Refactor command update API to take a list

Update the `urCommandBufferUpdateKernelLaunchExp` API for updating
commands in a command-buffer to take a list of commands.

The current API operates on a single command, this means that the
SYCL-Graph `update(std::vector<nodes>)` API needs to serialize
the list into N calls to the UR API. Given that both OpenCL
`clUpdateMutableCommandsKHR` and Level-Zero
`zeCommandListUpdateMutableCommandsExp` can operate on a list of
commands, this serialization at the UR layer of the stack
introduces extra host API calls.

This PR improves the `urCommandBufferUpdateKernelLaunchExp` API
so that a list of commands is passed all the way from SYCL to the
native backend API.

Author: Ewan Crawford

include/ur_api.h

@@ -9969,6 +9969,21 @@ typedef struct ur_exp_command_buffer_desc_t {
 
 } ur_exp_command_buffer_desc_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief A value that identifies a command inside of a command-buffer, used
+/// for
+///        defining dependencies between commands in the same command-buffer.
+typedef uint32_t ur_exp_command_buffer_sync_point_t;
+
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Handle of Command-Buffer object
+typedef struct ur_exp_command_buffer_handle_t_ *ur_exp_command_buffer_handle_t;
+
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Handle of a Command-Buffer command
+typedef struct ur_exp_command_buffer_command_handle_t_
+    *ur_exp_command_buffer_command_handle_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Descriptor type for updating a kernel command memobj argument.
 typedef struct ur_exp_command_buffer_update_memobj_arg_desc_t {
@@ -10032,6 +10047,8 @@ typedef struct ur_exp_command_buffer_update_kernel_launch_desc_t {
   ur_structure_type_t stype;
   /// [in][optional] pointer to extension-specific structure
   const void *pNext;
+  /// [in] Handle of the command-buffer kernel command to update.
+  ur_exp_command_buffer_command_handle_t hCommand;
   /// [in][optional] The new kernel handle. If this parameter is nullptr,
   /// the current kernel handle in `hCommand`
   /// will be used. If a kernel handle is passed, it must be a valid kernel
@@ -10081,21 +10098,6 @@ typedef struct ur_exp_command_buffer_update_kernel_launch_desc_t {
 
 } ur_exp_command_buffer_update_kernel_launch_desc_t;
 
-///////////////////////////////////////////////////////////////////////////////
-/// @brief A value that identifies a command inside of a command-buffer, used
-/// for
-///        defining dependencies between commands in the same command-buffer.
-typedef uint32_t ur_exp_command_buffer_sync_point_t;
-
-///////////////////////////////////////////////////////////////////////////////
-/// @brief Handle of Command-Buffer object
-typedef struct ur_exp_command_buffer_handle_t_ *ur_exp_command_buffer_handle_t;
-
-///////////////////////////////////////////////////////////////////////////////
-/// @brief Handle of a Command-Buffer command
-typedef struct ur_exp_command_buffer_command_handle_t_
-    *ur_exp_command_buffer_command_handle_t;
-
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Create a Command-Buffer object
 ///
@@ -11042,74 +11044,84 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferEnqueueExp(
 ///
 /// @details
 /// This entry-point is synchronous and may block if the command-buffer is
-/// executing when the entry-point is called.
+/// executing when the entry-point is called. On error, none of the updates to
+/// mutable-command objects are preserved.
 ///
 /// @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 == hCommand`
+///         + `NULL == hCommandBuffer`
+///         + `NULL == pUpdateKernelLaunch->hCommand`
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
 ///         + `NULL == pUpdateKernelLaunch`
+///     - ::UR_RESULT_ERROR_INVALID_COMMAND_BUFFER_EXP
+///     - ::UR_RESULT_ERROR_INVALID_SIZE
+///         + `numKernelUpdates == 0`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 ///         + If
 ///         ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_KERNEL_ARGUMENTS
-///         is not supported by the device, but any of
-///         `pUpdateKernelLaunch->numNewMemObjArgs`,
-///         `pUpdateKernelLaunch->numNewPointerArgs`, or
-///         `pUpdateKernelLaunch->numNewValueArgs` are not zero.
+///         is not supported by the device, and for any of any element of
+///         `pUpdateKernelLaunch` the `numNewMemObjArgs`, `numNewPointerArgs`,
+///         or `numNewValueArgs` members are not zero.
 ///         + If
 ///         ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_LOCAL_WORK_SIZE is
-///         not supported by the device but
-///         `pUpdateKernelLaunch->pNewLocalWorkSize` is not nullptr.
+///         not supported by the device, and for any element of
+///         `pUpdateKernelLaunch` the `pNewLocalWorkSize` member is not nullptr.
 ///         + If
 ///         ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_LOCAL_WORK_SIZE is
-///         not supported by the device but
-///         `pUpdateKernelLaunch->pNewLocalWorkSize` is nullptr and
-///         `pUpdateKernelLaunch->pNewGlobalWorkSize` is not nullptr.
+///         not supported by the device, and for any element of
+///         `pUpdateKernelLaunch` the `pNewLocalWorkSize` member is nullptr and
+///         `pNewGlobalWorkSize` is not nullptr.
 ///         + If
 ///         ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_GLOBAL_WORK_SIZE
-///         is not supported by the device but
-///         `pUpdateKernelLaunch->pNewGlobalWorkSize` is not nullptr
+///         is not supported by the device, and for any element of
+///         `pUpdateKernelLaunch` the `pNewGlobalWorkSize` member is not nullptr
 ///         + If
 ///         ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_GLOBAL_WORK_OFFSET
-///         is not supported by the device but
-///         `pUpdateKernelLaunch->pNewGlobalWorkOffset` is not nullptr.
+///         is not supported by the device, and for any element of
+///         `pUpdateKernelLaunch` the `pNewGlobalWorkOffset` member is not
+///         nullptr.
 ///         + If ::UR_DEVICE_COMMAND_BUFFER_UPDATE_CAPABILITY_FLAG_KERNEL_HANDLE
-///         is not supported by the device but `pUpdateKernelLaunch->hNewKernel`
-///         is not nullptr.
+///         is not supported by the device, and for any element of
+///         `pUpdateKernelLaunch` the `hNewKernel` member is not nullptr.
 ///     - ::UR_RESULT_ERROR_INVALID_OPERATION
 ///         + If ::ur_exp_command_buffer_desc_t::isUpdatable was not set to true
-///         on creation of the command-buffer `hCommand` belongs to.
-///         + If the command-buffer `hCommand` belongs to has not been
-///         finalized.
+///         on creation of the `hCommandBuffer`.
+///         + If `hCommandBuffer`  has not been finalized.
 ///     - ::UR_RESULT_ERROR_INVALID_COMMAND_BUFFER_COMMAND_HANDLE_EXP
-///         + If `hCommand` is not a kernel execution command.
+///         + If for any element of `pUpdateKernelLaunch` the `hCommand` member
+///         is not a kernel execution command.
+///         + If for any element of `pUpdateKernelLaunch` the `hCommand` member
+///         was not created from `hCommandBuffer`.
 ///     - ::UR_RESULT_ERROR_INVALID_MEM_OBJECT
 ///     - ::UR_RESULT_ERROR_INVALID_KERNEL_ARGUMENT_INDEX
 ///     - ::UR_RESULT_ERROR_INVALID_KERNEL_ARGUMENT_SIZE
 ///     - ::UR_RESULT_ERROR_INVALID_ENUMERATION
 ///     - ::UR_RESULT_ERROR_INVALID_WORK_DIMENSION
-///         + `pUpdateKernelLaunch->newWorkDim < 1 ||
-///         pUpdateKernelLaunch->newWorkDim > 3`
+///         + If for any element of `pUpdateKernelLaunch` the `newWorkDim`
+///         member is less than 1 or greater than 3.
 ///     - ::UR_RESULT_ERROR_INVALID_WORK_GROUP_SIZE
 ///     - ::UR_RESULT_ERROR_INVALID_VALUE
-///         + If `pUpdateKernelLaunch->hNewKernel` was not passed to the
-///         `hKernel` or `phKernelAlternatives` parameters of
-///         ::urCommandBufferAppendKernelLaunchExp when this command was
-///         created.
-///         + If `pUpdateKernelLaunch->newWorkDim` is different from the current
-///         workDim in `hCommand` and,
-///         `pUpdateKernelLaunch->pNewGlobalWorkSize`, or
-///         `pUpdateKernelLaunch->pNewGlobalWorkOffset` are nullptr.
+///         + If for any element of `pUpdateKernelLaunch` the `hNewKernel`
+///         member was not passed to the `hKernel` or `phKernelAlternatives`
+///         parameters of ::urCommandBufferAppendKernelLaunchExp when the
+///         command was created.
+///         + If for any element of `pUpdateKernelLaunch` the `newWorkDim`
+///         member is different from the current workDim in the `hCommand`
+///         member, and `pNewGlobalWorkSize` or `pNewGlobalWorkOffset` are
+///         nullptr.
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
 UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
-    /// [in] Handle of the command-buffer kernel command to update.
-    ur_exp_command_buffer_command_handle_t hCommand,
-    /// [in] Struct defining how the kernel command is to be updated.
+    /// [in] Handle of the command-buffer object.
+    ur_exp_command_buffer_handle_t hCommandBuffer,
+    /// [in] Length of pUpdateKernelLaunch.
+    uint32_t numKernelUpdates,
+    /// [in][range(0, numKernelUpdates)]  List of structs defining how a
+    /// kernel commands are to be updated.
     const ur_exp_command_buffer_update_kernel_launch_desc_t
         *pUpdateKernelLaunch);
 
@@ -14200,7 +14212,8 @@ typedef struct ur_command_buffer_enqueue_exp_params_t {
 /// @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_command_buffer_update_kernel_launch_exp_params_t {
-  ur_exp_command_buffer_command_handle_t *phCommand;
+  ur_exp_command_buffer_handle_t *phCommandBuffer;
+  uint32_t *pnumKernelUpdates;
   const ur_exp_command_buffer_update_kernel_launch_desc_t *
       *ppUpdateKernelLaunch;
 } ur_command_buffer_update_kernel_launch_exp_params_t;

include/ur_ddi.h

@@ -1599,7 +1599,7 @@ typedef ur_result_t(UR_APICALL *ur_pfnCommandBufferEnqueueExp_t)(
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function-pointer for urCommandBufferUpdateKernelLaunchExp
 typedef ur_result_t(UR_APICALL *ur_pfnCommandBufferUpdateKernelLaunchExp_t)(
-    ur_exp_command_buffer_command_handle_t,
+    ur_exp_command_buffer_handle_t, uint32_t,
     const ur_exp_command_buffer_update_kernel_launch_desc_t *);
 
 ///////////////////////////////////////////////////////////////////////////////

include/ur_print.hpp

@@ -11201,6 +11201,11 @@ inline std::ostream &operator<<(
 
   ur::details::printStruct(os, (params.pNext));
 
+  os << ", ";
+  os << ".hCommand = ";
+
+  ur::details::printPtr(os, (params.hCommand));
+
   os << ", ";
   os << ".hNewKernel = ";
 
@@ -18691,14 +18696,30 @@ inline std::ostream &
 operator<<(std::ostream &os, [[maybe_unused]] const struct
            ur_command_buffer_update_kernel_launch_exp_params_t *params) {
 
-  os << ".hCommand = ";
+  os << ".hCommandBuffer = ";
 
-  ur::details::printPtr(os, *(params->phCommand));
+  ur::details::printPtr(os, *(params->phCommandBuffer));
+
+  os << ", ";
+  os << ".numKernelUpdates = ";
+
+  os << *(params->pnumKernelUpdates);
 
   os << ", ";
   os << ".pUpdateKernelLaunch = ";
+  ur::details::printPtr(
+      os, reinterpret_cast<const void *>(*(params->ppUpdateKernelLaunch)));
+  if (*(params->ppUpdateKernelLaunch) != NULL) {
+    os << " {";
+    for (size_t i = 0; i < *params->pnumKernelUpdates; ++i) {
+      if (i != 0) {
+        os << ", ";
+      }
 
-  ur::details::printPtr(os, *(params->ppUpdateKernelLaunch));
+      os << (*(params->ppUpdateKernelLaunch))[i];
+    }
+    os << "}";
+  }
 
   return os;
 }

scripts/core/EXP-COMMAND-BUFFER.rst

@@ -309,7 +309,8 @@ ${x}CommandBufferUpdateKernelLaunchExp.
     ${x}_exp_command_buffer_update_kernel_launch_desc_t update {
         UR_STRUCTURE_TYPE_EXP_COMMAND_BUFFER_UPDATE_KERNEL_LAUNCH_DESC, // stype
         nullptr, // pNext
-        hNewKernel  // hNewKernel
+        hCommand, // hCommand
+        hNewKernel,  // hNewKernel
         2, // numNewMemobjArgs
         0, // numNewPointerArgs
         0, // numNewValueArgs
@@ -325,7 +326,7 @@ ${x}CommandBufferUpdateKernelLaunchExp.
     };
 
     // Perform the update
-    ${x}CommandBufferUpdateKernelLaunchExp(hCommand, &update);
+    ${x}CommandBufferUpdateKernelLaunchExp(hCommandBuffer, 1, &update);
 
 Command Event Update
 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
@@ -514,6 +515,8 @@ Changelog
 +-----------+-------------------------------------------------------+
 | 1.7       | Remove command handle reference counting and querying |
 +-----------+-------------------------------------------------------+
+| 1.8       | Change Kernel command update API to take a list       |
++-----------+-------------------------------------------------------+
 
 Contributors
 --------------------------------------------------------------------------------