several fixes; adding direct loading of alprofile library

test/util/hsa_rsrc_factory.cpp

@@ -24,6 +24,7 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include "util/hsa_rsrc_factory.h"
 
+#include <dlfcn.h>
 #include <hsa.h>
 #include <hsa_ext_finalize.h>
 #include <stdint.h>
@@ -116,6 +117,12 @@ HsaRsrcFactory::HsaRsrcFactory() {
   // Discover the set of Gpu devices available on the platform
   status = hsa_iterate_agents(GetHsaAgentsCallback, this);
   CHECK_STATUS("Error Calling hsa_iterate_agents", status);
+
+  // Get AqlProfile API table
+  aqlprofile_api_ = {0};
+  status = hsa_system_get_extension_table(HSA_EXTENSION_AMD_AQLPROFILE, 1, 0, &aqlprofile_api_);
+  if (status != HSA_STATUS_SUCCESS) status = LoadAqlProfileLib(&aqlprofile_api_);
+  CHECK_STATUS("aqlprofile API table load failed", status);
 }
 
 // Destructor of the class
@@ -124,6 +131,39 @@ HsaRsrcFactory::~HsaRsrcFactory() {
   CHECK_STATUS("Error in hsa_shut_down", status);
 }
 
+hsa_status_t HsaRsrcFactory::LoadAqlProfileLib(aqlprofile_pfn_t* api) {
+    void* handle = dlopen(kAqlProfileLib, RTLD_NOW);
+    if (handle == NULL) {
+      fprintf(stderr, "Loading '%s' failed, %s\n", kAqlProfileLib, dlerror());
+      return HSA_STATUS_ERROR;
+    }
+    dlerror(); /* Clear any existing error */
+
+    api->hsa_ven_amd_aqlprofile_error_string =
+      (decltype(::hsa_ven_amd_aqlprofile_error_string)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_error_string");
+    api->hsa_ven_amd_aqlprofile_validate_event =
+      (decltype(::hsa_ven_amd_aqlprofile_validate_event)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_validate_event");
+    api->hsa_ven_amd_aqlprofile_start =
+      (decltype(::hsa_ven_amd_aqlprofile_start)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_start");
+    api->hsa_ven_amd_aqlprofile_stop =
+      (decltype(::hsa_ven_amd_aqlprofile_stop)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_stop");
+    api->hsa_ven_amd_aqlprofile_legacy_get_pm4 =
+      (decltype(::hsa_ven_amd_aqlprofile_legacy_get_pm4)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_legacy_get_pm4");
+    api->hsa_ven_amd_aqlprofile_get_info =
+      (decltype(::hsa_ven_amd_aqlprofile_get_info)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_get_info");
+    api->hsa_ven_amd_aqlprofile_iterate_data =
+      (decltype(::hsa_ven_amd_aqlprofile_iterate_data)*)
+        dlsym(handle, "hsa_ven_amd_aqlprofile_iterate_data");
+
+  return HSA_STATUS_SUCCESS;
+}
+
 // Get the count of Hsa Gpu Agents available on the platform
 //
 // @return uint32_t Number of Gpu agents on platform

test/util/hsa_rsrc_factory.h

@@ -27,6 +27,7 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include <hsa.h>
 #include <hsa_ext_finalize.h>
+#include <hsa_ven_amd_aqlprofile.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -54,7 +55,7 @@ static const unsigned MEM_PAGE_MASK = MEM_PAGE_BYTES - 1;
 
 // Encapsulates information about a Hsa Agent such as its
 // handle, name, max queue size, max wavefront size, etc.
-typedef struct {
+struct AgentInfo {
   // Handle of Agent
   hsa_agent_t dev_id;
 
@@ -78,8 +79,7 @@ typedef struct {
 
   // Memory region supporting kernel arguments
   hsa_region_t kernarg_region;
-
-} AgentInfo;
+};
 
 class HsaRsrcFactory {
  public:
@@ -207,7 +207,14 @@ class HsaRsrcFactory {
   // Print the various fields of Hsa Gpu Agents
   bool PrintGpuAgents(const std::string& header);
 
+  // Return AqlProfile API table
+  typedef hsa_ven_amd_aqlprofile_1_00_pfn_t aqlprofile_pfn_t;
+  const aqlprofile_pfn_t* AqlProfileApi() const { return &aqlprofile_api_; }
+
  private:
+  // Load AQL profile HSA extension library directly
+  static hsa_status_t LoadAqlProfileLib(aqlprofile_pfn_t* api);
+
   // Constructor of the class. Will initialize the Hsa Runtime and
   // query the system topology to get the list of Cpu and Gpu devices
   HsaRsrcFactory();
@@ -229,6 +236,9 @@ class HsaRsrcFactory {
 
   // Used to maintain a list of Hsa Cpu Agent Info
   std::vector<AgentInfo*> cpu_list_;
+
+  // AqlProfile API table
+  aqlprofile_pfn_t aqlprofile_api_;
 };
 
 #endif  // TEST_UTIL_HSA_RSRC_FACTORY_H_

test/util/xml.h

@@ -27,6 +27,33 @@ class Xml {
 
   enum { DECL_STATE, BODY_STATE };
 
+  static Xml* Create(const char* file_name) {
+    Xml* xml = new Xml(file_name);
+    if (xml->fd_ == -1) {
+      delete xml;
+      xml = NULL;
+    }
+    return xml;
+  }
+
+  static void Destroy(Xml *xml) { delete xml; }
+
+  std::vector<level_t*> GetNodes(std::string global_tag) { return map_[global_tag]; }
+
+  void Print() const {
+    for (auto& elem : map_) {
+      for (auto node : elem.second) {
+        if (node->opts.size()) {
+          std::cout << elem.first << ":" << std::endl;
+          for (auto& opt : node->opts) {
+            std::cout << "  " << opt.first << " = " << opt.second << std::endl;
+          }
+        }
+      }
+    }
+  }
+
+ private:
   Xml(const char* file_name)
       : file_name_(file_name),
         file_line_(0),
@@ -39,7 +66,7 @@ class Xml {
 
     fd_ = open(file_name, O_RDONLY);
     if (fd_ == -1) {
-      std::cout << "XML file not found: " << file_name << std::endl;
+      perror("open XML file");
       return;
     }
 
@@ -117,22 +144,8 @@ class Xml {
     }
   }
 
-  std::vector<level_t*> GetNodes(std::string global_tag) { return map_[global_tag]; }
+  ~Xml() {}
 
-  void Print() const {
-    for (auto& elem : map_) {
-      for (auto node : elem.second) {
-        if (node->opts.size()) {
-          std::cout << elem.first << ":" << std::endl;
-          for (auto& opt : node->opts) {
-            std::cout << "  " << opt.first << " = " << opt.second << std::endl;
-          }
-        }
-      }
-    }
-  }
-
- private:
   bool LineEndCheck() {
     bool found = false;
     if (buffer_[index_] == '\n') {