SWDEV-351980 - Consolidate registration tables in the roctracer

Change-Id: I44cd1cc81cf6a529aed89ee8db1377c0aa67f0dc
trixirt · Sep 9, 2022 · 2673bf5 · 2673bf5
1 parent 57867e4
commit 2673bf5
Show file tree

Hide file tree

Showing 11 changed files with 653 additions and 695 deletions.
diff --git a/inc/roctracer_hip.h b/inc/roctracer_hip.h
@@ -41,7 +41,12 @@ inline static std::ostream& operator<<(std::ostream& out, const char& v) {
 
 #include <roctracer.h>
 
-enum { HIP_OP_ID_DISPATCH = 0, HIP_OP_ID_COPY = 1, HIP_OP_ID_BARRIER = 2, HIP_OP_ID_NUMBER = 3 };
+typedef enum {
+  HIP_OP_ID_DISPATCH = 0,
+  HIP_OP_ID_COPY = 1,
+  HIP_OP_ID_BARRIER = 2,
+  HIP_OP_ID_NUMBER = 3
+} hip_op_id_t;
 
 #ifdef __cplusplus
 extern "C" {

diff --git a/script/hsaap.py b/script/hsaap.py
@@ -332,7 +332,6 @@ def __init__(self, out_h_file, hsa_dir, api_table_h, api_headers, license):
     self.cpp_content += "/* Generated by " + os.path.basename(__file__) + " */\n" + license + "\n\n"
 
     self.cpp_content += '#include <hsa/hsa_api_trace.h>\n'
-    self.cpp_content += '#include \"util/callback_table.h\"\n\n'
     self.cpp_content += '#include <atomic>\n'
     self.cpp_content += 'namespace roctracer::hsa_support::detail {\n'
 
@@ -409,40 +408,52 @@ def gen_arg_struct(self, n, name, call, struct):
   def gen_callbacks(self, n, name, call, struct):
     content = ''
     if n == -1:
-      content += 'static util::CallbackTable<ACTIVITY_DOMAIN_HSA_API, HSA_API_ID_NUMBER> cb_table;\n'
+      content += '/* section: Static declarations */\n'
       content += '\n'
     if call != '-':
       call_id = self.api_id[call];
       ret_type = struct['ret']
       content += 'static ' + ret_type + ' ' + call + '_callback(' + struct['args'] + ') {\n'
-      content += '  hsa_api_data_t api_data{};\n'
+
+      content += '  hsa_trace_data_t trace_data;\n'
+      content += '  bool enabled{false};\n'
+      content += '\n'
+      content += '  if (auto function = report_activity.load(std::memory_order_relaxed); function &&\n'
+      content += '      (enabled =\n'
+      content += '           function(ACTIVITY_DOMAIN_HSA_API, ' + call_id + ', &trace_data) == 0)) {\n'
+      content += '    if (trace_data.phase_enter != nullptr) {\n'
+
       for var in struct['alst']:
         item = struct['astr'][var];
         if re.search(r'char\* ', item):
-          content += '  api_data.args.' + call + '.' + var + ' = ' + '(' + var + ' != NULL) ? strdup(' + var + ')' + ' : NULL;\n'
+          # FIXME: we should not strdup the char* arguments here, as the callback will not outlive the scope of this function. Instead, we
+          # should generate a helper function to capture the content of the arguments similar to hipApiArgsInit for HIP. We also need a
+          # helper to free the memory that is allocated to capture the content.
+          content += '      trace_data.api_data.args.' + call + '.' + var + ' = ' + '(' + var + ' != NULL) ? strdup(' + var + ')' + ' : NULL;\n'
         else:
-          content += '  api_data.args.' + call + '.' + var + ' = ' + var + ';\n'
+          content += '      trace_data.api_data.args.' + call + '.' + var + ' = ' + var + ';\n'
           if call == 'hsa_amd_memory_async_copy_rect' and var == 'range':
-            content += '  api_data.args.' + call + '.' + var + '__val = ' + '*(' + var + ');\n'
-      content += '  auto [ api_callback_fun, api_callback_arg ] = cb_table.Get(' + call_id + ');\n'
-      content += '  if (api_callback_fun) {\n'
-      content += '    api_data.phase = ACTIVITY_API_PHASE_ENTER;\n'
-      content += '    api_data.correlation_id = CorrelationIdPush();\n'
-      content += '    api_callback_fun(ACTIVITY_DOMAIN_HSA_API, ' + call_id + ', &api_data, api_callback_arg);\n'
+            content += '      trace_data.api_data.args.' + call + '.' + var + '__val = ' + '*(' + var + ');\n'
+
+      content += '      trace_data.phase_enter(' + call_id + ', &trace_data);\n'
+      content += '    }\n'
       content += '  }\n'
+      content += '\n'
+
       if ret_type != 'void':
+        # FIXME: we should capture the return value and store it in the api_data
         content += '  ' + ret_type + ' ret ='
       content += '  ' + name + '_saved_before_cb.' + call + '_fn(' + ', '.join(struct['alst']) + ');\n'
-      content += '  if (api_callback_fun) {\n'
-      if ret_type != 'void':
-        content += '    api_data.' + ret_type + '_retval = ret;\n'
-      content += '    api_data.phase = ACTIVITY_API_PHASE_EXIT;\n'
-      content += '    api_callback_fun(ACTIVITY_DOMAIN_HSA_API, ' + call_id + ', &api_data, api_callback_arg);\n'
-      content += '    CorrelationIdPop();\n'
-      content += '  }\n'
+
+      content += '\n'
+      content += '  if (enabled && trace_data.phase_exit != nullptr)\n'
+      content += '    trace_data.phase_exit(' + call_id + ', &trace_data);\n'
+
       if ret_type != 'void':
+        content += '\n'
         content += '  return ret;\n'
       content += '}\n'
+
     return content
 
   # generate API intercepting code