Make profiling info optional and update tests

This patch turns all of the values returned by urEventGetProfilingInfo
to be optional and updates adapters to handle this by returning the
appropriate enum when it is not supported.

The tests have also been updated, to ensure that returning a counter of
"0" or values equal to the previous profiling event is no longer
considered a failure.

include/ur_api.h

@@ -7089,21 +7089,21 @@ typedef enum ur_event_info_t {
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Profiling query information type
 typedef enum ur_profiling_info_t {
-  /// [uint64_t] A 64-bit value of current device counter in nanoseconds
-  /// when the event is enqueued
+  /// [uint64_t][optional-query] A 64-bit value of current device counter in
+  /// nanoseconds when the event is enqueued
   UR_PROFILING_INFO_COMMAND_QUEUED = 0,
-  /// [uint64_t] A 64-bit value of current device counter in nanoseconds
-  /// when the event is submitted
+  /// [uint64_t][optional-query] A 64-bit value of current device counter in
+  /// nanoseconds when the event is submitted
   UR_PROFILING_INFO_COMMAND_SUBMIT = 1,
-  /// [uint64_t] A 64-bit value of current device counter in nanoseconds
-  /// when the event starts execution
+  /// [uint64_t][optional-query] A 64-bit value of current device counter in
+  /// nanoseconds when the event starts execution
   UR_PROFILING_INFO_COMMAND_START = 2,
-  /// [uint64_t] A 64-bit value of current device counter in nanoseconds
-  /// when the event has finished execution
+  /// [uint64_t][optional-query] A 64-bit value of current device counter in
+  /// nanoseconds when the event has finished execution
   UR_PROFILING_INFO_COMMAND_END = 3,
-  /// [uint64_t] A 64-bit value of current device counter in nanoseconds
-  /// when the event and any child events enqueued by this event on the
-  /// device have finished execution
+  /// [uint64_t][optional-query] A 64-bit value of current device counter in
+  /// nanoseconds when the event and any child events enqueued by this event
+  /// on the device have finished execution
   UR_PROFILING_INFO_COMMAND_COMPLETE = 4,
   /// @cond
   UR_PROFILING_INFO_FORCE_UINT32 = 0x7fffffff
@@ -7177,6 +7177,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetInfo(
 ///     - ::UR_RESULT_ERROR_INVALID_EVENT
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION
+///         + If `propName` is not supported by the adapter.
 UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
     /// [in] handle of the event object
     ur_event_handle_t hEvent,

scripts/core/event.yml

@@ -121,15 +121,15 @@ name: $x_profiling_info_t
 typed_etors: True
 etors:
     - name: COMMAND_QUEUED
-      desc: "[uint64_t] A 64-bit value of current device counter in nanoseconds when the event is enqueued"
+      desc: "[uint64_t][optional-query] A 64-bit value of current device counter in nanoseconds when the event is enqueued"
     - name: COMMAND_SUBMIT
-      desc: "[uint64_t] A 64-bit value of current device counter in nanoseconds when the event is submitted"
+      desc: "[uint64_t][optional-query] A 64-bit value of current device counter in nanoseconds when the event is submitted"
     - name: COMMAND_START
-      desc: "[uint64_t] A 64-bit value of current device counter in nanoseconds when the event starts execution"
+      desc: "[uint64_t][optional-query] A 64-bit value of current device counter in nanoseconds when the event starts execution"
     - name: COMMAND_END
-      desc: "[uint64_t] A 64-bit value of current device counter in nanoseconds when the event has finished execution"
+      desc: "[uint64_t][optional-query] A 64-bit value of current device counter in nanoseconds when the event has finished execution"
     - name: COMMAND_COMPLETE
-      desc: "[uint64_t] A 64-bit value of current device counter in nanoseconds when the event and any child events enqueued by this event on the device have finished execution"
+      desc: "[uint64_t][optional-query] A 64-bit value of current device counter in nanoseconds when the event and any child events enqueued by this event on the device have finished execution"
 --- #--------------------------------------------------------------------------
 type: function
 desc: "Get event object information"
@@ -198,6 +198,8 @@ returns:
     - $X_RESULT_ERROR_INVALID_EVENT
     - $X_RESULT_ERROR_OUT_OF_RESOURCES
     - $X_RESULT_ERROR_OUT_OF_HOST_MEMORY
+    - $X_RESULT_ERROR_UNSUPPORTED_ENUMERATION:
+      - "If `propName` is not supported by the adapter."
 --- #--------------------------------------------------------------------------
 type: function
 desc: "Wait for a list of events to finish."

source/adapters/cuda/event.cpp

@@ -213,6 +213,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
     return ReturnValue(static_cast<uint64_t>(hEvent->getStartTime()));
   case UR_PROFILING_INFO_COMMAND_END:
     return ReturnValue(static_cast<uint64_t>(hEvent->getEndTime()));
+  case UR_PROFILING_INFO_COMMAND_COMPLETE:
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   default:
     break;
   }

source/adapters/hip/event.cpp

@@ -234,6 +234,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
     return ReturnValue(static_cast<uint64_t>(hEvent->getStartTime()));
   case UR_PROFILING_INFO_COMMAND_END:
     return ReturnValue(static_cast<uint64_t>(hEvent->getEndTime()));
+  case UR_PROFILING_INFO_COMMAND_COMPLETE:
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   default:
     break;
   }

source/adapters/level_zero/event.cpp

@@ -609,6 +609,10 @@ ur_result_t urEventGetProfilingInfo(
 
       return ReturnValue(ContextEndTime);
     }
+    case UR_PROFILING_INFO_COMMAND_COMPLETE:
+      logger::error("urEventGetProfilingInfo: "
+                    "UR_PROFILING_INFO_COMMAND_COMPLETE not supported");
+      return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
     default:
       logger::error("urEventGetProfilingInfo: not supported ParamName");
       return UR_RESULT_ERROR_INVALID_VALUE;
@@ -672,6 +676,10 @@ ur_result_t urEventGetProfilingInfo(
         ContextEndTime *= ZeTimerResolution;
         return ReturnValue(ContextEndTime);
       }
+      case UR_PROFILING_INFO_COMMAND_COMPLETE:
+        logger::error("urEventGetProfilingInfo: "
+                      "UR_PROFILING_INFO_COMMAND_COMPLETE not supported");
+        return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
       default:
         logger::error("urEventGetProfilingInfo: not supported ParamName");
         return UR_RESULT_ERROR_INVALID_VALUE;
@@ -715,6 +723,10 @@ ur_result_t urEventGetProfilingInfo(
     // enqueue.
     //
     return ReturnValue(uint64_t{0});
+  case UR_PROFILING_INFO_COMMAND_COMPLETE:
+    logger::error("urEventGetProfilingInfo: UR_PROFILING_INFO_COMMAND_COMPLETE "
+                  "not supported");
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   default:
     logger::error("urEventGetProfilingInfo: not supported ParamName");
     return UR_RESULT_ERROR_INVALID_VALUE;

source/adapters/native_cpu/event.cpp

@@ -52,6 +52,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
   case UR_PROFILING_INFO_COMMAND_QUEUED:
   case UR_PROFILING_INFO_COMMAND_SUBMIT:
   case UR_PROFILING_INFO_COMMAND_COMPLETE:
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   default:
     break;
   }

source/loader/ur_libapi.cpp

@@ -4725,6 +4725,8 @@ ur_result_t UR_APICALL urEventGetInfo(
 ///     - ::UR_RESULT_ERROR_INVALID_EVENT
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION
+///         + If `propName` is not supported by the adapter.
 ur_result_t UR_APICALL urEventGetProfilingInfo(
     /// [in] handle of the event object
     ur_event_handle_t hEvent,

source/ur_api.cpp

@@ -4130,6 +4130,8 @@ ur_result_t UR_APICALL urEventGetInfo(
 ///     - ::UR_RESULT_ERROR_INVALID_EVENT
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION
+///         + If `propName` is not supported by the adapter.
 ur_result_t UR_APICALL urEventGetProfilingInfo(
     /// [in] handle of the event object
     ur_event_handle_t hEvent,

test/conformance/event/urEventGetProfilingInfo.cpp

@@ -13,35 +13,19 @@ using urEventGetProfilingInfoTest =
 TEST_P(urEventGetProfilingInfoTest, Success) {
 
   ur_profiling_info_t info_type = getParam();
-
-  if (info_type == UR_PROFILING_INFO_COMMAND_COMPLETE) {
-    UUR_KNOWN_FAILURE_ON(uur::CUDA{}, uur::HIP{}, uur::LevelZero{},
-                         uur::NativeCPU{});
-  }
-
-  if (info_type == UR_PROFILING_INFO_COMMAND_QUEUED) {
-    UUR_KNOWN_FAILURE_ON(uur::LevelZero{}, uur::LevelZeroV2{},
-                         uur::NativeCPU{});
-  }
-
-  if (info_type == UR_PROFILING_INFO_COMMAND_SUBMIT) {
-    UUR_KNOWN_FAILURE_ON(uur::LevelZero{}, uur::LevelZeroV2{},
-                         uur::NativeCPU{});
-  }
-
   size_t size;
   ASSERT_SUCCESS_OR_OPTIONAL_QUERY(
       urEventGetProfilingInfo(event, info_type, 0, nullptr, &size), info_type);
   ASSERT_EQ(size, 8);
 
-  std::vector<uint8_t> data(size);
+  uint64_t time = 0x12341234;
   ASSERT_SUCCESS(
-      urEventGetProfilingInfo(event, info_type, size, data.data(), nullptr));
+      urEventGetProfilingInfo(event, info_type, size, &time, nullptr));
 
-  if (sizeof(size_t) == size) {
-    auto returned_value = reinterpret_cast<size_t *>(data.data());
-    ASSERT_NE(*returned_value, 0);
-  }
+  // Note: In theory it's possible for this test to run when the counter happens
+  // to equal this value, but I assume that's so unlikely as to not worry about
+  // it
+  ASSERT_NE(time, 0x12341234);
 }
 
 UUR_DEVICE_TEST_SUITE_P(urEventGetProfilingInfoTest,
@@ -55,48 +39,41 @@ UUR_DEVICE_TEST_SUITE_P(urEventGetProfilingInfoTest,
 using urEventGetProfilingInfoWithTimingComparisonTest = uur::event::urEventTest;
 
 TEST_P(urEventGetProfilingInfoWithTimingComparisonTest, Success) {
-  UUR_KNOWN_FAILURE_ON(uur::CUDA{}, uur::HIP{}, uur::LevelZero{},
-                       uur::LevelZeroV2{}, uur::NativeCPU{});
-
-  uint8_t size = 8;
-
-  std::vector<uint8_t> queued_data(size);
-  ASSERT_SUCCESS(urEventGetProfilingInfo(event,
-                                         UR_PROFILING_INFO_COMMAND_QUEUED, size,
-                                         queued_data.data(), nullptr));
-  auto queued_timing = reinterpret_cast<size_t *>(queued_data.data());
-  ASSERT_NE(*queued_timing, 0);
-
-  std::vector<uint8_t> submit_data(size);
-  ASSERT_SUCCESS(urEventGetProfilingInfo(event,
-                                         UR_PROFILING_INFO_COMMAND_SUBMIT, size,
-                                         submit_data.data(), nullptr));
-  auto submit_timing = reinterpret_cast<size_t *>(submit_data.data());
-  ASSERT_NE(*submit_timing, 0);
-
-  std::vector<uint8_t> start_data(size);
-  ASSERT_SUCCESS(urEventGetProfilingInfo(event, UR_PROFILING_INFO_COMMAND_START,
-                                         size, start_data.data(), nullptr));
-  auto start_timing = reinterpret_cast<size_t *>(start_data.data());
-  ASSERT_NE(*start_timing, 0);
-
-  std::vector<uint8_t> end_data(size);
-  ASSERT_SUCCESS(urEventGetProfilingInfo(event, UR_PROFILING_INFO_COMMAND_END,
-                                         size, end_data.data(), nullptr));
-  auto end_timing = reinterpret_cast<size_t *>(end_data.data());
-  ASSERT_NE(*end_timing, 0);
-
-  std::vector<uint8_t> complete_data(size);
-  ASSERT_SUCCESS(urEventGetProfilingInfo(event,
-                                         UR_PROFILING_INFO_COMMAND_COMPLETE,
-                                         size, complete_data.data(), nullptr));
-  auto complete_timing = reinterpret_cast<size_t *>(complete_data.data());
-  ASSERT_NE(*complete_timing, 0);
-
-  ASSERT_LE(*queued_timing, *submit_timing);
-  ASSERT_LT(*submit_timing, *start_timing);
-  ASSERT_LT(*start_timing, *end_timing);
-  ASSERT_LE(*end_timing, *complete_timing);
+  // AMD devices may report a "start" time before the "submit" time
+  UUR_KNOWN_FAILURE_ON(uur::HIP{});
+
+  // If a and b are supported, asserts that a <= b
+  auto test_timing = [=](ur_profiling_info_t a, ur_profiling_info_t b) {
+    std::stringstream trace{"Profiling Info: "};
+    trace << a << " <= " << b;
+    SCOPED_TRACE(trace.str());
+    uint64_t a_time;
+    auto result =
+        urEventGetProfilingInfo(event, a, sizeof(a_time), &a_time, nullptr);
+    if (result == UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION) {
+      return;
+    }
+    ASSERT_SUCCESS(result);
+
+    uint64_t b_time;
+    result =
+        urEventGetProfilingInfo(event, b, sizeof(b_time), &b_time, nullptr);
+    if (result == UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION) {
+      return;
+    }
+    ASSERT_SUCCESS(result);
+
+    // Note: This assumes that the counter doesn't overflow
+    ASSERT_LE(a_time, b_time);
+  };
+
+  test_timing(UR_PROFILING_INFO_COMMAND_QUEUED,
+              UR_PROFILING_INFO_COMMAND_SUBMIT);
+  test_timing(UR_PROFILING_INFO_COMMAND_SUBMIT,
+              UR_PROFILING_INFO_COMMAND_START);
+  test_timing(UR_PROFILING_INFO_COMMAND_START, UR_PROFILING_INFO_COMMAND_END);
+  test_timing(UR_PROFILING_INFO_COMMAND_END,
+              UR_PROFILING_INFO_COMMAND_COMPLETE);
 }
 
 UUR_INSTANTIATE_DEVICE_TEST_SUITE_P(

test/conformance/testing/include/uur/optional_queries.h

@@ -116,4 +116,16 @@ template <> inline bool isQueryOptional(ur_queue_info_t query) {
                    query) != optional_ur_queue_info_t.end();
 }
 
+constexpr std::array optional_ur_profiling_info_t = {
+    UR_PROFILING_INFO_COMMAND_QUEUED,   UR_PROFILING_INFO_COMMAND_SUBMIT,
+    UR_PROFILING_INFO_COMMAND_START,    UR_PROFILING_INFO_COMMAND_END,
+    UR_PROFILING_INFO_COMMAND_COMPLETE,
+};
+
+template <> inline bool isQueryOptional(ur_profiling_info_t query) {
+  return std::find(optional_ur_profiling_info_t.begin(),
+                   optional_ur_profiling_info_t.end(),
+                   query) != optional_ur_profiling_info_t.end();
+}
+
 } // namespace uur