(ACCL::ACCL& accl, options_t options) {
 
     //run test here:
@@ -344,8 +298,6 @@ int main(int argc, char *argv[]) {
 
     auto accl = test_vadd_put(options);
     MPI_Barrier(MPI_COMM_WORLD);
-    test_copy(*accl, options);
-    MPI_Barrier(MPI_COMM_WORLD);
     if(!options.hardware){
         std::srand(42);
         for(int i=0; i<options.nruns; i++){
diff --git a/test/host/xrt/test.cpp b/test/host/xrt/test.cpp
index 98caa90d..c5edfb47 100644
--- a/test/host/xrt/test.cpp
+++ b/test/host/xrt/test.cpp
@@ -119,6 +119,37 @@ void test_copy(ACCL::ACCL &accl, options_t &options) {
   }
 }
 
+void test_copy_stream(ACCL::ACCL &accl, options_t &options) {
+  std::cout << "Start copy stream test..." << std::endl;
+  unsigned int count = options.count;
+  auto op_buf = accl.create_buffer<float>(count, dataType::float32);
+  auto res_buf = accl.create_buffer<float>(count, dataType::float32);
+  random_array(op_buf->buffer(), count);
+
+  test_debug("Copy data from buffer to stream", options);
+  accl.copy_to_stream(*op_buf, count, false);
+  test_debug("Copy data from stream to buffer", options);
+  accl.copy_from_stream(*res_buf, count, false);
+  int errors = 0;
+  for (unsigned int i = 0; i < count; ++i) {
+    float ref = (*op_buf)[i];
+    float res = (*res_buf)[i];
+    if (res != ref) {
+      std::cout << i + 1
+                << "th item is incorrect! (" + std::to_string(res) +
+                       " != " + std::to_string(ref) + ")"
+                << std::endl;
+      errors += 1;
+    }
+  }
+
+  if (errors > 0) {
+    std::cout << errors << " errors!" << std::endl;
+  } else {
+    std::cout << "Test succesfull!" << std::endl;
+  }
+}
+
 void test_copy_p2p(ACCL::ACCL &accl, options_t &options) {
   std::cout << "Start copy p2p test..." << std::endl;
   unsigned int count = options.count;
@@ -1304,7 +1335,7 @@ int start_test(options_t options) {
     auto cclo_ip = xrt::ip(device, xclbin_uuid,
                            "ccl_offload:{ccl_offload_" + cclo_id + "}");
     auto hostctrl_ip = xrt::kernel(device, xclbin_uuid,
-                                   "hostctrl:{hostctrl_" + cclo_id + "_0}",
+                                   "hostctrl:{hostctrl_" + cclo_id + "}",
                                    xrt::kernel::cu_access_mode::exclusive);
 
     int devicemem;
@@ -1350,6 +1381,8 @@ int start_test(options_t options) {
   MPI_Barrier(MPI_COMM_WORLD);
   test_copy(*accl, options);
   MPI_Barrier(MPI_COMM_WORLD);
+  test_copy_stream(*accl, options);
+  MPI_Barrier(MPI_COMM_WORLD);
   test_copy_p2p(*accl, options);
   MPI_Barrier(MPI_COMM_WORLD);
   test_combine_sum(*accl, options);