[EXP][CMDBUF] Support for Node Profiling

include/ur_api.h

@@ -215,6 +215,7 @@ typedef enum ur_function_t {
     UR_FUNCTION_COMMAND_BUFFER_APPEND_USM_ADVISE_EXP = 213,                    ///< Enumerator for ::urCommandBufferAppendUSMAdviseExp
     UR_FUNCTION_ENQUEUE_COOPERATIVE_KERNEL_LAUNCH_EXP = 214,                   ///< Enumerator for ::urEnqueueCooperativeKernelLaunchExp
     UR_FUNCTION_KERNEL_SUGGEST_MAX_COOPERATIVE_GROUP_COUNT_EXP = 215,          ///< Enumerator for ::urKernelSuggestMaxCooperativeGroupCountExp
+    UR_FUNCTION_EVENT_GET_SYNC_POINT_PROFILING_INFO_EXP = 218,                 ///< Enumerator for ::urEventGetSyncPointProfilingInfoExp
     /// @cond
     UR_FUNCTION_FORCE_UINT32 = 0x7fffffff
     /// @endcond
@@ -8341,6 +8342,40 @@ urCommandBufferEnqueueExp(
                                                    ///< command-buffer execution instance.
 );
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Get profiling information for the sync point execution associated with
+///        an event object
+///
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_UNINITIALIZED
+///     - ::UR_RESULT_ERROR_DEVICE_LOST
+///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
+///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
+///         + `NULL == hEvent`
+///     - ::UR_RESULT_ERROR_INVALID_ENUMERATION
+///         + `::UR_PROFILING_INFO_COMMAND_COMPLETE < propName`
+///     - ::UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE
+///         + If `hEvent`s associated queue was not created with `::UR_QUEUE_FLAG_PROFILING_ENABLE`.
+///     - ::UR_RESULT_ERROR_INVALID_VALUE
+///         + `pPropValue && propSize == 0`
+///     - ::UR_RESULT_ERROR_INVALID_EVENT
+///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
+///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
+///     - ::UR_RESULT_ERROR_INVALID_COMMAND_BUFFER_SYNC_POINT_EXP - "If `syncPoint` does not reference a command in the command-buffer submission associated with `hEvent.`"
+UR_APIEXPORT ur_result_t UR_APICALL
+urEventGetSyncPointProfilingInfoExp(
+    ur_event_handle_t hEvent,                     ///< [in] handle of the event object
+    ur_exp_command_buffer_sync_point_t syncPoint, ///< [in] Sync point referencing the node (i.e. command) from which we want
+                                                  ///< to get profile information
+    ur_profiling_info_t propName,                 ///< [in] the name of the profiling property to query
+    size_t propSize,                              ///< [in] size in bytes of the profiling property value
+    void *pPropValue,                             ///< [out][optional][typename(propName, propSize)] value of the profiling
+                                                  ///< property
+    size_t *pPropSizeRet                          ///< [out][optional] pointer to the actual size in bytes returned in
+                                                  ///< propValue
+);
+
 #if !defined(__GNUC__)
 #pragma endregion
 #endif
@@ -9026,6 +9061,19 @@ typedef struct ur_event_set_callback_params_t {
     void **ppUserData;
 } ur_event_set_callback_params_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function parameters for urEventGetSyncPointProfilingInfoExp
+/// @details Each entry is a pointer to the parameter passed to the function;
+///     allowing the callback the ability to modify the parameter's value
+typedef struct ur_event_get_sync_point_profiling_info_exp_params_t {
+    ur_event_handle_t *phEvent;
+    ur_exp_command_buffer_sync_point_t *psyncPoint;
+    ur_profiling_info_t *ppropName;
+    size_t *ppropSize;
+    void **ppPropValue;
+    size_t **ppPropSizeRet;
+} ur_event_get_sync_point_profiling_info_exp_params_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function parameters for urProgramCreateWithIL
 /// @details Each entry is a pointer to the parameter passed to the function;

include/ur_ddi.h

@@ -269,6 +269,43 @@ typedef ur_result_t(UR_APICALL *ur_pfnGetEventProcAddrTable_t)(
     ur_api_version_t,
     ur_event_dditable_t *);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function-pointer for urEventGetSyncPointProfilingInfoExp
+typedef ur_result_t(UR_APICALL *ur_pfnEventGetSyncPointProfilingInfoExp_t)(
+    ur_event_handle_t,
+    ur_exp_command_buffer_sync_point_t,
+    ur_profiling_info_t,
+    size_t,
+    void *,
+    size_t *);
+
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Table of EventExp functions pointers
+typedef struct ur_event_exp_dditable_t {
+    ur_pfnEventGetSyncPointProfilingInfoExp_t pfnGetSyncPointProfilingInfoExp;
+} ur_event_exp_dditable_t;
+
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Exported function for filling application's EventExp table
+///        with current process' addresses
+///
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_UNINITIALIZED
+///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_VERSION
+UR_DLLEXPORT ur_result_t UR_APICALL
+urGetEventExpProcAddrTable(
+    ur_api_version_t version,          ///< [in] API version requested
+    ur_event_exp_dditable_t *pDdiTable ///< [in,out] pointer to table of DDI function pointers
+);
+
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function-pointer for urGetEventExpProcAddrTable
+typedef ur_result_t(UR_APICALL *ur_pfnGetEventExpProcAddrTable_t)(
+    ur_api_version_t,
+    ur_event_exp_dditable_t *);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function-pointer for urProgramCreateWithIL
 typedef ur_result_t(UR_APICALL *ur_pfnProgramCreateWithIL_t)(
@@ -2306,6 +2343,7 @@ typedef struct ur_dditable_t {
     ur_platform_dditable_t Platform;
     ur_context_dditable_t Context;
     ur_event_dditable_t Event;
+    ur_event_exp_dditable_t EventExp;
     ur_program_dditable_t Program;
     ur_program_exp_dditable_t ProgramExp;
     ur_kernel_dditable_t Kernel;

include/ur_print.h

@@ -1162,6 +1162,14 @@ UR_APIEXPORT ur_result_t UR_APICALL urPrintEventCreateWithNativeHandleParams(con
 ///         - `buff_size < out_size`
 UR_APIEXPORT ur_result_t UR_APICALL urPrintEventSetCallbackParams(const struct ur_event_set_callback_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print ur_event_get_sync_point_profiling_info_exp_params_t struct
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_INVALID_SIZE
+///         - `buff_size < out_size`
+UR_APIEXPORT ur_result_t UR_APICALL urPrintEventGetSyncPointProfilingInfoExpParams(const struct ur_event_get_sync_point_profiling_info_exp_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print ur_program_create_with_il_params_t struct
 /// @returns

include/ur_print.hpp

@@ -879,6 +879,9 @@ inline std::ostream &operator<<(std::ostream &os, ur_function_t value) {
     case UR_FUNCTION_KERNEL_SUGGEST_MAX_COOPERATIVE_GROUP_COUNT_EXP:
         os << "UR_FUNCTION_KERNEL_SUGGEST_MAX_COOPERATIVE_GROUP_COUNT_EXP";
         break;
+    case UR_FUNCTION_EVENT_GET_SYNC_POINT_PROFILING_INFO_EXP:
+        os << "UR_FUNCTION_EVENT_GET_SYNC_POINT_PROFILING_INFO_EXP";
+        break;
     default:
         os << "unknown enumerator";
         break;
@@ -9888,6 +9891,45 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
     return os;
 }
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print operator for the ur_event_get_sync_point_profiling_info_exp_params_t type
+/// @returns
+///     std::ostream &
+inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct ur_event_get_sync_point_profiling_info_exp_params_t *params) {
+
+    os << ".hEvent = ";
+
+    ur::details::printPtr(os,
+                          *(params->phEvent));
+
+    os << ", ";
+    os << ".syncPoint = ";
+
+    os << *(params->psyncPoint);
+
+    os << ", ";
+    os << ".propName = ";
+
+    os << *(params->ppropName);
+
+    os << ", ";
+    os << ".propSize = ";
+
+    os << *(params->ppropSize);
+
+    os << ", ";
+    os << ".pPropValue = ";
+    ur::details::printTagged(os, *(params->ppPropValue), *(params->ppropName), *(params->ppropSize));
+
+    os << ", ";
+    os << ".pPropSizeRet = ";
+
+    ur::details::printPtr(os,
+                          *(params->ppPropSizeRet));
+
+    return os;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print operator for the ur_program_create_with_il_params_t type
 /// @returns
@@ -15965,6 +16007,9 @@ inline ur_result_t UR_APICALL printFunctionParams(std::ostream &os, ur_function_
     case UR_FUNCTION_EVENT_SET_CALLBACK: {
         os << (const struct ur_event_set_callback_params_t *)params;
     } break;
+    case UR_FUNCTION_EVENT_GET_SYNC_POINT_PROFILING_INFO_EXP: {
+        os << (const struct ur_event_get_sync_point_profiling_info_exp_params_t *)params;
+    } break;
     case UR_FUNCTION_PROGRAM_CREATE_WITH_IL: {
         os << (const struct ur_program_create_with_il_params_t *)params;
     } break;

scripts/core/EXP-COMMAND-BUFFER.rst

@@ -132,6 +132,23 @@ were obtained from.
                                            pLocalWorkSize, 1, &syncPoint, 
                                            nullptr);
 
+    // Finalize the CommandBuffer to launch it
+    ${x}CommandBufferFinalizeExp(hCommandBuffer);
+
+    // Execute the CommandBuffer and obtain the event associated to this
+    // execution
+    ${x}_event_handle_t event;
+    ${x}CommandBufferEnqueueExp(hCommandBuffer, hQueue, 0, nullptr,
+                                &event);
+
+    // Get SyncPoint profiling information
+    ${x}_profiling_info_t propName;
+    size_t propSize;
+    void* pPropValue;
+    size_t pPropSizeRet;
+    ${x}EventGetSyncPointProfilingInfoExp(event, syncPoint, propName, propSize,
+                                          pPropValue, &pPropSizeRet);
+
 Enqueueing Command-Buffers
 --------------------------------------------------------------------------------
 
@@ -211,6 +228,7 @@ Functions
 * ${x}CommandBufferAppendUSMPrefetchExp
 * ${x}CommandBufferAppendUSMAdviseExp
 * ${x}CommandBufferEnqueueExp
+* ${x}EventGetSyncPointProfilingInfoExp
 
 Changelog
 --------------------------------------------------------------------------------
@@ -227,6 +245,9 @@ Changelog
 | 1.3       | Add function definitions for Prefetch and Advise      |
 |           | commands                                              |
 +-----------+-------------------------------------------------------+
+| 1.4       | Add function definitions for getting sync point       |
+|           | profiling information                                 |
++-----------+-------------------------------------------------------+
 
 Contributors
 --------------------------------------------------------------------------------

scripts/core/exp-command-buffer.yml

@@ -704,3 +704,38 @@ returns:
         - "If event objects in phEventWaitList are not valid events."
     - $X_RESULT_ERROR_OUT_OF_HOST_MEMORY
     - $X_RESULT_ERROR_OUT_OF_RESOURCES
+--- #--------------------------------------------------------------------------
+type: function
+desc: "Get profiling information for the sync point execution associated with an event object"
+class: $xEvent
+name: GetSyncPointProfilingInfoExp
+ordinal: "0"
+params:
+    - type: $x_event_handle_t
+      name: hEvent
+      desc: "[in] handle of the event object"
+    - type: $x_exp_command_buffer_sync_point_t
+      name: syncPoint
+      desc: "[in] Sync point referencing the node (i.e. command) from which we want to get profile information"
+    - type: $x_profiling_info_t
+      name: propName
+      desc: "[in] the name of the profiling property to query"
+    - type: size_t
+      name: propSize
+      desc: "[in] size in bytes of the profiling property value"
+    - type: void*
+      name: pPropValue
+      desc: "[out][optional][typename(propName, propSize)] value of the profiling property"
+    - type: size_t*
+      name: pPropSizeRet
+      desc: "[out][optional] pointer to the actual size in bytes returned in propValue"
+returns:
+    - $X_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE:
+      - "If `hEvent`s associated queue was not created with `$X_QUEUE_FLAG_PROFILING_ENABLE`."
+    - $X_RESULT_ERROR_INVALID_VALUE:
+      - "`pPropValue && propSize == 0`"
+    - $X_RESULT_ERROR_INVALID_EVENT
+    - $X_RESULT_ERROR_OUT_OF_RESOURCES
+    - $X_RESULT_ERROR_OUT_OF_HOST_MEMORY
+    - $X_RESULT_ERROR_INVALID_COMMAND_BUFFER_SYNC_POINT_EXP
+      - "If `syncPoint` does not reference a command in the command-buffer submission associated with `hEvent.`"

scripts/core/registry.yml

@@ -559,6 +559,9 @@ etors:
 - name: KERNEL_SUGGEST_MAX_COOPERATIVE_GROUP_COUNT_EXP
   desc: Enumerator for $xKernelSuggestMaxCooperativeGroupCountExp
   value: '215'
+- name: EVENT_GET_SYNC_POINT_PROFILING_INFO_EXP
+  desc: Enumerator for $xEventGetSyncPointProfilingInfoExp
+  value: '218'
 ---
 type: enum
 desc: Defines structure types

source/adapters/adapter.def.in

@@ -7,6 +7,7 @@ EXPORTS
 	urGetEnqueueProcAddrTable
 	urGetEnqueueExpProcAddrTable
 	urGetEventProcAddrTable
+	urGetEventExpProcAddrTable
 	urGetKernelProcAddrTable
 	urGetKernelExpProcAddrTable
 	urGetMemProcAddrTable

source/adapters/adapter.map.in

@@ -7,6 +7,7 @@
 		urGetEnqueueProcAddrTable;
 		urGetEnqueueExpProcAddrTable;
 		urGetEventProcAddrTable;
+		urGetEventExpProcAddrTable;
 		urGetKernelProcAddrTable;
 		urGetKernelExpProcAddrTable;
 		urGetMemProcAddrTable;

source/adapters/cuda/command_buffer.cpp

@@ -762,3 +762,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferEnqueueExp(
 
   return Result;
 }
+
+UR_APIEXPORT ur_result_t UR_APICALL urEventGetSyncPointProfilingInfoExp(
+    ur_event_handle_t hEvent, ur_exp_command_buffer_sync_point_t SyncPoint,
+    ur_profiling_info_t PropName, size_t PropValueSize, void *PropValue,
+    size_t *PropValueSizeRet) {
+  (void)hEvent;
+  (void)SyncPoint;
+  (void)PropName;
+  (void)PropValueSize;
+  (void)PropValue;
+  (void)PropValueSizeRet;
+  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+}

source/adapters/cuda/ur_interface_loader.cpp

@@ -298,6 +298,20 @@ UR_DLLEXPORT ur_result_t UR_APICALL urGetCommandBufferExpProcAddrTable(
   return retVal;
 }
 
+UR_DLLEXPORT ur_result_t UR_APICALL urGetEventExpProcAddrTable(
+    ur_api_version_t version, ///< [in] API version requested
+    ur_event_exp_dditable_t
+        *pDdiTable ///< [in,out] pointer to table of DDI function pointers
+) {
+  auto retVal = validateProcInputs(version, pDdiTable);
+  if (UR_RESULT_SUCCESS != retVal) {
+    return retVal;
+  }
+  pDdiTable->pfnGetSyncPointProfilingInfoExp =
+      urEventGetSyncPointProfilingInfoExp;
+  return retVal;
+}
+
 UR_DLLEXPORT ur_result_t UR_APICALL urGetUsmP2PExpProcAddrTable(
     ur_api_version_t version, ur_usm_p2p_exp_dditable_t *pDdiTable) {
   auto retVal = validateProcInputs(version, pDdiTable);

source/adapters/hip/command_buffer.cpp

@@ -162,3 +162,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferEnqueueExp(
                   "implemented for HIP adapter.");
   return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }
+
+UR_APIEXPORT ur_result_t UR_APICALL urEventGetSyncPointProfilingInfoExp(
+    ur_event_handle_t hEvent, ur_exp_command_buffer_sync_point_t SyncPoint,
+    ur_profiling_info_t PropName, size_t PropValueSize, void *PropValue,
+    size_t *PropValueSizeRet) {
+  (void)hEvent;
+  (void)SyncPoint;
+  (void)PropName;
+  (void)PropValueSize;
+  (void)PropValue;
+  (void)PropValueSizeRet;
+  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+}

source/adapters/hip/ur_interface_loader.cpp

@@ -295,6 +295,20 @@ UR_DLLEXPORT ur_result_t UR_APICALL urGetCommandBufferExpProcAddrTable(
   return retVal;
 }
 
+UR_DLLEXPORT ur_result_t UR_APICALL urGetEventExpProcAddrTable(
+    ur_api_version_t version, ///< [in] API version requested
+    ur_event_exp_dditable_t
+        *pDdiTable ///< [in,out] pointer to table of DDI function pointers
+) {
+  auto retVal = validateProcInputs(version, pDdiTable);
+  if (UR_RESULT_SUCCESS != retVal) {
+    return retVal;
+  }
+  pDdiTable->pfnGetSyncPointProfilingInfoExp =
+      urEventGetSyncPointProfilingInfoExp;
+  return retVal;
+}
+
 UR_DLLEXPORT ur_result_t UR_APICALL urGetUsmP2PExpProcAddrTable(
     ur_api_version_t version, ur_usm_p2p_exp_dditable_t *pDdiTable) {
   auto retVal = validateProcInputs(version, pDdiTable);