diff --git a/openmp/libomptarget/plugins/hero/CMakeLists.txt b/openmp/libomptarget/plugins/hero/CMakeLists.txt
index 9309e57c25e196..b425ac48fcc756 100644
--- a/openmp/libomptarget/plugins/hero/CMakeLists.txt
+++ b/openmp/libomptarget/plugins/hero/CMakeLists.txt
@@ -17,7 +17,7 @@ if(NOT LIBOMPTARGET_HERO_BUILD)
   return()
 endif()
 
-if(NOT EXISTS $ENV{HERO_DEV_INC_DIR}/libhero/herodev.h)
+if(NOT EXISTS $ENV{HERO_DEV_INC_DIR}/libhero/hero_api.h)
   libomptarget_say("Not building HERO offloading plugin: HERO device SDK not found (HERO_DEV_INC_DIR to the correct path)")
   return()
 endif()
diff --git a/openmp/libomptarget/plugins/hero/plugin-hero-dev.h b/openmp/libomptarget/plugins/hero/plugin-hero-dev.h
index a0558aa0719399..40f8c92330552b 100644
--- a/openmp/libomptarget/plugins/hero/plugin-hero-dev.h
+++ b/openmp/libomptarget/plugins/hero/plugin-hero-dev.h
@@ -56,7 +56,7 @@
 
 extern "C" {
 
-#include "libhero/herodev.h"
+#include "libhero/hero_api.h"
 
 #define HERO_DEV_DEFAULT_CLUSTER_ID (0x1U)
 #define HERO_DEV_DEFAULT_FREQ (HERO_DEV_DEFAULT_FREQ_MHZ)
@@ -126,19 +126,16 @@ static int init_hero_device() {
 
   // reserve virtual addresses overlapping with HERO Device's internal physical address
   // space
-  hero_dev_reserve_v_addr(hd);
   ret = hero_dev_mmap(hd);
   if (ret < 0) {
     TRACE("ERROR: cannot load device!");
     return ret;
   }
 
-  currFreq = hero_dev_clking_set_freq(hd, HERO_DEV_DEFAULT_FREQ);
-  if (currFreq > 0)
-    TRACE("HERO Device running @ %d MHz.", currFreq);
-  //  else
-  //    GOMP_PLUGIN_fatal("HERO device init failed!");
-  hero_dev_rab_free(hd, 0x0);
+  //currFreq = hero_dev_clking_set_freq(hd, HERO_DEV_DEFAULT_FREQ);
+  //if (currFreq > 0)
+  //  TRACE("HERO Device running @ %d MHz.", currFreq);
+
   hero_dev_reset(hd, 0x1);
 
   // initialization of HERO Device, static RAB rules (mbox, L2, ...)
@@ -156,7 +153,7 @@ static int init_hero_device() {
 }
 
 static int deinit_hero_device() {
-  hero_dev_munmap(hd);
+  //hero_dev_munmap(NULL);
 }
 
 extern "C" bool GOMP_OFFLOAD_init_device(int n __attribute__((unused))) {
@@ -174,14 +171,7 @@ extern "C" bool GOMP_OFFLOAD_fini_device(int n __attribute__((unused))) {
 
   TRACE("Waiting for EOC...");
   hero_dev_exe_wait(hd, HERO_DEV_DEFAULT_TIMEOUT);
-
   hero_dev_exe_stop(hd);
-
-  if (GOMP_OFFLOAD_get_caps() & GOMP_OFFLOAD_CAP_SHARED_MEM)
-    hero_dev_rab_soc_mh_disable(hd);
-
-  hero_dev_rab_free(hd, 0);
-  hero_dev_free_v_addr(hd);
   hero_dev_munmap(hd);
 
   return 1;
diff --git a/openmp/libomptarget/plugins/hero/rtl.cpp b/openmp/libomptarget/plugins/hero/rtl.cpp
index cc0e4b812f7034..0dfc6da87af369 100644
--- a/openmp/libomptarget/plugins/hero/rtl.cpp
+++ b/openmp/libomptarget/plugins/hero/rtl.cpp
@@ -96,7 +96,7 @@ int32_t __tgt_rtl_init_device(int32_t device_id) {
 int32_t __tgt_rtl_unregister_lib(__tgt_bin_desc *Desc) {
   DP("__tgt_rtl_unregister_lib\n");
   if (initialized) {
-    deinit_hero_device();
+    hero_dev_munmap(hd);
     initialized = false;
   }
   return OFFLOAD_SUCCESS;
@@ -325,7 +325,7 @@ bool load_and_execute_image(__tgt_device_image *image) {
 
 __tgt_target_table *__tgt_rtl_load_binary(int32_t device_id,
                                           __tgt_device_image *image) {
-  hero_add_timestamp("load_binary",__func__,0);
+  // hero_add_timestamp("load_binary",__func__,0);
   DP("__tgt_rtl_load_binary(%d, " DPxMOD ")\n", device_id, DPxPTR(image));
 
   DP("Dev %d: load binary from " DPxMOD " image\n", device_id,
@@ -343,7 +343,7 @@ __tgt_target_table *__tgt_rtl_load_binary(int32_t device_id,
     return NULL;
   }
 
-  hero_add_timestamp("init_omp",__func__,0);
+  // hero_add_timestamp("init_omp",__func__,0);
 
   // init argument buffers
   host_arg_buf.reserve(ARG_BUF_SIZE); // memory leak
@@ -427,7 +427,7 @@ int32_t __tgt_rtl_data_retrieve(int32_t device_id, void *hst_ptr, void *tgt_ptr,
     assert(hst_ptr == tgt_ptr);
     return OFFLOAD_SUCCESS;
   }
-  //char toprint[128];
+  char toprint[128];
   //snprintf(toprint, 128, "retrieve-%x", hst_ptr);
   //hero_add_timestamp(toprint,__func__,0);
   ret = GOMP_OFFLOAD_dev2host(device_id, hst_ptr, tgt_ptr, size)
@@ -445,7 +445,7 @@ int32_t __tgt_rtl_data_delete(int32_t device_id, void *tgt_ptr) {
   if (device_id == HERODEV_SVM) {
     return OFFLOAD_SUCCESS;
   }
-  //char toprint[128];
+  char toprint[128];
   //snprintf(toprint, 128, "dealloc-%x", tgt_ptr);
   //hero_add_timestamp(toprint,__func__,0);
   ret = GOMP_OFFLOAD_free(device_id, tgt_ptr) ? OFFLOAD_SUCCESS : OFFLOAD_FAIL;
@@ -460,7 +460,7 @@ int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr,
                                          int32_t arg_num, int32_t team_num,
                                          int32_t thread_limit,
                                          uint64_t loop_tripcount) {
-  hero_add_timestamp("send_args",__func__,0);
+  //hero_add_timestamp("send_args",__func__,0);
   for (int32_t i = 0; i < arg_num; i++) {
     if (tgt_offsets[i] != 0) {
       fprintf(stderr, "Unimplemented: non-zero offset");
@@ -477,7 +477,7 @@ int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr,
   size_t size = sizeof(uint64_t) * host_arg_buf.size();
   __tgt_rtl_data_submit(device_id, dev_arg_buf, host_buf, size);
 
-  hero_add_timestamp("offload_wait",__func__,0);
+  // hero_add_timestamp("offload_wait",__func__,0);
 
   hero_dev_mbox_write(hd, MBOX_DEVICE_START);
   hero_dev_mbox_write(hd, (uint32_t)tgt_entry_ptr);
@@ -501,7 +501,7 @@ int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr,
   DP("Done offloading, cycles to execute kernel: %d!\n", (int)ret[1]);
   hero_device_cycles[hero_num_device_cycles++] = (uint32_t) ret[1];
 
-  hero_add_timestamp("offload_return",__func__,0);
+  // hero_add_timestamp("offload_return",__func__,0);
 
   return OFFLOAD_SUCCESS;
 }
diff --git a/openmp/libomptarget/src/omptarget.cpp b/openmp/libomptarget/src/omptarget.cpp
index 703355c70bcccb..f4382ab91bd618 100644
--- a/openmp/libomptarget/src/omptarget.cpp
+++ b/openmp/libomptarget/src/omptarget.cpp
@@ -22,7 +22,7 @@
 
 #ifdef LIBOMPTARGET_HERO
 extern "C" {
-#include "libhero/herodev.h"
+#include "libhero/hero_api.h"
 }
 #endif
 
@@ -1286,7 +1286,7 @@ static int processDataBefore(ident_t *loc, int64_t DeviceId, void *HostPtr,
                              AsyncInfoTy &AsyncInfo) {
   TIMESCOPE_WITH_NAME_AND_IDENT("mappingBeforeTargetRegion", loc);
 #ifdef LIBOMPTARGET_HERO
-  hero_add_timestamp((char*)"all",(char*)__func__,0);
+  //hero_add_timestamp((char*)"all",(char*)__func__,0);
 #endif
   DeviceTy &Device = *PM->Devices[DeviceId];
   int Ret = targetDataBegin(loc, Device, ArgNum, ArgBases, Args, ArgSizes,
@@ -1417,7 +1417,7 @@ static int processDataAfter(ident_t *loc, int64_t DeviceId, void *HostPtr,
                             PrivateArgumentManagerTy &PrivateArgumentManager,
                             AsyncInfoTy &AsyncInfo) {
 #ifdef LIBOMPTARGET_HERO
-  hero_add_timestamp((char*)"all",(char*)__func__,0);
+  //hero_add_timestamp((char*)"all",(char*)__func__,0);
 #endif
   //TIMESCOPE_WITH_NAME_AND_IDENT("mappingAfterTargetRegion", loc);
   DeviceTy &Device = *PM->Devices[DeviceId];