[UR][L0][CUDA][HIP] Add enqueue timestamp recording extension

include/ur_api.h

@@ -222,6 +222,7 @@ typedef enum ur_function_t {
     UR_FUNCTION_COMMAND_BUFFER_UPDATE_KERNEL_LAUNCH_EXP = 220,                 ///< Enumerator for ::urCommandBufferUpdateKernelLaunchExp
     UR_FUNCTION_COMMAND_BUFFER_GET_INFO_EXP = 221,                             ///< Enumerator for ::urCommandBufferGetInfoExp
     UR_FUNCTION_COMMAND_BUFFER_COMMAND_GET_INFO_EXP = 222,                     ///< Enumerator for ::urCommandBufferCommandGetInfoExp
+    UR_FUNCTION_ENQUEUE_TIMESTAMP_RECORDING_EXP = 223,                         ///< Enumerator for ::urEnqueueTimestampRecordingExp
     /// @cond
     UR_FUNCTION_FORCE_UINT32 = 0x7fffffff
     /// @endcond
@@ -1641,6 +1642,7 @@ typedef enum ur_device_info_t {
                                                                      ///< backed 3D sampled image data.
     UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_EXP = 0x2017,     ///< [::ur_bool_t] returns true if the device is capable of fetching
                                                                      ///< non-USM backed 3D sampled image data.
+    UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP = 0x2018,         ///< [::ur_bool_t] returns true if the device supports timestamp recording
     /// @cond
     UR_DEVICE_INFO_FORCE_UINT32 = 0x7fffffff
     /// @endcond
@@ -1666,7 +1668,7 @@ typedef enum ur_device_info_t {
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == hDevice`
 ///     - ::UR_RESULT_ERROR_INVALID_ENUMERATION
-///         + `::UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_EXP < propName`
+///         + `::UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP < propName`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION
 ///         + If `propName` is not supported by the adapter.
 ///     - ::UR_RESULT_ERROR_INVALID_SIZE
@@ -5618,6 +5620,7 @@ typedef enum ur_command_t {
     UR_COMMAND_COMMAND_BUFFER_ENQUEUE_EXP = 0x1000,   ///< Event created by ::urCommandBufferEnqueueExp
     UR_COMMAND_INTEROP_SEMAPHORE_WAIT_EXP = 0x2000,   ///< Event created by ::urBindlessImagesWaitExternalSemaphoreExp
     UR_COMMAND_INTEROP_SEMAPHORE_SIGNAL_EXP = 0x2001, ///< Event created by ::urBindlessImagesSignalExternalSemaphoreExp
+    UR_COMMAND_TIMESTAMP_RECORDING_EXP = 0x2002,      ///< Event created by ::urEnqueueTimestampRecordingExp
     /// @cond
     UR_COMMAND_FORCE_UINT32 = 0x7fffffff
     /// @endcond
@@ -8890,6 +8893,46 @@ urKernelSuggestMaxCooperativeGroupCountExp(
     uint32_t *pGroupCountRet        ///< [out] pointer to maximum number of groups
 );
 
+#if !defined(__GNUC__)
+#pragma endregion
+#endif
+// Intel 'oneAPI' Unified Runtime Experimental APIs for enqueuing timestamp recordings
+#if !defined(__GNUC__)
+#pragma region enqueue timestamp recording(experimental)
+#endif
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Enqueue a command for recording the device timestamp
+///
+/// @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 == hQueue`
+///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
+///         + `NULL == phEvent`
+///     - ::UR_RESULT_ERROR_INVALID_EVENT_WAIT_LIST
+UR_APIEXPORT ur_result_t UR_APICALL
+urEnqueueTimestampRecordingExp(
+    ur_queue_handle_t hQueue,                 ///< [in] handle of the queue object
+    bool blocking,                            ///< [in] indicates whether the call to this function should block until
+                                              ///< until the device timestamp recording command has executed on the
+                                              ///< device.
+    uint32_t numEventsInWaitList,             ///< [in] size of the event wait list
+    const ur_event_handle_t *phEventWaitList, ///< [in][optional][range(0, numEventsInWaitList)] pointer to a list of
+                                              ///< events that must be complete before the kernel execution.
+                                              ///< If nullptr, the numEventsInWaitList must be 0, indicating no wait
+                                              ///< events.
+    ur_event_handle_t *phEvent                ///< [in,out] return an event object that identifies this particular kernel
+                                              ///< execution instance. Profiling information can be queried
+                                              ///< from this event as if `hQueue` had profiling enabled. Querying
+                                              ///< `UR_PROFILING_INFO_COMMAND_QUEUED` or `UR_PROFILING_INFO_COMMAND_SUBMIT`
+                                              ///< reports the timestamp at the time of the call to this function.
+                                              ///< Querying `UR_PROFILING_INFO_COMMAND_START` or `UR_PROFILING_INFO_COMMAND_END`
+                                              ///< reports the timestamp recorded when the command is executed on the device.
+);
+
 #if !defined(__GNUC__)
 #pragma endregion
 #endif
@@ -10600,6 +10643,18 @@ typedef struct ur_enqueue_cooperative_kernel_launch_exp_params_t {
     ur_event_handle_t **pphEvent;
 } ur_enqueue_cooperative_kernel_launch_exp_params_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function parameters for urEnqueueTimestampRecordingExp
+/// @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_enqueue_timestamp_recording_exp_params_t {
+    ur_queue_handle_t *phQueue;
+    bool *pblocking;
+    uint32_t *pnumEventsInWaitList;
+    const ur_event_handle_t **pphEventWaitList;
+    ur_event_handle_t **pphEvent;
+} ur_enqueue_timestamp_recording_exp_params_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function parameters for urBindlessImagesUnsampledImageHandleDestroyExp
 /// @details Each entry is a pointer to the parameter passed to the function;

include/ur_ddi.h

@@ -1448,10 +1448,20 @@ typedef ur_result_t(UR_APICALL *ur_pfnEnqueueCooperativeKernelLaunchExp_t)(
     const ur_event_handle_t *,
     ur_event_handle_t *);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function-pointer for urEnqueueTimestampRecordingExp
+typedef ur_result_t(UR_APICALL *ur_pfnEnqueueTimestampRecordingExp_t)(
+    ur_queue_handle_t,
+    bool,
+    uint32_t,
+    const ur_event_handle_t *,
+    ur_event_handle_t *);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Table of EnqueueExp functions pointers
 typedef struct ur_enqueue_exp_dditable_t {
     ur_pfnEnqueueCooperativeKernelLaunchExp_t pfnCooperativeKernelLaunchExp;
+    ur_pfnEnqueueTimestampRecordingExp_t pfnTimestampRecordingExp;
 } ur_enqueue_exp_dditable_t;
 
 ///////////////////////////////////////////////////////////////////////////////

include/ur_print.h

@@ -1954,6 +1954,14 @@ UR_APIEXPORT ur_result_t UR_APICALL urPrintEnqueueWriteHostPipeParams(const stru
 ///         - `buff_size < out_size`
 UR_APIEXPORT ur_result_t UR_APICALL urPrintEnqueueCooperativeKernelLaunchExpParams(const struct ur_enqueue_cooperative_kernel_launch_exp_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print ur_enqueue_timestamp_recording_exp_params_t struct
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_INVALID_SIZE
+///         - `buff_size < out_size`
+UR_APIEXPORT ur_result_t UR_APICALL urPrintEnqueueTimestampRecordingExpParams(const struct ur_enqueue_timestamp_recording_exp_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print ur_bindless_images_unsampled_image_handle_destroy_exp_params_t struct
 /// @returns

include/ur_print.hpp

@@ -916,6 +916,9 @@ inline std::ostream &operator<<(std::ostream &os, enum ur_function_t value) {
     case UR_FUNCTION_COMMAND_BUFFER_COMMAND_GET_INFO_EXP:
         os << "UR_FUNCTION_COMMAND_BUFFER_COMMAND_GET_INFO_EXP";
         break;
+    case UR_FUNCTION_ENQUEUE_TIMESTAMP_RECORDING_EXP:
+        os << "UR_FUNCTION_ENQUEUE_TIMESTAMP_RECORDING_EXP";
+        break;
     default:
         os << "unknown enumerator";
         break;
@@ -2571,6 +2574,9 @@ inline std::ostream &operator<<(std::ostream &os, enum ur_device_info_t value) {
     case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_EXP:
         os << "UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_EXP";
         break;
+    case UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP:
+        os << "UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP";
+        break;
     default:
         os << "unknown enumerator";
         break;
@@ -4280,6 +4286,18 @@ inline ur_result_t printTagged(std::ostream &os, const void *ptr, ur_device_info
 
         os << ")";
     } break;
+    case UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP: {
+        const ur_bool_t *tptr = (const ur_bool_t *)ptr;
+        if (sizeof(ur_bool_t) > size) {
+            os << "invalid size (is: " << size << ", expected: >=" << sizeof(ur_bool_t) << ")";
+            return UR_RESULT_ERROR_INVALID_SIZE;
+        }
+        os << (const void *)(tptr) << " (";
+
+        os << *tptr;
+
+        os << ")";
+    } break;
     default:
         os << "unknown enumerator";
         return UR_RESULT_ERROR_INVALID_ENUMERATION;
@@ -8788,6 +8806,9 @@ inline std::ostream &operator<<(std::ostream &os, enum ur_command_t value) {
     case UR_COMMAND_INTEROP_SEMAPHORE_SIGNAL_EXP:
         os << "UR_COMMAND_INTEROP_SEMAPHORE_SIGNAL_EXP";
         break;
+    case UR_COMMAND_TIMESTAMP_RECORDING_EXP:
+        os << "UR_COMMAND_TIMESTAMP_RECORDING_EXP";
+        break;
     default:
         os << "unknown enumerator";
         break;
@@ -14104,6 +14125,48 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
     return os;
 }
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print operator for the ur_enqueue_timestamp_recording_exp_params_t type
+/// @returns
+///     std::ostream &
+inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct ur_enqueue_timestamp_recording_exp_params_t *params) {
+
+    os << ".hQueue = ";
+
+    ur::details::printPtr(os,
+                          *(params->phQueue));
+
+    os << ", ";
+    os << ".blocking = ";
+
+    os << *(params->pblocking);
+
+    os << ", ";
+    os << ".numEventsInWaitList = ";
+
+    os << *(params->pnumEventsInWaitList);
+
+    os << ", ";
+    os << ".phEventWaitList = {";
+    for (size_t i = 0; *(params->pphEventWaitList) != NULL && i < *params->pnumEventsInWaitList; ++i) {
+        if (i != 0) {
+            os << ", ";
+        }
+
+        ur::details::printPtr(os,
+                              (*(params->pphEventWaitList))[i]);
+    }
+    os << "}";
+
+    os << ", ";
+    os << ".phEvent = ";
+
+    ur::details::printPtr(os,
+                          *(params->pphEvent));
+
+    return os;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print operator for the ur_bindless_images_unsampled_image_handle_destroy_exp_params_t type
 /// @returns
@@ -17126,6 +17189,9 @@ inline ur_result_t UR_APICALL printFunctionParams(std::ostream &os, ur_function_
     case UR_FUNCTION_ENQUEUE_COOPERATIVE_KERNEL_LAUNCH_EXP: {
         os << (const struct ur_enqueue_cooperative_kernel_launch_exp_params_t *)params;
     } break;
+    case UR_FUNCTION_ENQUEUE_TIMESTAMP_RECORDING_EXP: {
+        os << (const struct ur_enqueue_timestamp_recording_exp_params_t *)params;
+    } break;
     case UR_FUNCTION_BINDLESS_IMAGES_UNSAMPLED_IMAGE_HANDLE_DESTROY_EXP: {
         os << (const struct ur_bindless_images_unsampled_image_handle_destroy_exp_params_t *)params;
     } break;

scripts/core/EXP-ENQUEUE-TIMESTAMP-RECORDING.rst

@@ -0,0 +1,70 @@
+<%
+    OneApi=tags['$OneApi']
+    x=tags['$x']
+    X=x.upper()
+%>
+
+.. _experimental-enqueue-timestamp-recording:
+
+================================================================================
+Enqueue Timestamp Recording
+================================================================================
+
+.. warning::
+
+    Experimental features:
+
+    *   May be replaced, updated, or removed at any time.
+    *   Do not require maintaining API/ABI stability of their own additions over
+        time.
+    *   Do not require conformance testing of their own additions.
+
+
+Motivation
+--------------------------------------------------------------------------------
+Currently, the only way to get timestamp information is through enabling
+profiling on a queue and retrieving the information from events coming from
+commands submitted to it. However, not all systems give full control of the
+queue construction to the programmer wanting the profiling information. To amend
+this, this extension adds the ability to enqueue a timestamp recording on any
+queue, with or without profiling enabled. This event can in turn be queried for
+the usual profiling information.
+
+
+API
+--------------------------------------------------------------------------------
+
+Enums
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+* ${x}_device_info_t
+    * ${X}_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP
+
+* ${x}_command_t
+    * ${X}_COMMAND_TIMESTAMP_RECORDING_EXP
+
+Functions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* ${x}EnqueueTimestampRecordingExp
+
+Changelog
+--------------------------------------------------------------------------------
+
++-----------+------------------------+
+| Revision  | Changes                |
++===========+========================+
+| 1.0       | Initial Draft          |
++-----------+------------------------+
+
+
+Support
+--------------------------------------------------------------------------------
+
+Adapters which support this experimental feature *must* return true for the new
+`${X}_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP` device info query.
+
+
+Contributors
+--------------------------------------------------------------------------------
+
+* Steffen Larsen `steffen.larsen@intel.com <steffen.larsen@intel.com>`_

scripts/core/exp-enqueue-timestamp-recording.yml

@@ -0,0 +1,66 @@
+#
+# Copyright (C) 2024 Intel Corporation
+#
+# Part of the Unified-Runtime Project, under the Apache License v2.0 with LLVM Exceptions.
+# See LICENSE.TXT
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# See YaML.md for syntax definition
+#
+--- #--------------------------------------------------------------------------
+type: header
+desc: "Intel $OneApi Unified Runtime Experimental APIs for enqueuing timestamp recordings"
+ordinal: "99"
+--- #--------------------------------------------------------------------------
+type: enum
+extend: true
+typed_etors: true
+desc: "Extension enums to $x_device_info_t to support timestamp recordings."
+name: $x_device_info_t
+etors: 
+    - name: TIMESTAMP_RECORDING_SUPPORT_EXP
+      value: "0x2018"
+      desc: "[$x_bool_t] returns true if the device supports timestamp recording"
+--- #--------------------------------------------------------------------------
+type: enum
+extend: true
+desc: "Command Type experimental enumerations."
+name: $x_command_t
+etors:
+    - name: TIMESTAMP_RECORDING_EXP
+      value: "0x2002"
+      desc: Event created by $xEnqueueTimestampRecordingExp
+--- #--------------------------------------------------------------------------
+type: function
+desc: "Enqueue a command for recording the device timestamp"
+class: $xEnqueue
+name: TimestampRecordingExp
+params:
+    - type: $x_queue_handle_t
+      name: hQueue
+      desc: "[in] handle of the queue object"
+    - type: bool
+      name: blocking
+      desc: |
+            [in] indicates whether the call to this function should block until
+            until the device timestamp recording command has executed on the
+            device.
+    - type: uint32_t
+      name: numEventsInWaitList
+      desc: "[in] size of the event wait list"
+    - type: "const $x_event_handle_t*"
+      name: phEventWaitList
+      desc: |
+            [in][optional][range(0, numEventsInWaitList)] pointer to a list of events that must be complete before the kernel execution.
+            If nullptr, the numEventsInWaitList must be 0, indicating no wait events. 
+    - type: $x_event_handle_t*
+      name: phEvent
+      desc: |
+            [in,out] return an event object that identifies this particular kernel execution instance. Profiling information can be queried
+            from this event as if `hQueue` had profiling enabled. Querying `UR_PROFILING_INFO_COMMAND_QUEUED` or `UR_PROFILING_INFO_COMMAND_SUBMIT`
+            reports the timestamp at the time of the call to this function. Querying `UR_PROFILING_INFO_COMMAND_START` or `UR_PROFILING_INFO_COMMAND_END`
+            reports the timestamp recorded when the command is executed on the device.
+returns:
+    - $X_RESULT_ERROR_INVALID_NULL_HANDLE
+    - $X_RESULT_ERROR_INVALID_NULL_POINTER
+    - $X_RESULT_ERROR_INVALID_EVENT_WAIT_LIST

scripts/core/registry.yml

@@ -580,6 +580,9 @@ etors:
 - name: COMMAND_BUFFER_COMMAND_GET_INFO_EXP
   desc: Enumerator for $xCommandBufferCommandGetInfoExp
   value: '222'
+- name: ENQUEUE_TIMESTAMP_RECORDING_EXP
+  desc: Enumerator for $xEnqueueTimestampRecordingExp
+  value: '223'
 ---
 type: enum
 desc: Defines structure types