From f6d8fa548760b861ed9d9cdcb5eaa11d85cf4988 Mon Sep 17 00:00:00 2001
From: Evgeny <evgeny.shcherbakov@amd.com>
Date: Sun, 11 Nov 2018 23:44:16 -0600
Subject: [PATCH] specialized spawner object

[ROCm/clr commit: 1e49cff87afe7749b763a517cc0e098567744f01]
---
 .../include/hip/hcc_detail/hip_prof_api.h     | 17 ++++++++++-------
 .../include/hip/hcc_detail/hip_prof_str.h     | 19 +++++++++++++++----
 2 files changed, 25 insertions(+), 11 deletions(-)

diff --git a/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_api.h b/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_api.h
index 8589bc0ee6..dd3c3efec9 100644
--- a/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_api.h
+++ b/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_api.h
@@ -127,21 +127,19 @@ typedef activity_sync_callback_t hip_act_callback_t;
 
 // HIP API callbacks spawner object macro
 #define HIP_CB_SPAWNER_OBJECT(CB_ID) \
-  if (HIP_API_ID_##CB_ID < HIP_API_ID_NUMBER) { \
-    hip_api_data_t api_data{}; \
-    INIT_CB_ARGS_DATA(CB_ID, api_data); \
-    api_callbacks_spawner_t __api_tracer(HIP_API_ID_##CB_ID, api_data); \
-  }
+  hip_api_data_t api_data{}; \
+  INIT_CB_ARGS_DATA(CB_ID, api_data); \
+  api_callbacks_spawner_t<HIP_API_ID_##CB_ID> __api_tracer(HIP_API_ID_##CB_ID, api_data);
 
 typedef api_callbacks_table_templ<hip_api_record_t,
                                   hip_api_callback_t,
                                   hip_act_callback_t> api_callbacks_table_t;
 extern api_callbacks_table_t callbacks_table;
 
+template <int cid_>
 class api_callbacks_spawner_t {
  public:
   api_callbacks_spawner_t(const hip_api_id_t& cid, hip_api_data_t& api_data) :
-    cid_(cid),
     api_data_(api_data),
     record_({})
   {
@@ -174,7 +172,6 @@ class api_callbacks_spawner_t {
     return callbacks_table.entry(id);
   }
 
-  const hip_api_id_t cid_;
   hip_api_data_t& api_data_;
   hip_api_record_t record_;
 
@@ -184,6 +181,12 @@ class api_callbacks_spawner_t {
   void* arg;
 };
 
+template <>
+class api_callbacks_spawner_t<HIP_API_ID_NUMBER> {
+ public:
+  api_callbacks_spawner_t(const hip_api_id_t& cid, hip_api_data_t& api_data) {}
+};
+
 #else
 
 #define HIP_CB_SPAWNER_OBJECT(x) do {} while(0)
diff --git a/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_str.h b/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_str.h
index 40cd865c2e..512e8a3188 100644
--- a/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_str.h
+++ b/projects/clr/hipamd/include/hip/hcc_detail/hip_prof_str.h
@@ -5,7 +5,6 @@
 #include <string>
 
 // Dummy API callbacks definition
-#define INIT_hipHccModuleLaunchKernel_CB_ARGS_DATA(cb_data) {};
 #define INIT_hipHccGetAccelerator_CB_ARGS_DATA(cb_data) {};
 #define INIT_hipHccGetAcceleratorView_CB_ARGS_DATA(cb_data) {};
 #define INIT_hipDeviceCanAccessPeer2_CB_ARGS_DATA(cb_data) {};
@@ -177,10 +176,10 @@ enum hip_api_id_t {
   HIP_API_ID_hipModuleGetFunction = 135,
   HIP_API_ID_hipGetDevice = 136,
   HIP_API_ID_hipGetDeviceCount = 137,
-  HIP_API_ID_NUMBER = 138,
-  HIP_API_ID_ANY = 139,
+  HIP_API_ID_hipHccModuleLaunchKernel = 138,
+  HIP_API_ID_NUMBER = 139,
+  HIP_API_ID_ANY = 140,
 
-  HIP_API_ID_hipHccModuleLaunchKernel = HIP_API_ID_NUMBER,
   HIP_API_ID_hipHccGetAccelerator = HIP_API_ID_NUMBER,
   HIP_API_ID_hipHccGetAcceleratorView = HIP_API_ID_NUMBER,
   HIP_API_ID_hipDeviceCanAccessPeer2 = HIP_API_ID_NUMBER,
@@ -340,6 +339,7 @@ static const char* hip_api_name(const uint32_t& id) {
     case HIP_API_ID_hipCtxEnablePeerAccess: return "hipCtxEnablePeerAccess";
     case HIP_API_ID_hipMemcpyDtoHAsync: return "hipMemcpyDtoHAsync";
     case HIP_API_ID_hipModuleLaunchKernel: return "hipModuleLaunchKernel";
+    case HIP_API_ID_hipHccModuleLaunchKernel: return "hipHccModuleLaunchKernel";
     case HIP_API_ID_hipModuleGetTexRef: return "hipModuleGetTexRef";
     case HIP_API_ID_hipRemoveActivityCallback: return "hipRemoveActivityCallback";
     case HIP_API_ID_hipDeviceGetLimit: return "hipDeviceGetLimit";
@@ -907,6 +907,9 @@ struct hip_api_data_t {
       void** kernelParams;
       void** extra;
     } hipModuleLaunchKernel;
+    struct {
+      hipFunction_t f;
+    } hipHccModuleLaunchKernel;
     struct {
       textureReference** texRef;
       hipModule_t hmod;
@@ -1526,6 +1529,9 @@ struct hip_api_data_t {
   cb_data.args.hipModuleLaunchKernel.kernelParams = (void**)kernelParams; \
   cb_data.args.hipModuleLaunchKernel.extra = (void**)extra; \
 };
+#define INIT_hipHccModuleLaunchKernel_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipModuleLaunchKernel.f = (hipFunction_t)f; \
+};
 #define INIT_hipModuleGetTexRef_CB_ARGS_DATA(cb_data) { \
   cb_data.args.hipModuleGetTexRef.texRef = (textureReference**)texRef; \
   cb_data.args.hipModuleGetTexRef.hmod = (hipModule_t)hmod; \
@@ -2396,6 +2402,11 @@ const char* hipApiString(hip_api_id_t id, const hip_api_data_t* data) {
           << " extra=" << data->args.hipModuleLaunchKernel.extra
           << ")";
     break;
+    case HIP_API_ID_hipHccModuleLaunchKernel:
+      oss << "hipHccModuleLaunchKernel("
+          << " f=" << data->args.hipHccModuleLaunchKernel.f << ","
+          << ")";
+    break;
     case HIP_API_ID_hipModuleGetTexRef:
       oss << "hipModuleGetTexRef("
           << " texRef=" << data->args.hipModuleGetTexRef.texRef << ","