Move base event into device

intel · Sep 25, 2023 · c2a3cf7 · c2a3cf7
1 parent 0de2831
commit c2a3cf7
Show file tree

Hide file tree

Showing 4 changed files with 31 additions and 42 deletions.
diff --git a/sycl/plugins/unified_runtime/ur/adapters/hip/device.cpp b/sycl/plugins/unified_runtime/ur/adapters/hip/device.cpp
@@ -18,6 +18,18 @@ int getAttribute(ur_device_handle_t Device, hipDeviceAttribute_t Attribute) {
   return Value;
 }
 
+uint64_t ur_device_handle_t_::getElapsedTime(hipEvent_t Event) const {
+  float MilliSeconds = 0.0f;
+
+  // hipEventSynchronize waits till the event is ready for call to
+  // hipEventElapsedTime.
+  UR_CHECK_ERROR(hipEventSynchronize(EvBase));
+  UR_CHECK_ERROR(hipEventSynchronize(Event));
+  UR_CHECK_ERROR(hipEventElapsedTime(&MilliSeconds, EvBase, Event));
+
+  return static_cast<uint64_t>(MilliSeconds * 1.0e6);
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
                                                     ur_device_info_t propName,
                                                     size_t propSize,

diff --git a/sycl/plugins/unified_runtime/ur/adapters/hip/device.hpp b/sycl/plugins/unified_runtime/ur/adapters/hip/device.hpp
@@ -23,13 +23,15 @@ struct ur_device_handle_t_ {
   std::atomic_uint32_t RefCount;
   ur_platform_handle_t Platform;
   hipCtx_t HIPContext;
+  hipEvent_t EvBase;
   size_t DeviceIndex; // The index of the device in the UR context
 
 public:
   ur_device_handle_t_(native_type HipDevice, hipCtx_t Context,
-                      ur_platform_handle_t Platform, size_t DeviceIndex)
+                      ur_platform_handle_t Platform, size_t DeviceIndex,
+                      hipEvent_t EvBase)
       : HIPDevice(HipDevice), RefCount{1}, Platform(Platform),
-        HIPContext(Context), DeviceIndex(DeviceIndex) {}
+        HIPContext(Context), EvBase(EvBase), DeviceIndex(DeviceIndex) {}
 
   ~ur_device_handle_t_() {
     UR_CHECK_ERROR(hipDevicePrimaryCtxRelease(HIPDevice));
@@ -41,7 +43,9 @@ struct ur_device_handle_t_ {
 
   ur_platform_handle_t getPlatform() const noexcept { return Platform; };
 
-  hipCtx_t getNativeContext() { return HIPContext; };
+  uint64_t getElapsedTime(hipEvent_t) const;
+
+  hipCtx_t getNativeContext() const { return HIPContext; };
 
   // Returns the index of the device in question relative to the other devices
   // in the platform

diff --git a/sycl/plugins/unified_runtime/ur/adapters/hip/event.cpp b/sycl/plugins/unified_runtime/ur/adapters/hip/event.cpp
@@ -88,44 +88,13 @@ bool ur_event_handle_t_::isCompleted() const noexcept {
 }
 
 uint64_t ur_event_handle_t_::getQueuedTime() const {
-  float MilliSeconds = 0.0f;
   assert(isStarted());
-
-  // hipEventSynchronize waits till the event is ready for call to
-  // hipEventElapsedTime.
-  UR_CHECK_ERROR(hipEventSynchronize(EvStart));
-  UR_CHECK_ERROR(hipEventSynchronize(EvEnd));
-
-  UR_CHECK_ERROR(hipEventElapsedTime(&MilliSeconds, EvStart, EvEnd));
-  return static_cast<uint64_t>(MilliSeconds * 1.0e6);
-}
-
-uint64_t ur_event_handle_t_::getStartTime() const {
-  float MiliSeconds = 0.0f;
-  assert(isStarted());
-
-  // hipEventSynchronize waits till the event is ready for call to
-  // hipEventElapsedTime.
-  UR_CHECK_ERROR(hipEventSynchronize(ur_platform_handle_t_::EvBase));
-  UR_CHECK_ERROR(hipEventSynchronize(EvStart));
-
-  UR_CHECK_ERROR(hipEventElapsedTime(&MiliSeconds,
-                                     ur_platform_handle_t_::EvBase, EvStart));
-  return static_cast<uint64_t>(MiliSeconds * 1.0e6);
+  return Queue->getDevice()->getElapsedTime(EvStart);
 }
 
 uint64_t ur_event_handle_t_::getEndTime() const {
-  float MiliSeconds = 0.0f;
   assert(isStarted() && isRecorded());
-
-  // hipEventSynchronize waits till the event is ready for call to
-  // hipEventElapsedTime.
-  UR_CHECK_ERROR(hipEventSynchronize(ur_platform_handle_t_::EvBase));
-  UR_CHECK_ERROR(hipEventSynchronize(EvEnd));
-
-  UR_CHECK_ERROR(
-      hipEventElapsedTime(&MiliSeconds, ur_platform_handle_t_::EvBase, EvEnd));
-  return static_cast<uint64_t>(MiliSeconds * 1.0e6);
+  return Queue->getDevice()->getElapsedTime(EvEnd);
 }
 
 ur_result_t ur_event_handle_t_::record() {
@@ -249,11 +218,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
   UrReturnHelper ReturnValue(propValueSize, pPropValue, pPropValueSizeRet);
   switch (propName) {
   case UR_PROFILING_INFO_COMMAND_QUEUED:
-  case UR_PROFILING_INFO_COMMAND_SUBMIT:
     // Note: No user for this case
-    return ReturnValue(static_cast<uint64_t>(hEvent->getQueuedTime()));
+  case UR_PROFILING_INFO_COMMAND_SUBMIT:
   case UR_PROFILING_INFO_COMMAND_START:
-    return ReturnValue(static_cast<uint64_t>(hEvent->getStartTime()));
+    return ReturnValue(static_cast<uint64_t>(hEvent->getQueuedTime()));
   case UR_PROFILING_INFO_COMMAND_END:
     return ReturnValue(static_cast<uint64_t>(hEvent->getEndTime()));
   default:

diff --git a/sycl/plugins/unified_runtime/ur/adapters/hip/platform.cpp b/sycl/plugins/unified_runtime/ur/adapters/hip/platform.cpp
@@ -74,12 +74,17 @@ urPlatformGet(ur_adapter_handle_t *, uint32_t, uint32_t NumEntries,
           try {
             for (int i = 0; i < NumDevices; ++i) {
               hipDevice_t Device;
-              Err = UR_CHECK_ERROR(hipDeviceGet(&Device, i));
+              UR_CHECK_ERROR(hipDeviceGet(&Device, i));
               hipCtx_t Context;
-              Err = UR_CHECK_ERROR(hipDevicePrimaryCtxRetain(&Context, Device));
+              UR_CHECK_ERROR(hipDevicePrimaryCtxRetain(&Context, Device));
+              UR_CHECK_ERROR(hipCtxSetCurrent(Context));
+              hipEvent_t EvBase;
+              UR_CHECK_ERROR(hipEventCreateWithFlags(&EvBase, hipEventDefault));
+              UR_CHECK_ERROR(hipEventRecord(EvBase, 0));
+
               Platform.Devices.emplace_back(
                   std::make_unique<ur_device_handle_t_>(Device, Context,
-                                                        &Platform, i));
+                                                        &Platform, i, EvBase));
             }
           } catch (const std::bad_alloc &) {
             Platform.Devices.clear();