Merge branch 'llvm' into review/yang/invalid_arguments

Author: AllanZyne

.github/workflows/e2e_level_zero.yml

@@ -21,9 +21,9 @@ jobs:
       config: ""
       unit: "gpu"
       # Failing tests
-      xfail: "ESIMD/regression/minmax.cpp;ESIMD/preemption.cpp;Matrix/SG32/element_wise_all_ops.cpp;Matrix/SG32/get_coord_int8_matB.cpp;Matrix/element_wise_all_ops.cpp;Matrix/element_wise_all_ops_1d.cpp;Matrix/element_wise_all_ops_1d_cont.cpp;Matrix/element_wise_all_ops_scalar.cpp;Matrix/element_wise_ops.cpp;Matrix/get_coord_int8_matB.cpp;Matrix/joint_matrix_apply_bf16.cpp;Matrix/joint_matrix_apply_two_matrices.cpp;Matrix/joint_matrix_bfloat16.cpp;Matrix/joint_matrix_bfloat16_array.cpp;Matrix/joint_matrix_rowmajorA_rowmajorB.cpp;ProgramManager/uneven_kernel_split.cpp"
+      xfail: "Matrix/SG32/get_coord_int8_matB.cpp;Matrix/element_wise_all_ops_1d.cpp;Matrix/element_wise_all_ops_1d_cont.cpp;Matrix/element_wise_all_ops_scalar.cpp;Matrix/get_coord_int8_matB.cpp;Matrix/joint_matrix_rowmajorA_rowmajorB.cpp;Plugin/level_zero_barrier_optimization.cpp"
       # Flaky tests
       filter_out: "UserDefinedReductions/user_defined_reductions.cpp"
       # These runners by default spawn upwards of 260 workers.
       # We also add a time out just in case some test hangs
-      extra_lit_flags: "-sv -j 100 --max-time 600"
+      extra_lit_flags: "--param gpu-intel-pvc=True -sv -j 100 --max-time 600"

.github/workflows/multi_device.yml

@@ -55,7 +55,8 @@ jobs:
 
     - name: Test adapter specific
       working-directory: ${{github.workspace}}/build
-      run: ctest -C ${{matrix.build_type}} --output-on-failure -L "adapter-specific" --timeout 180
+      run: ctest -C ${{matrix.build_type}} --output-on-failure -L "adapter-specific" -E "test-adapter-level_zero_multi_queue" --timeout 180
+      # TODO: investigate why test-adapter-level_zero_multi_queue fails on newer driver
 
     - name: Test adapters
       working-directory: ${{github.workspace}}/build

include/ur_api.h

@@ -2401,16 +2401,19 @@ typedef struct ur_context_native_properties_t {
 ///     - ::UR_RESULT_ERROR_UNINITIALIZED
 ///     - ::UR_RESULT_ERROR_DEVICE_LOST
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
+///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
+///         + `NULL == hAdapter`
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
-///         + `NULL == phDevices`
 ///         + `NULL == phContext`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 ///         + If the adapter has no underlying equivalent handle.
 UR_APIEXPORT ur_result_t UR_APICALL
 urContextCreateWithNativeHandle(
     ur_native_handle_t hNativeContext,                 ///< [in][nocheck] the native handle of the context.
+    ur_adapter_handle_t hAdapter,                      ///< [in] handle of the adapter that owns the native handle
     uint32_t numDevices,                               ///< [in] number of devices associated with the context
-    const ur_device_handle_t *phDevices,               ///< [in][range(0, numDevices)] list of devices associated with the context
+    const ur_device_handle_t *phDevices,               ///< [in][optional][range(0, numDevices)] list of devices associated with
+                                                       ///< the context
     const ur_context_native_properties_t *pProperties, ///< [in][optional] pointer to native context properties struct
     ur_context_handle_t *phContext                     ///< [out] pointer to the handle of the context object created.
 );
@@ -5278,6 +5281,8 @@ typedef struct ur_kernel_native_properties_t {
 ///     - The application may call this function from simultaneous threads for
 ///       the same context.
 ///     - The implementation of this function should be thread-safe.
+///     - The implementation may require a valid program handle to return the
+///       native kernel handle
 ///
 /// @returns
 ///     - ::UR_RESULT_SUCCESS
@@ -5286,7 +5291,7 @@ typedef struct ur_kernel_native_properties_t {
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == hContext`
-///         + `NULL == hProgram`
+///         + If `hProgram == NULL` and the implementation requires a valid program.
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
 ///         + `NULL == phKernel`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
@@ -5295,7 +5300,7 @@ UR_APIEXPORT ur_result_t UR_APICALL
 urKernelCreateWithNativeHandle(
     ur_native_handle_t hNativeKernel,                 ///< [in][nocheck] the native handle of the kernel.
     ur_context_handle_t hContext,                     ///< [in] handle of the context object
-    ur_program_handle_t hProgram,                     ///< [in] handle of the program associated with the kernel
+    ur_program_handle_t hProgram,                     ///< [in][optional] handle of the program associated with the kernel
     const ur_kernel_native_properties_t *pProperties, ///< [in][optional] pointer to native kernel properties struct
     ur_kernel_handle_t *phKernel                      ///< [out] pointer to the handle of the kernel object created.
 );
@@ -5625,7 +5630,6 @@ typedef struct ur_queue_native_properties_t {
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == hContext`
-///         + `NULL == hDevice`
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
 ///         + `NULL == phQueue`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
@@ -5634,7 +5638,7 @@ UR_APIEXPORT ur_result_t UR_APICALL
 urQueueCreateWithNativeHandle(
     ur_native_handle_t hNativeQueue,                 ///< [in][nocheck] the native handle of the queue.
     ur_context_handle_t hContext,                    ///< [in] handle of the context object
-    ur_device_handle_t hDevice,                      ///< [in] handle of the device object
+    ur_device_handle_t hDevice,                      ///< [in][optional] handle of the device object
     const ur_queue_native_properties_t *pProperties, ///< [in][optional] pointer to native queue properties struct
     ur_queue_handle_t *phQueue                       ///< [out] pointer to the handle of the queue object created.
 );
@@ -9822,6 +9826,7 @@ typedef struct ur_context_get_native_handle_params_t {
 ///     allowing the callback the ability to modify the parameter's value
 typedef struct ur_context_create_with_native_handle_params_t {
     ur_native_handle_t *phNativeContext;
+    ur_adapter_handle_t *phAdapter;
     uint32_t *pnumDevices;
     const ur_device_handle_t **pphDevices;
     const ur_context_native_properties_t **ppProperties;

include/ur_ddi.h

@@ -135,6 +135,7 @@ typedef ur_result_t(UR_APICALL *ur_pfnContextGetNativeHandle_t)(
 /// @brief Function-pointer for urContextCreateWithNativeHandle
 typedef ur_result_t(UR_APICALL *ur_pfnContextCreateWithNativeHandle_t)(
     ur_native_handle_t,
+    ur_adapter_handle_t,
     uint32_t,
     const ur_device_handle_t *,
     const ur_context_native_properties_t *,

include/ur_print.hpp

@@ -10668,6 +10668,12 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
     ur::details::printPtr(os, reinterpret_cast<void *>(
                                   *(params->phNativeContext)));
 
+    os << ", ";
+    os << ".hAdapter = ";
+
+    ur::details::printPtr(os,
+                          *(params->phAdapter));
+
     os << ", ";
     os << ".numDevices = ";
 

scripts/core/context.yml

@@ -228,12 +228,15 @@ params:
       name: hNativeContext
       desc: |
             [in][nocheck] the native handle of the context.
+    - type: $x_adapter_handle_t
+      name: hAdapter
+      desc: "[in] handle of the adapter that owns the native handle"
     - type: uint32_t
       name: numDevices
       desc: "[in] number of devices associated with the context"
     - type: "const $x_device_handle_t*"
       name: phDevices
-      desc: "[in][range(0, numDevices)] list of devices associated with the context"
+      desc: "[in][optional][range(0, numDevices)] list of devices associated with the context"
     - type: "const $x_context_native_properties_t*"
       name: pProperties
       desc: "[in][optional] pointer to native context properties struct"

scripts/core/kernel.yml

@@ -513,6 +513,7 @@ details:
     - "Creates runtime kernel handle from native driver kernel handle."
     - "The application may call this function from simultaneous threads for the same context."
     - "The implementation of this function should be thread-safe."
+    - "The implementation may require a valid program handle to return the native kernel handle"
 params:
     - type: $x_native_handle_t
       name: hNativeKernel
@@ -523,7 +524,7 @@ params:
       desc: "[in] handle of the context object"
     - type: $x_program_handle_t
       name: hProgram
-      desc: "[in] handle of the program associated with the kernel"
+      desc: "[in][optional] handle of the program associated with the kernel"
     - type: "const $x_kernel_native_properties_t*"
       name: pProperties
       desc: "[in][optional] pointer to native kernel properties struct"
@@ -534,6 +535,8 @@ params:
 returns:
     - $X_RESULT_ERROR_UNSUPPORTED_FEATURE:
         - "If the adapter has no underlying equivalent handle."
+    - $X_RESULT_ERROR_INVALID_NULL_HANDLE:
+        - "If `hProgram == NULL` and the implementation requires a valid program."
 --- #--------------------------------------------------------------------------
 type: function
 desc: "Get the suggested local work size for a kernel."

scripts/core/queue.yml

@@ -286,7 +286,7 @@ params:
       desc: "[in] handle of the context object"
     - type: $x_device_handle_t
       name: hDevice
-      desc: "[in] handle of the device object"
+      desc: "[in][optional] handle of the device object"
     - type: "const $x_queue_native_properties_t*"
       name: pProperties
       desc: "[in][optional] pointer to native queue properties struct"

source/adapters/cuda/context.cpp

@@ -143,6 +143,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextGetNativeHandle(
 
 UR_APIEXPORT ur_result_t UR_APICALL urContextCreateWithNativeHandle(
     [[maybe_unused]] ur_native_handle_t hNativeContext,
+    [[maybe_unused]] ur_adapter_handle_t hAdapter,
     [[maybe_unused]] uint32_t numDevices,
     [[maybe_unused]] const ur_device_handle_t *phDevices,
     [[maybe_unused]] const ur_context_native_properties_t *pProperties,

source/adapters/hip/context.cpp

@@ -122,6 +122,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextGetNativeHandle(
 
 UR_APIEXPORT ur_result_t UR_APICALL urContextCreateWithNativeHandle(
     [[maybe_unused]] ur_native_handle_t hNativeContext,
+    [[maybe_unused]] ur_adapter_handle_t hAdapter,
     [[maybe_unused]] uint32_t numDevices,
     [[maybe_unused]] const ur_device_handle_t *phDevices,
     [[maybe_unused]] const ur_context_native_properties_t *pProperties,