Level Zero UR Reference Document
Contents
Level Zero UR Reference Document#
This document gives general guidelines on differences in the UR L0 adapter for customer usecases.
Environment Variables#
Environment Variable |
Description |
Possible Values |
Default Value |
|---|---|---|---|
UR_L0_USE_COPY_ENGINE |
Controls the use of copy engines. |
“0”: Copy engines will not be used. “1”: All available copy engines can be used. “lower_index:upper_index”: Specifies a range of copy engines to be used. |
“1” |
UR_L0_USE_IMMEDIATE_COMMANDLISTS |
Determines the mode of immediate command lists. |
“0”: Immediate command lists are not used. “1”: Immediate command lists are used per queue. “2”: Immediate command lists are used per thread per queue. |
“0” |
UR_L0_USE_RELAXED_ALLOCATION_LIMITS |
Controls the use of relaxed allocation limits. |
“0”: Relaxed allocation limits are not used. “1”: Relaxed allocation limits are used. |
“0” |
UR_L0_USE_DRIVER_INORDER_LISTS |
Controls the use of in-order lists from the driver. |
“0”: In-order lists from the driver are not used. “1”: In-order lists from the driver are used. |
“0” |
UR_L0_USM_ALLOCATOR_TRACE |
Enables tracing for the USM allocator. |
“0”: Tracing is disabled. “1”: Tracing is enabled. |
“0” |
UR_L0_USM_ALLOCATOR |
Configures the USM allocator. |
Specifies the configuration for the USM allocator. |
All Configs |
UR_L0_DEBUG_BASIC |
Enables basic debugging for Level Zero. |
“0”: Debugging is disabled. “1”: Debugging is enabled. |
“0” |
UR_L0_ENABLE_SYSMAN_ENV_DEFAULT |
Controls the default SysMan environment initialization. |
“1” or unset: Enables SysMan environment initialization. “0”: Disables SysMan environment initialization. |
“1” |
UR_L0_ENABLE_ZESINIT_DEFAULT |
Controls the default SysMan initialization with zesInit. |
“1”: Enables SysMan initialization with zesInit. “0” or unset: Disables SysMan initialization with zesInit. |
“0” |
SYCL_ENABLE_PCI |
Deprecated and no longer needed. |
Any value: Triggers a warning message. |
None |
UR_L0_USE_COPY_ENGINE_FOR_FILL |
Controls the use of copy engines for memory fill operations. |
“0”: Copy engines will not be used for fill operations. “1”: Copy engines will be used for fill operations. |
“0” |
UR_L0_DEBUG |
Controls the debug level for Level Zero. |
“0”: No debug information. “1”: Basic debug information. “2”: Validation debug information. “-1”: All debug information. |
“0” |
UR_L0_LEAKS_DEBUG |
Enables debugging for memory leaks. |
“0”: Memory leaks debugging is disabled. “1”: Memory leaks debugging is enabled. |
“0” |
UR_L0_INIT_ALL_DRIVERS |
Controls the initialization of all Level Zero drivers. |
“0”: Only currently used drivers are initialized. “1”: All drivers on the system are initialized. |
“0” |
UR_L0_SERIALIZE |
Controls serialization of Level Zero calls. |
“0”: No locking or blocking. “1”: Locking around each UR_CALL. “2”: Blocking UR calls where supported. |
“0” |
UR_L0_QUEUE_SYNCHRONIZE_NON_BLOCKING |
Controls non-blocking synchronization of queues. |
“0”: Non-blocking synchronization is disabled. “1”: Non-blocking synchronization is enabled. |
“0” |
UR_L0_OOQ_INTEGRATED_SIGNAL_EVENT |
Controls signal events for commands on integrated GPUs. |
“0”: Signal events are not created. “1”: Signal events are created. |
“0” |
UR_L0_TRACK_INDIRECT_ACCESS_MEMORY |
Enables tracking of indirect access memory. |
“0”: Tracking is disabled. “1”: Tracking is enabled. |
“0” |
UR_L0_EXPOSE_CSLICE_IN_AFFINITY_PARTITIONING |
Controls exposure of CSlice in affinity partitioning. |
“0”: CSlice is not exposed. “1”: CSlice is exposed. |
“0” |
UR_L0_MAX_NUMBER_OF_EVENTS_PER_EVENT_POOL |
Sets the maximum number of events per event pool. |
Any positive integer: Specifies the maximum number of events per event pool. |
256 |
UR_L0_COMMANDLISTS_CLEANUP_THRESHOLD |
Sets the threshold for command lists cleanup. |
Any positive integer: Specifies the threshold for cleanup. Negative value: Disables the threshold. |
20 |
UR_L0_USE_NATIVE_USM_MEMCPY2D |
Controls the use of native USM memcpy2D operations. |
“0”: Native USM memcpy2D operations are not used. “1”: Native USM memcpy2D operations are used. |
“0” |
UR_L0_ENABLE_USM_HOSTPTR_IMPORT |
Enables USM host pointer import. |
“0”: USM host pointer import is disabled. “1”: USM host pointer import is enabled. |
“0” |
UR_L0_ENABLE_USM_HOSTPTR_RELEASE |
Enables USM host pointer release. |
“0”: USM host pointer release is disabled. “1”: USM host pointer release is enabled. |
“0” |
UR_L0_ENABLE_USM_HOST_UNIFIED_MEMORY |
Enables USM host unified memory. |
“0”: USM host unified memory is disabled. “1”: USM host unified memory is enabled. |
“0” |
UR_L0_USE_MULTIPLE_COMMANDLIST_BARRIERS |
Controls the use of multiple command lists for barriers. |
“0”: Multiple command lists are not used. “1”: Multiple command lists are used. |
“0” |
UR_L0_IN_ORDER_BARRIER_BY_SIGNAL |
Controls if in-order barriers are implemented by signal. |
“0”: Barriers are implemented by true barrier command. “1”: Barriers are implemented by signal. |
“0” |
UR_L0_DISABLE_EVENTS_CACHING |
Controls the caching of events in the context. |
“0” or unset: Event caching is enabled. “1”: Event caching is disabled. |
“0” |
UR_L0_REUSE_DISCARDED_EVENTS |
Controls the reuse of uncompleted events in in-order queues. |
“0”: Reuse of discarded events is disabled. “1” or unset: Reuse of discarded events is enabled. |
“1” |
SYCL_PI_LEVEL_ZERO_FILTER_EVENT_WAIT_LIST |
Controls filtering of event wait lists. |
“0” or unset: Filtering is disabled. “1”: Filtering is enabled. |
“0” |
UR_L0_DEVICE_SCOPE_EVENTS |
Controls the scope of device events. |
“0”: All events are host-visible. “1”: On-demand host-visible proxy events. “2”: Last command in batch host-visible. |
“0” |
UR_L0_USE_COPY_ENGINE_FOR_D2D_COPY |
Controls the use of copy engines for device-to-device copy operations. |
“0”: Copy engines will not be used for D2D copy operations. “1”: Copy engines will be used for D2D copy operations. |
“0” |
UR_L0_BATCH_SIZE |
Controls the batch size for command lists. |
“0”: Dynamic batch size adjustment. Any positive integer: Specifies the fixed batch size. |
“0” |
UR_L0_COPY_BATCH_SIZE |
Controls the batch size for copy command lists. |
“0”: Dynamic batch size adjustment. Any positive integer: Specifies the fixed batch size. |
“0” |
UR_L0_IMMEDIATE_COMMANDLISTS_BATCH_MAX |
Sets the maximum number of immediate command lists batches. |
Any positive integer: Specifies the maximum number of batches |
10 |
UR_L0_IMMEDIATE_COMMANDLISTS_EVENTS_PER_BATCH |
Sets the number of events per batch for immediate command lists. |
Any positive integer: Specifies the number of events per batch. |
256 |
UR_L0_USE_COMPUTE_ENGINE |
Controls the use of compute engines. |
“0”: Only the first compute engine is used. Any positive integer: Specifies the index of the compute engine to be used. Negative value: All available compute engines may be used. |
“0” |
UR_L0_USM_RESIDENT |
Controls memory residency for USM allocations. |
“0xHSD”: Specifies residency for host, shared, and device allocations. |
0x2 |
UR_L0_DISABLE_USM_ALLOCATOR |
Controls the use of the USM allocator. |
“0”: USM allocator is enabled. Any other value: USM allocator is disabled. |
“0” |
UR_L0_CMD_BUFFER_USE_IMMEDIATE_APPEND_PATH |
Controls which command-buffer implementation path is used. The paths rely on different APIs to enqueue command-buffers. The immediate append path relies on zeCommandListImmediateAppendCommandListsExp(). The wait-event path relies on zeCommandQueueExecuteCommandLists() |
“1”: the immediate append path will always be enabled as long as the pre-requisites are met. “0”: the immediate append path will always be disabled. “unset”: The default behaviour will be used which enables the immediate append path only for some devices when the pre-requisites are met. |
Unset |
UR_L0_VECTOR_WIDTH_SIZE |
Specifies the size (in bits) of the vector width supported by the Level Zero device. This value indicates the maximum number of data elements that can be processed simultaneously in a single instruction, which is useful for optimizing data-parallel workloads and understanding device caps. |
Any positive integer: Indicates the maximum number of data elements that can be processed simultaneously in a single instruction. The value entered by user user is the desired width size to config. If this width size is not supported, then the default “max” will be used. |
Device-specific |
Contributors#
Neil Spruit neil.r.spruit@intel.com
Fábio Mestre fabio.mestre@codeplay.com