diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2838aa8a..b62db6a6 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1022,7 +1022,10 @@ enable_testing()
 # Recurse into the "Apex" and "Demo" subdirectories. This does not actually
 # cause another cmake executable to run. The same process will walk through
 # the project's entire directory structure.
-add_subdirectory (src/perfetto_sdk)
+if(APEX_WITH_PERFETTO)
+    add_definitions("-DAPEX_WITH_PERFETTO")
+    add_subdirectory (src/perfetto_sdk)
+endif()
 add_subdirectory (src/apex)
 #add_subdirectory (src/ITTNotify)
 add_subdirectory (src/comm)
diff --git a/cmake/Modules/APEX_DefaultOptions.cmake b/cmake/Modules/APEX_DefaultOptions.cmake
index 2a014c45..61a655c8 100644
--- a/cmake/Modules/APEX_DefaultOptions.cmake
+++ b/cmake/Modules/APEX_DefaultOptions.cmake
@@ -16,6 +16,7 @@ option (APEX_WITH_STARPU "Enable APEX StarPU support" FALSE)
 option (APEX_WITH_TCMALLOC "Enable TCMalloc heap management" FALSE)
 option (APEX_WITH_JEMALLOC "Enable JEMalloc heap management" FALSE)
 option (APEX_WITH_LM_SENSORS "Enable LM Sensors support" FALSE)
+option (APEX_WITH_PERFETTO "Enable native Perfetto trace support" TRUE)
 option (APEX_BUILD_TESTS "Build APEX tests (for 'make test')" FALSE)
 option (APEX_CUDA_TESTS "Build APEX CUDA tests (for 'make test')" FALSE)
 option (APEX_HIP_TESTS "Build APEX HIP tests (for 'make test')" FALSE)
diff --git a/cmake/Modules/FindHWLOC.cmake b/cmake/Modules/FindHWLOC.cmake
index dcdbbc90..bf0e7a04 100644
--- a/cmake/Modules/FindHWLOC.cmake
+++ b/cmake/Modules/FindHWLOC.cmake
@@ -54,9 +54,23 @@ endif()
 if(NOT DEFINED $HWLOC_DIR)
     if(DEFINED $HWLOC_ROOT)
         set(HWLOC_DIR $HWLOC_ROOT)
+    else()
+        # Search using CUDAToolkit_ROOT
+        find_program(HWLOC_LSTOPO_EXECUTABLE
+            NAMES lstopo lstopo-no-graphics hwloc-ls
+            PATHS ENV{PATH}
+            PATH_SUFFIXES bin
+        )
+        if(HWLOC_LSTOPO_EXECUTABLE)
+            get_filename_component(hwloc_dir "${HWLOC_LSTOPO_EXECUTABLE}" DIRECTORY)
+            set(HWLOC_BIN_DIR "${hwloc_dir}" CACHE PATH "" FORCE)
+            unset(hwloc_dir)
+            get_filename_component(HWLOC_DIR ${HWLOC_BIN_DIR} DIRECTORY ABSOLUTE)
+        endif()
     endif()
 endif()
 
+
 include(CheckStructHasMember)
 include(CheckCSourceCompiles)
 if (NOT HWLOC_FOUND)
diff --git a/src/apex/CMakeLists.hpx b/src/apex/CMakeLists.hpx
index aae50402..2c9629c6 100644
--- a/src/apex/CMakeLists.hpx
+++ b/src/apex/CMakeLists.hpx
@@ -255,6 +255,13 @@ if(APEX_WITH_OTF2)
   set(otf2_sources otf2_listener.cpp otf2_listener_hpx.cpp otf2_listener_nompi.cpp)
 endif()
 
+if(APEX_WITH_PERFETTO)
+  set(perfetto_headers perfetto_listener.hpp perfetto_static.hpp)
+  set(perfetto_sources
+    perfetto_listener.cpp perfetto_static.cpp ../perfetto_sdk/perfetto.cc)
+  target_compile_definitions(apex_flags INTERFACE APEX_WITH_PERFETTO)
+endif()
+
 if(HPX_WITH_HPXMP)
   set(ompt_sources apex_ompt.cpp)
   include(APEX_SetupHPXMP)
@@ -317,6 +324,7 @@ set(apex_headers
     task_wrapper.hpp
     tau_listener.hpp
     utils.hpp
+    ${perfetto_headers}
     ${proc_headers}
     ${otf2_headers}
     )
@@ -338,9 +346,6 @@ set(apex_sources
     gzstream.cpp
     handler.cpp
     memory_wrapper.cpp
-    ../perfetto_sdk/perfetto.cc
-    perfetto_listener.cpp
-    perfetto_static.cpp
     policy_handler.cpp
     profile_reducer.cpp
     profiler_listener.cpp
@@ -352,6 +357,7 @@ set(apex_sources
     thread_instance.cpp
     trace_event_listener.cpp
     utils.cpp
+    ${perfetto_sources}
     ${proc_sources}
     ${bfd_sources}
     ${sensor_sources}
@@ -364,7 +370,7 @@ set(apex_sources
 
 include(GNUInstallDirs)
 
-if (WIN32)
+if (WIN32 AND APEX_WITH_PERFETTO)
 # Enable standards-compliant mode when using the Visual Studio compiler.
     if (MSVC)
         SET_SOURCE_FILES_PROPERTIES( perfetto.cc PROPERTIES COMPILE_FLAGS
diff --git a/src/apex/CMakeLists.standalone b/src/apex/CMakeLists.standalone
index 24757bb5..31532b42 100644
--- a/src/apex/CMakeLists.standalone
+++ b/src/apex/CMakeLists.standalone
@@ -79,6 +79,15 @@ if (APEX_WITH_RAJA AND RAJA_FOUND)
     set(RAJA_SOURCE apex_raja.cpp)
 endif (APEX_WITH_RAJA AND RAJA_FOUND)
 
+if(APEX_WITH_PERFETTO)
+    set(perfetto_headers perfetto_listener.hpp)
+    set(perfetto_sources
+        perfetto_listener.cpp perfetto_static.cpp)
+    set(perfetto_target perfetto)
+    add_definitions(-DAPEX_WITH_PERFETTO)
+    include_directories(${APEX_SOURCE_DIR}/src/perfetto_sdk)
+endif()
+
 # Try to keep this in alphabetical order
 SET(all_SOURCE
 apex_preload.cpp
@@ -108,8 +117,7 @@ exhaustive.cpp
 handler.cpp
 memory_wrapper.cpp
 ${OTF2_SOURCE}
-perfetto_listener.cpp
-perfetto_static.cpp
+${perfetto_sources}
 perftool_implementation.cpp
 policy_handler.cpp
 ${PROC_SOURCE}
@@ -129,8 +137,7 @@ ${ZLIB_SOURCE}
 
 add_library (apex ${all_SOURCE})
 add_library (taudummy tau_dummy.cpp)
-add_dependencies (apex taudummy perfetto)
-include_directories(${APEX_SOURCE_DIR}/src/perfetto_sdk)
+add_dependencies (apex taudummy ${perfetto_target})
 
 if(ACTIVEHARMONY_FOUND)
     add_dependencies (apex project_activeharmony)
@@ -168,10 +175,10 @@ endif(APEX_WITH_CUDA)
 # so that we can LD_PRELOAD this library with all requirements met.
 if (NOT BUILD_STATIC_EXECUTABLES)
     if(APPLE)
-        target_link_libraries(apex perfetto ${LIBS})
+        target_link_libraries(apex ${perfetto_target} ${LIBS})
         set_target_properties(apex PROPERTIES LINK_FLAGS "${CMAKE_CURRENT_BINARY_DIR}/libtaudummy.dylib -flat_namespace")
     else(APPLE)
-        target_link_libraries(apex ${LIBS} perfetto taudummy)
+        target_link_libraries(apex ${LIBS} ${perfetto_target} taudummy)
     endif(APPLE)
 endif()
 
diff --git a/src/apex/apex.cpp b/src/apex/apex.cpp
index be9c0757..fff1f0bb 100644
--- a/src/apex/apex.cpp
+++ b/src/apex/apex.cpp
@@ -43,7 +43,9 @@
 #include "tau_listener.hpp"
 #include "profiler_listener.hpp"
 #include "trace_event_listener.hpp"
+#if defined(APEX_WITH_PERFETTO)
 #include "perfetto_listener.hpp"
+#endif
 #if defined(APEX_DEBUG) || defined(APEX_ERROR_HANDLING)
 // #define APEX_DEBUG_disabled
 #include "apex_error_handling.hpp"
@@ -311,10 +313,13 @@ void apex::_initialize()
             listeners.push_back(the_otf2_listener);
         }
 #endif
+#if defined(APEX_WITH_PERFETTO)
         if (apex_options::use_perfetto()) {
             the_perfetto_listener = new perfetto_listener();
             listeners.push_back(the_perfetto_listener);
-        } else if (apex_options::use_trace_event()) {
+        }
+#endif
+        if (apex_options::use_trace_event()) {
             the_trace_event_listener = new trace_event_listener();
             listeners.push_back(the_trace_event_listener);
         }
diff --git a/src/apex/apex_ompt.cpp b/src/apex/apex_ompt.cpp
index 6511a724..8b2f6d2d 100644
--- a/src/apex/apex_ompt.cpp
+++ b/src/apex/apex_ompt.cpp
@@ -23,7 +23,9 @@
 #include "event_listener.hpp"
 #include "async_thread_node.hpp"
 #include "apex.hpp"
+#if defined(APEX_WITH_PERFETTO)
 #include "perfetto_listener.hpp"
+#endif
 #include "trace_event_listener.hpp"
 #ifdef APEX_HAVE_OTF2
 #include "otf2_listener.hpp"
@@ -175,13 +177,16 @@ void stop_async_task(std::shared_ptr<apex::task_wrapper> tt, uint64_t start, uin
     if (correlationId > 0) {
         as_data = Globals::find_data(correlationId);
     }
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         as_data.cat = "ControlFlow";
         as_data.reverse_flow = false;
         tel->on_async_event(node, prof, as_data);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         as_data.cat = "ControlFlow";
@@ -219,11 +224,14 @@ void store_profiler_data(const std::string &name,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         tel->on_async_event(node, prof, as_data);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         tel->on_async_event(node, prof, as_data);
@@ -262,11 +270,14 @@ void store_counter_data(const char * name, const std::string& ctx,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         tel->on_async_metric(node, prof);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         tel->on_async_metric(node, prof);
diff --git a/src/apex/cupti_trace.cpp b/src/apex/cupti_trace.cpp
index 389c1157..8bc82e9a 100644
--- a/src/apex/cupti_trace.cpp
+++ b/src/apex/cupti_trace.cpp
@@ -24,7 +24,9 @@
 #include "profiler.hpp"
 #include "thread_instance.hpp"
 #include "apex_options.hpp"
+#if defined(APEX_WITH_PERFETTO)
 #include "perfetto_listener.hpp"
+#endif
 #include "trace_event_listener.hpp"
 #include "apex_nvml.hpp"
 #ifdef APEX_HAVE_OTF2
@@ -525,13 +527,16 @@ void store_profiler_data(const std::string &name, uint32_t correlationId,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         as_data.cat = category;
         as_data.reverse_flow = reverseFlow;
         tel->on_async_event(node, prof, as_data);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         as_data.cat = category;
@@ -590,11 +595,14 @@ void store_counter_data(const char * name, const std::string& ctx,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         tel->on_async_metric(node, prof);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         tel->on_async_metric(node, prof);
diff --git a/src/apex/hip_trace.cpp b/src/apex/hip_trace.cpp
index 93ab6e8e..2e47e0d3 100644
--- a/src/apex/hip_trace.cpp
+++ b/src/apex/hip_trace.cpp
@@ -40,7 +40,9 @@ using namespace std;
 #include "address_resolution.hpp"
 #endif
 #include "async_thread_node.hpp"
+#if defined(APEX_WITH_PERFETTO)
 #include "perfetto_listener.hpp"
+#endif
 #include "trace_event_listener.hpp"
 #ifdef APEX_HAVE_OTF2
 #include "otf2_listener.hpp"
@@ -888,13 +890,16 @@ void store_profiler_data(const std::string &name, uint32_t correlationId,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         as_data.cat = category;
         as_data.reverse_flow = reverse_flow;
         tel->on_async_event(node, prof, as_data);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         as_data.cat = category;
@@ -935,11 +940,14 @@ void store_counter_data(const char * name, const std::string& ctx,
     // fake out the profiler_listener
     instance->the_profiler_listener->push_profiler_public(prof);
     // Handle tracing, if necessary
+#if defined(APEX_WITH_PERFETTO)
     if (apex::apex_options::use_perfetto()) {
         apex::perfetto_listener * tel =
             (apex::perfetto_listener*)instance->the_perfetto_listener;
         tel->on_async_metric(node, prof);
-    } else if (apex::apex_options::use_trace_event()) {
+    }
+#endif
+    if (apex::apex_options::use_trace_event()) {
         apex::trace_event_listener * tel =
             (apex::trace_event_listener*)instance->the_trace_event_listener;
         tel->on_async_metric(node, prof);
diff --git a/src/perfetto_sdk/perfetto.cc b/src/perfetto_sdk/perfetto.cc
index bc924ccc..525e7cd2 100644
--- a/src/perfetto_sdk/perfetto.cc
+++ b/src/perfetto_sdk/perfetto.cc
@@ -1,5 +1,5 @@
 #ifdef __NVCOMPILER
-#warning "nvc++ takes a long time to compile this file..."
+#warning "nvc++ takes a long time to compile this file...If you don't plan to use native perfetto trace output support, disable it with the CMake option -DAPEX_WITH_PERFETTO=FALSE"
 #endif
 // Copyright (C) 2019 The Android Open Source Project
 //