adding allocate/device/memcopy/submit callbacks

Change-Id: Ie83ae3cd57cdf5038e83db70891b103439c37d55

test/run.sh

@@ -52,18 +52,22 @@ eval_test() {
   test_number=$((test_number + 1))
 }
 
-# enable tools load failure reporting
-export HSA_TOOLS_REPORT_LOAD_FAILURE=1
 # paths to ROC profiler and oher libraries
 export LD_LIBRARY_PATH=$PWD
+
+# enable tools load failure reporting
+export HSA_TOOLS_REPORT_LOAD_FAILURE=1
 # enable error messages logging to '/tmp/rocprofiler_log.txt'
 export ROCPROFILER_LOG=1
-# ROC profiler metrics config file
+# enable error messages logging to '/tmp/aql_profile_log.txt'
+export HSA_VEN_AMD_AQLPROFILE_LOG=1
+# test trace
+export ROC_TEST_TRACE=1
+
+# Disabple profiler own proxy queue
 unset ROCP_PROXY_QUEUE
 # ROC profiler metrics config file
 export ROCP_METRICS=metrics.xml
-# test trace
-export ROC_TEST_TRACE=1
 
 ## C test
 eval_test "C test" ./test/c_test
@@ -93,6 +97,9 @@ export ROCP_TIMESTAMP_ON=1
 # and SQTT trace files 'thread_trace.se<n>.out'
 export ROCP_OUTPUT_DIR=./RESULTS
 
+# enable HSA intercepting
+export ROCP_HSA_INTERC=1
+
 if [ ! -e $ROCP_TOOL_LIB ] ; then
   export ROCP_TOOL_LIB=test/libtool.so
 fi

test/tool/tool.cpp

@@ -826,6 +826,37 @@ static inline void check_env_var(const char* var_name, uint64_t& val) {
   if (str != NULL ) val = atoll(str);
 }
 
+// HSA intercepting routines
+
+// HSA unified callback function
+hsa_status_t hsa_unified_callback(
+  rocprofiler_hsa_cb_id_t id,
+  const rocprofiler_hsa_callback_data_t* data,
+  void* arg)
+{
+  printf("hsa_unified_callback(%d, %p, %p\n", (int)id, data, arg);
+
+  if (id == ROCPROFILER_HSA_CB_ID_SUBMIT) {
+    if (data->submit.kernel_code != NULL) {
+      printf("  submit:\n");
+      printf("    kernel_code_entry_byte_offset %lx\n", data->submit.kernel_code->kernel_code_entry_byte_offset);
+      printf("    kernel_code_prefetch_byte_offset %lx\n", data->submit.kernel_code->kernel_code_prefetch_byte_offset);
+      printf("    kernel_code_prefetch_byte_size %lx\n", data->submit.kernel_code->kernel_code_prefetch_byte_size);
+    }
+  }
+
+  fflush(stdout);
+  return HSA_STATUS_SUCCESS;
+}
+
+// HSA callbacks structure
+rocprofiler_hsa_callbacks_t hsa_callbacks {
+  hsa_unified_callback,
+  hsa_unified_callback,
+  hsa_unified_callback,
+  hsa_unified_callback
+};
+
 // Tool constructor
 extern "C" PUBLIC_API void OnLoadToolProp(rocprofiler_settings_t* settings)
 {
@@ -908,6 +939,8 @@ extern "C" PUBLIC_API void OnLoadToolProp(rocprofiler_settings_t* settings)
   check_env_var("ROCP_OBJ_TRACKING", settings->code_obj_tracking);
   // Set memcopies tracking
   check_env_var("ROCP_MCOPY_TRACKING", settings->memcopy_tracking);
+  // Set HSA intercepting
+  check_env_var("ROCP_HSA_INTERC", settings->hsa_intercepting);
 
   is_trace_local = settings->trace_local;