DECLARE_DEBUG_VARIABLE(std::string, OverrideGdiPath, std::string("unk"), "When different value than \"unk\", will override default path to gdi library.")
DECLARE_DEBUG_VARIABLE(std::string, AubDumpAddMmioRegistersList, std::string("unk"), "Semicolon separated sequence of additional MMIO registers offset;values pairs i.e. 0x111;0x123;0x222;0x456")
DECLARE_DEBUG_VARIABLE(bool, AUBDumpAllocsOnEnqueueReadOnly, false, "Force dumping buffers and images on clEnqueueReadBuffer/Image only (blocking calls)")
DECLARE_DEBUG_VARIABLE(bool, TrackParentEvents, false, "Events track their parents")
DECLARE_DEBUG_VARIABLE(bool, RebuildPrecompiledKernels, false, "Forces driver to recompile precompiled kernels from sources; applies to builtin and user kernels")
DECLARE_DEBUG_VARIABLE(bool, EnableResourceTags, false, "Enable resource tagging in GmmLib")
DECLARE_DEBUG_VARIABLE(bool, DoNotFreeResources, false, "Do not free resources")
DECLARE_DEBUG_VARIABLE(bool, AllowMixingRegularAndCooperativeKernels, false, "Allow mixing regular and cooperative kernels in a single command list and in a single execute")
DECLARE_DEBUG_VARIABLE(bool, AllowPatchingVfeStateInCommandLists, false, "Allow programming MEDIA_VFE_STATE in a command list")
DECLARE_DEBUG_VARIABLE(bool, ForceTheoreticalMaxWorkGroupCount, false, "Do not apply any limitation to max cooperative/concurrent work-group count queries")
DECLARE_DEBUG_VARIABLE(bool, DontDisableZebinIfVmeUsed, false, "When enabled, driver will not add -cl-intel-disable-zebin internal option when vme is used")
DECLARE_DEBUG_VARIABLE(bool, AppendMemoryPrefetchForKmdMigratedSharedAllocations, true, "Allow prefetching shared memory to the device associated with the specified command list")
DECLARE_DEBUG_VARIABLE(std::string, FilterDeviceId, std::string("unk"), "Device id filter, adapter matching device id will be opened; ignored when unk")
DECLARE_DEBUG_VARIABLE(std::string, FilterBdfPath, std::string("unk"), "Linux-only, BDF path filter, only matching paths will be opened; ignored when unk")
DECLARE_DEBUG_VARIABLE(std::string, LoadBinarySipFromFile, std::string("unk"), "Select binary file to load SIP kernel raw binary; when file named *_header.* exists, it is used as header")
DECLARE_DEBUG_VARIABLE(std::string, InjectInternalBuildOptions, std::string("unk"), "Append provided string to internal build options for user modules; ignored when unk")
DECLARE_DEBUG_VARIABLE(std::string, InjectApiBuildOptions, std::string("unk"), "Append provided string to api build options for user modules; ignored when unk")
DECLARE_DEBUG_VARIABLE(std::string, OverrideDeviceName, std::string("unk"), "Override device name to provided string; ignored when unk")
DECLARE_DEBUG_VARIABLE(int64_t, OverrideMultiStoragePlacement, -1, "Place memory only in selected tiles indicated by bit mask; ignore when -1")
DECLARE_DEBUG_VARIABLE(int64_t, ForceCompressionDisabledForCompressedBlitCopies, -1, "If compression is required, set AUX_CCS_E, but force CompressionEnable filed; 0 should result in uncompressed read/write; values = -1: default, 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, ForceL1Caching, -1, "Program L1 cache policy for surface state and stateless accesses; values = -1: default, 0: disable, 1: enable")
DECLARE_DEBUG_VARIABLE(int32_t, ForceAuxTranslationEnabled, -1, "Require AUX translation for kernels; values = -1: default, 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideStatelessMocsIndex, -1, "Program provided MOCS index for stateless accesses in state base address for regular buffers; ignore when -1")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideMocsIndexForScratchSpace, -1, "Program provided MOCS index for stateful accesses in surface state for scratch space; ignore when -1")
DECLARE_DEBUG_VARIABLE(int32_t, CFEFusedEUDispatch, -1, "Set Fused EU dispatch in FrontEnd State command; values = -1: default, 0: enabled, 1: disabled")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideGpuAddressSpace, -1, "Set GPU address space range in bits; ignore when -1")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideMaxWorkgroupSize, -1, "Set max workgroup size; ignore when -1")
DECLARE_DEBUG_VARIABLE(int32_t, DoCpuCopyOnReadBuffer, -1, "Override CPU copy behavior for buffer reads; values = -1: default, 0: do not use CPU copy, 1: triggers CPU copy path for Read Buffer calls, only supported for some basic use cases (no blocked user events in dependencies tree)")
DECLARE_DEBUG_VARIABLE(int32_t, DoCpuCopyOnWriteBuffer, -1, "Override CPU copy behavior for buffer writes; values = -1: default, 0: do not use CPU copy, 1: triggers CPU copy path for Write Buffer calls, only supported for some basic use cases (no blocked user events in dependencies tree)")
DECLARE_DEBUG_VARIABLE(int32_t, PauseOnEnqueue, -1, "-1: default, -2: always, x: pause on enqueue number x and ask for user confirmation before and after execution, counted from 0")
DECLARE_DEBUG_VARIABLE(int32_t, PauseOnBlitCopy, -1, "-1: default, -2: always, x: pause on blit enqueue number x and ask for user confirmation before and after execution, counted from 0. Note that single blit enqueue may have multiple copy instructions")
DECLARE_DEBUG_VARIABLE(int32_t, PauseOnGpuMode, -1, "-1: default (before and after), 0: before only, 1: after only")
DECLARE_DEBUG_VARIABLE(int32_t, MultiStorageGranularity, -1, "Forces granularity for chunk size and mapping colouring policies: -1: default 64 kB, >0 - enabled chunk size based colouring with granularity in kb (should be multiple of 64kb page size, because of alignment restrictions)")
DECLARE_DEBUG_VARIABLE(int32_t, ForceCsrLockInBcsEnqueueOnlyForGpgpuSubmission, -1, "-1: Default, 1: Force gpgpu command stream receiver lock for bcs enqueue only when gpgpu submission, 0: Do nothing, if possible")
DECLARE_DEBUG_VARIABLE(int32_t, FormatForStatelessCompressionWithUnifiedMemory, 0xF, "Format for stateless compression with unified memory")
DECLARE_DEBUG_VARIABLE(int32_t, ForceMultiGpuPartialWritesInComputeMode, -1, "-1: default - 0 for multiOsContext capable, 0: program value 0 in MultiGpuPartialWrites bit in STATE_COMPUTE_MODE, 1: program value 1 in MultiGpuPartialWrites bit in STATE_COMPUTE_MODE,")
DECLARE_DEBUG_VARIABLE(int32_t, ForceMultiGpuPartialWrites, -1, "-1: default - 0 for multiOsContext capable, 0: program value 0 in MultiGpuPartialWrites controls 1: program value 1 in MultiGpuPartialWrites controls")
DECLARE_DEBUG_VARIABLE(int32_t, ForceMultiGpuAtomicsInComputeMode, -1, "-1: default - 0 for multiOsContext capable, 0: program value 0 in MultiGpuAtomics bit in STATE_COMPUTE_MODE, 1: program value 1 in MultiGpuAtomics bit in STATE_COMPUTE_MODE")
DECLARE_DEBUG_VARIABLE(int32_t, ForceMultiGpuAtomics, -1, "-1: default - 0 for multiOsContext capable, 0: program value 0 in MultiGpuAtomics controls 1: program value 1 in MultiGpuAtomics controls")
DECLARE_DEBUG_VARIABLE(int32_t, ForceBufferCompressionFormat, -1, "-1: default, >0: Format value")
DECLARE_DEBUG_VARIABLE(int32_t, EnableHwGenerationLocalIds, -1, "-1: default, 0: disable, 1: enable : Enables generation of local ids on HW")
DECLARE_DEBUG_VARIABLE(int32_t, WalkerPartitionPreferHighestDimension, -1, "-1: default, 0: prefer biggest dimension, 1: prefer Z over Y over X if they divide partition count evenly")
DECLARE_DEBUG_VARIABLE(int32_t, SetMinimalPartitionSize, -1, "-1 default value set to 512 workgroups, 0 - disabled, >0 - minimal partition size in workgroups (should be power of 2)")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideBlitterTargetMemory, -1, "-1:default 0: overwrites to System 1: overwrites to Local")
DECLARE_DEBUG_VARIABLE(int32_t, EnableImmediateVmBindExt, -1, "Use immediate bind extension to a new residency model on Linux (requires kernel support), -1: default (enabled with direct submission), 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, ForceExecutionTile, -1, "-1: default, 0+: given tile is chosen as submission, must be used with EnableWalkerPartition = 0.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideL1CacheControlInSurfaceState, -1, "-1: feature inactive, >=0 : following L1 cache control value will be programmed in render surface state (for regular buffers)")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideL1CacheControlInSurfaceStateForScratchSpace, -1, "-1: feature inactive, >=0 : following L1 cache control value will be programmed in render surface state for scratch space")
DECLARE_DEBUG_VARIABLE(int32_t, OverridePreferredSlmAllocationSizePerDss, -1, "-1: default, >=0: program value for preferred SLM allocation size per dss")
DECLARE_DEBUG_VARIABLE(int32_t, ForceThreadArbitrationPolicyProgrammingWithScm, -1, "-1: default, 0: do not program thread arbitration policy with SCM, 1: program thread arbitration policy with SCM")
DECLARE_DEBUG_VARIABLE(int32_t, ForceZPassAsyncComputeThreadLimit, -1, "-1: default, >0: Limit value in STATE_COMPUTE_MODE")
DECLARE_DEBUG_VARIABLE(int32_t, ForcePixelAsyncComputeThreadLimit, -1, "-1: default, >0: Limit value in STATE_COMPUTE_MODE")
DECLARE_DEBUG_VARIABLE(int32_t, DecompressInL3ForImage2dFromBuffer, -1, "-1: default, 0: WA Disabled, 1: WA enabled - Enable DecompressInL3 for image 2d from compressed buffer")
DECLARE_DEBUG_VARIABLE(int32_t, ToggleHintKernelDisableCompression, -1, "-1: default - use kernel as source of hint, 0: provide hint to disable compression, 1: provide hint to enable compression")
DECLARE_DEBUG_VARIABLE(int32_t, CFEComputeDispatchAllWalkerEnable, -1, "Set Compute Dispatch All Walker flag in CFE_STATE on XE_HPC_CORE platforms, -1: do not set")
DECLARE_DEBUG_VARIABLE(int32_t, ComputeDispatchAllWalkerEnableInComputeWalker, -1, "Set Compute Dispatch All Walker flag in COMPUTE_WALKER on PVC+ platforms, -1: do not set")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramGlobalFenceAsMiMemFenceCommandInCommandStream, -1, "-1: default, 0: disable, 1: enable, System Memory Fence executed as MI_MEM_FENCE command in command stream")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramGlobalFenceAsPostSyncOperationInComputeWalker, -1, "-1: default, 0: disable, 1: enable, System Memory Fence executed as post-sync operation in COMPUTE_WALKER")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramGlobalFenceAsKernelInstructionInEUKernel, -1, "-1: default, 0: disable, 1: enable, System Memory Fence executed as kernel instruction in EU kernel")
DECLARE_DEBUG_VARIABLE(int32_t, DoNotReportTile1BscWaActive, -1, "-1: default, 0: WA Disabled, 1: WA enabled")
DECLARE_DEBUG_VARIABLE(int32_t, ForceTile0PlacementForTile1ResourcesWaActive, -1, "-1: default, 0: WA Disabled, 1: WA enabled")
DECLARE_DEBUG_VARIABLE(int32_t, ClosNumCacheWays, -1, "-1: default, >= 0 Number of cache ways to allocate within previously reserved CLOS regions (0-32)")
DECLARE_DEBUG_VARIABLE(int32_t, EngineUsageHint, -1, "-1: default, >=0: engine usage value to use when creating command queue on user selected engine")
DECLARE_DEBUG_VARIABLE(int32_t, UseDrmVirtualEnginesForCcs, -1, "-1: default, 0: disable, 1: enable, Combine all CCS nodes to single VE (per context)")
DECLARE_DEBUG_VARIABLE(int32_t, UseDrmVirtualEnginesForBcs, -1, "-1: default, 0: disable, 1: enable, Combine all BCS nodes to single VE (per context)")
DECLARE_DEBUG_VARIABLE(int32_t, LimitEngineCountForVirtualBcs, -1, "-1: default, >0 Only use VirtualEngine with limited amount of engines, not max ")
DECLARE_DEBUG_VARIABLE(int32_t, LimitEngineCountForVirtualCcs, -1, "-1: default, >0 Only use VirtualEngine with limited amount of engines, not max ")
DECLARE_DEBUG_VARIABLE(int32_t, KMDSupportForCrossTileMigrationPolicy, -1, "-1: default, 0: no KMD support, 1: KMD support for cross-tile migrations of shared allocations")
DECLARE_DEBUG_VARIABLE(int32_t, UseTileMemoryBankInVirtualMemoryCreation, -1, "-1: default - on, 0: do not assign tile memory bank to virtual memory space, 1: assign tile memory bank to virtual memory space")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideTimestampEvents, -1, "-1: default (based on user settings), 0: Force disable timestamp events (no timestamps will be reported), 1: Force enable timestamp events")
DECLARE_DEBUG_VARIABLE(int32_t, BatchBufferStartPrepatchingWaEnabled, -1, "-1: default , 0: disabled, 1: enabled. WA applies valid VA pointing to 'self' instead of 0x0. This mitigates incorrect VA preparsing.")
DECLARE_DEBUG_VARIABLE(int32_t, ReadBackCommandBufferAllocation, -1, "Read command buffer allocation back on the host side. -1: default, 0 - disabled, 1 - local memory only, 2 - local and system memory")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideCmdListCmdBufferSizeInKb, -1, "-1: default, 0: disable, >0: size in KB. Override cmd list command buffer size in KB.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideL1CachePolicyInSurfaceStateAndStateless, -1, "-1: default, >=0 : following policy will be programmed in render surface state (for regular buffers) and stateless L1 caching")
DECLARE_DEBUG_VARIABLE(int32_t, ForceEvictOnlyIfNecessaryFlag, -1, "-1: default - driver selects when to use, 0: force never use this flag, 1: force always use this flag")
DECLARE_DEBUG_VARIABLE(int32_t, CopyHostPtrOnCpu, -1, "-1: default, 0: disable, 1:enable, In clCreateBuffer with CL_MEM_COPY_HOST_PTR, copy memory using locked ptr on cpu")
DECLARE_DEBUG_VARIABLE(int32_t, AdjustThreadGroupDispatchSize, -1, "-1: default, 0: do not adjust thread group dispatch size 1: adjust thread group dispatch size (PVC)")
DECLARE_DEBUG_VARIABLE(int32_t, EnableMultipleRegularContextForBcs, -1, "-1: default, 0: disabled, 1: Use NumberOfRegularContextsPerEngine to create multiple Regular contexts on the same engine")
DECLARE_DEBUG_VARIABLE(int32_t, DisableScratchPages, -1, "-1: default, 0: do not disable scratch pages during VM creations, 1: disable scratch pages during VM creations")
DECLARE_DEBUG_VARIABLE(int32_t, DebugSetMemoryDiagnosticsDelay, -1, "-1: default, >=0: delay time in minutes necessary for completion of Memory diagnostics")
DECLARE_DEBUG_VARIABLE(int32_t, EnableDeviceStateVerification, -1, "-1: default, 0: disable, 1: enable check of device state before submit on Windows")
DECLARE_DEBUG_VARIABLE(int32_t, PrintTimestampPacketUsage, -1, "-1: default, 0: Disabled, 1: Print when TSP is allocated, initialized, returned to pool, etc.")
DECLARE_DEBUG_VARIABLE(int32_t, TrackNumCsrClientsOnSyncPoints, -1, "-1: default, 0: Disabled, 1: If set, synchronization points like zeEventHostSynchronize will unregister CmdQ from CSR clients")
DECLARE_DEBUG_VARIABLE(int32_t, WaitForUserFenceOnEventHostSynchronize, -1, "-1: default, 0: Disabled, 1: Enabled. If enabled, use WaitUserFence KMD call for in-order Events instead of active polling on host.")
DECLARE_DEBUG_VARIABLE(int32_t, DisableSystemPointerKernelArgument, -1, "-1: default, 0: Disabled, 1: using a system pointer for kernel argument returns an error.")
DECLARE_DEBUG_VARIABLE(bool, PrintDebugSettings, false, "Dump all debug variables settings to text file. Print to stdout if value is different than default.")
DECLARE_DEBUG_VARIABLE(bool, PrintProgramBinaryProcessingTime, false, "prints execution time of Program::processGenBinary() method during program building")
DECLARE_DEBUG_VARIABLE(bool, DisableZeroCopyForBuffers, false, "When active all buffer allocations will not share memory with CPU.")
DECLARE_DEBUG_VARIABLE(bool, DisableDcFlushInEpilogue, false, "Disable DC flush in epilogue")
DECLARE_DEBUG_VARIABLE(bool, EnableNullHardware, false, "works on Windows only, sets the Null Hardware flag that makes all Command buffers completed while GPU does nothing")
DECLARE_DEBUG_VARIABLE(bool, ForceLinearImages, false, "Force linear images. Default is Y-tiled.")
DECLARE_DEBUG_VARIABLE(bool, ForceSLML3Config, false, "Forces L3Config with SLM for all kernels")
DECLARE_DEBUG_VARIABLE(bool, Force32bitAddressing, false, "Forces 32 bit addresses to be used in 64 bit dll")
DECLARE_DEBUG_VARIABLE(bool, ForceCsrFlushing, false, "Forces flushing of command stream receiver")
DECLARE_DEBUG_VARIABLE(bool, ForceCsrReprogramming, false, "Forces reprogramming of command stream receiver")
DECLARE_DEBUG_VARIABLE(bool, UseNoRingFlushesKmdMode, true, "Windows only, passes flag to KMD that informs KMD to not emit any ring buffer flushes.")
DECLARE_DEBUG_VARIABLE(bool, DisableZeroCopyForUseHostPtr, false, "When active all buffer allocations created with CL_MEM_USE_HOST_PTR flag will not share memory with CPU.")
DECLARE_DEBUG_VARIABLE(int32_t, MaxHwThreadsPercent, 0, "If not zero then maximum number of used HW threads is capped to max * MaxHwThreadsPercent / 100")
DECLARE_DEBUG_VARIABLE(int32_t, MinHwThreadsUnoccupied, 0, "If not zero then maximum number of used HW threads is reduced by MinHwThreadsUnoccupied")
DECLARE_DEBUG_VARIABLE(int32_t, EventWaitOnHost, -1, "Wait for events on host instead of program semaphores for them, works for append kernel launch with immediate command list, -1: default, 0: disable, 1: enable")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideKernelSizeLimitForSmallDispatch, -1, "-1: default, >=0: on XEHP+ changes the threshold for treating kernel as small during NULL LWS selection")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideUseKmdWaitFunction, -1, "-1: default (L0: disabled), 0: disabled, 1: enabled. It uses only busy loop to wait or busy loop with KMD wait function, when KMD fallback is enabled")
DECLARE_DEBUG_VARIABLE(int32_t, MakeIndirectAllocationsResidentAsPack, -1, "-1: default, 0:disabled, 1: enabled. If enabled, driver handles all indirect allocations as one pack instead of making them resident individually.")
DECLARE_DEBUG_VARIABLE(int32_t, DetectIndirectAccessInKernel, -1, "-1: default, 0:disabled, 1: enabled. If enabled and indirect accesses are not detected in kernel, indirect allocations will not be allowed even if set by API.")
DECLARE_DEBUG_VARIABLE(int32_t, MakeEachAllocationResident, -1, "-1: default, 0: disabled, 1: bind every allocation at creation time, 2: bind all created allocations in flush")
DECLARE_DEBUG_VARIABLE(int32_t, PreferInternalBcsEngine, -1, "-1: default, 0:disabled, 1: enabled. When enabled use internal BCS engine for internal transfers, when disabled use regular engine")
DECLARE_DEBUG_VARIABLE(int32_t, SplitBcsCopy, -1, "-1: default, 0:disabled, 1: enabled. When enqueues copy to main copy engine then split between even linked copy engines")
DECLARE_DEBUG_VARIABLE(int32_t, SetAmountOfReusableAllocations, -1, "-1: default, 0:disabled, > 1: enabled. If enabled, driver will fill reusable allocation lists with given amount of command buffers and heaps at initialization of immediate command list.")
DECLARE_DEBUG_VARIABLE(int32_t, UseHighAlignmentForHeapExtended, -1, "-1: default, 0:disabled, > 1: enabled. If enabled, driver aligns HEAP_EXTENDED allocations to GPU VA that is next power of 2 for a given size, if disables GPU VA is using 2MB/64KB alignment.")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionDrmContext, -1, "Create special drm context: -1: default, when new residency model available, 0: disable, 1: enable")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionOverrideBlitterSupport, -1, "Overrides default blitter support: -1: do not override, 0: disable engine support, 1: enable engine support with init start, 2: enable engine support without init start")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionOverrideRenderSupport, -1, "Overrides default render support: -1: do not override, 0: disable engine support, 1: enable engine support with init start, 2: enable engine support without init start")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionOverrideComputeSupport, -1, "Overrides default compute support: -1: do not override, 0: disable engine support, 1: enable engine support with init start, 2: enable engine support without init start")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerMaxTimeout, -1, "Set direct submission controller max timeout - timeout will increase up to given value, -1: default 5000 us, >=0: max timeout in us")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionForceLocalMemoryStorageMode, -1, "Force local memory storage for command/ring/semaphore buffer, -1: default - for all engines, 0: disabled, 1: for multiOsContextCapable engine, 2: for all engines")
DECLARE_DEBUG_VARIABLE(int32_t, EnableRingSwitchTagUpdateWa, -1, "-1: default, 0 - disable, 1 - enable. If enabled, completionFences wont be updated if ring is not running.")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionMaxRingBuffers, -1, "-1: default, >0: max ring buffer count, During switch ring buffer, if there is no available ring, wait for completion instead of allocating new one if DirectSubmissionMaxRingBuffers is reached")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrdering, -1, "-1: default, 0 - disable, 1 - enable. If enabled, tasks sent to direct submission ring may be dispatched out of order")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrderingQueueSizeLimit, -1, "-1: default, >0: Max gpu queue size. If limit is reached, scheduler wont consume new work")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionRelaxedOrderingMinNumberOfClients, -1, "-1: default, >0: Enables RelaxedOrdering mode only if specified number of clients is assigned to given CSR.")
DECLARE_DEBUG_VARIABLE(bool, UseDeprecatedClDeviceIpVersion, false, "When enabled, the deprecated ip version scheme distinguishing between families and integrated devices will be queried in OCL")
DECLARE_DEBUG_VARIABLE(int32_t, ReturnSubDevicesAsApiDevices, -1, "Expose each subdevice as a separate device during clGetDeviceIDs or zeDeviceGet API call")
DECLARE_DEBUG_VARIABLE(int32_t, EnableCacheFlushAfterWalker, -1, "-1: platform behavior, 0: disabled, 1: enabled. Adds dedicated cache flush command after WALKER command when surfaces used by kernel require to flush the cache")
DECLARE_DEBUG_VARIABLE(int32_t, LimitAmountOfReturnedDevices, 0, "0: default - disable, 1+: Driver will limit the number of devices returned from clGetDeviceIds to N.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideEnableQuickKmdSleep, -1, "-1: don't override, 0: disable, 1: enable. It works only when Kmd Notify is enabled.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideEnableQuickKmdSleepForSporadicWaits, -1, "-1: don't override, 0: disable, 1: enable. It works only when QuickKmdSleep is enabled.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideDelayQuickKmdSleepForSporadicWaitsMicroseconds, -1, "-1: don't override, >0: timeout in microseconds")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideEnableQuickKmdSleepForDirectSubmission, -1, "-1: don't override, 0: disable, 1: enable. It works only when QuickKmdSleep is enabled.")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideDelayQuickKmdSleepForDirectSubmissionMicroseconds, -1, "-1: don't override, >0: timeout in microseconds")
DECLARE_DEBUG_VARIABLE(int32_t, PowerSavingMode, 0, "0: default 1: enable. Whenever driver waits on GPU and its not ready, put waiting thread to sleep and wait for notification.")
DECLARE_DEBUG_VARIABLE(int32_t, ForceFineGrainedSVMSupport, -1, "-1: default, 0: Do not report Fine Grained SVM capabilities 1: Report SVM Fine Grained capabilities if device supports SVM")
DECLARE_DEBUG_VARIABLE(int32_t, AlignLocalMemoryVaTo2MB, -1, "Allow 2MB pages for allocations with size>=2MB. On Linux it means aligned VA, on Windows it means aligned size. -1: default, 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, EnableUserFenceUseCtxId, -1, "-1: default (disabled), 0: disable, 1: enable : Use Context Id in Wait User Fence when waiting for completion tag")
DECLARE_DEBUG_VARIABLE(int32_t, OverrideNotifyEnableForTagUpdatePostSync, -1, "-1: default (usage determined by user fence wait call), 0: disable use of NotifyEnable flag, 1: enable use NotifyEnable flag")
DECLARE_DEBUG_VARIABLE(int32_t, UsePipeControlMultiKernelEventSync, -1, "Use single PIPE_CONTROL for event signal of multi-kernel append operations instead multi-packet POSTSYNC_DATA from each COMPUTE_WALKER, -1: default , 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, UseDynamicEventPacketsCount, -1, "Use dynamic estimation for event packet count based on a given device configuration, -1: default , 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, SignalAllEventPackets, -1, "All packets of event are signaled, reset and waited/synchronized, -1: default, 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, EnableBcsSwControlWa, -1, "Enable BCS WA via BCSSWCONTROL MMIO. -1: default, 0: disabled, 1: if src in system mem, 2: if dst in system mem, 3: if src and dst in system mem, 4: always")
DECLARE_DEBUG_VARIABLE(int32_t, EnableImplicitScaling, -1, "-1: default for API, 0: disable implicit scaling support on a given API, 1: enable implicit scaling support on a given API.")
DECLARE_DEBUG_VARIABLE(int32_t, SynchronizeWalkerInWparidMode, -1, "-1: default, 0: do not synchronize 1: synchronize all tiles prior to doing work distribution")
DECLARE_DEBUG_VARIABLE(int32_t, SynchronizeWithSemaphores, -1, "-1: default (disabled), 1: Emit Semaphores waiting after Walker completion in WPARID mode 0: do not emit semaphores after Walker")
DECLARE_DEBUG_VARIABLE(int32_t, UseAtomicsForSelfCleanupSection, -1, "-1: default (disabled), 0: use store data op, 1: use atomic op")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramWalkerPartitionSelfCleanup, -1, "-1: default (API dependent), 0: Do not program self cleanup, 1: program self cleanup")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramStallCommandForSelfCleanup, -1, "-1: default (disabled), 0: Do not add extra PIPE_CONTROL for each self cleanup x-tile sync, 1: Add extra PIPE_CONTROL for each self cleanup x-tile sync")
DECLARE_DEBUG_VARIABLE(int32_t, ExperimentalSetWalkerPartitionCount, 0, "Experimental implementation: Set number of COMPUTE_WALKERs for a given Partition Type, 0 - do not set the feature.")
DECLARE_DEBUG_VARIABLE(int32_t, ExperimentalSetWalkerPartitionType, -1, "Experimental implementation: Set COMPUTE_WALKER Partition Type. Valid values for types from 1 to 3")
DECLARE_DEBUG_VARIABLE(int32_t, ExperimentalEnableCustomLocalMemoryAlignment, 0, "Align local memory allocations to a given value. Works only with allocations at least as big as the value. 0: no effect, 2097152: 2 megabytes, 1073741824: 1 gigabyte")
DECLARE_DEBUG_VARIABLE(int32_t, ExperimentalForceCopyThroughLock, -1, "Force copy through lock pointer on zeAppendMemoryCopy for all cases -1: default 0: disable 1: enable ")
DECLARE_DEBUG_VARIABLE(int32_t, ExperimentalCopyThroughLockWaitlistSizeThreshold, -1, "If less than given value, driver will wait for Waitlist on host, instead of sending appendBarrier. If 0, always use barrier.")
DECLARE_DEBUG_VARIABLE(bool, ExperimentalEnableL0DebuggerForOpenCL, false, "Experimentally enable debugging OCL with L0 Debug API. When enabled - Level Zero debugging is disabled.")
DECLARE_DEBUG_VARIABLE(bool, UseMaxSimdSizeToDeduceMaxWorkgroupSize, false, "With this flag on, max workgroup size is deduced using SIMD32 instead of SIMD8, this causes the max wkg size to be 4 times bigger")
DECLARE_DEBUG_VARIABLE(bool, ReturnRawGpuTimestamps, false, "Driver returns raw GPU timestamps instead of calculated ones.")
DECLARE_DEBUG_VARIABLE(bool, SkipFlushingEventsOnGetStatusCalls, false, "When set to 1, events are not causing internal flush when querying for CL_EVENT_COMMAND_EXECUTION_STATUS")
DECLARE_DEBUG_VARIABLE(bool, AllowUnrestrictedSize, false, "Allow allocating memory with greater size than MAX_MEM_ALLOC_SIZE")
DECLARE_DEBUG_VARIABLE(bool, ForceDefaultThreadArbitrationPolicyIfNotSpecified, false, "When executing kernel without thread arbitration hint specified, ensure the default setting is used")
DECLARE_DEBUG_VARIABLE(bool, ForceAllResourcesUncached, false, "When set, all memory operations for all resources are forced to UC. This overrides all caching-related debug variables and globally disables all caches")
DECLARE_DEBUG_VARIABLE(bool, EnableCpuCacheForResources, false, "When true, driver will set gmm flag cacheable related to caching on cpu, for resources where it is allowed")
DECLARE_DEBUG_VARIABLE(bool, ForceDefaultGrfCompilationMode, false, "Adds build option -cl-intel-128-GRF-per-thread to force kernel compilation in Default-GRF mode")
DECLARE_DEBUG_VARIABLE(bool, ForceLargeGrfCompilationMode, false, "Adds build option -cl-intel-256-GRF-per-thread to force kernel compilation in Large-GRF mode")
DECLARE_DEBUG_VARIABLE(bool, EnableConcurrentSharedCrossP2PDeviceAccess, false, "Enables the concurrent use between host and peer devices of shared-allocations ")
DECLARE_DEBUG_VARIABLE(int32_t, EnableBOChunking, 0, "Enables use of chunking of BOs in the KMD, mask: 0 = no chunking, 1 = shared allocations only, 2 = device allocations only, 3 = shared and device allocations .")
DECLARE_DEBUG_VARIABLE(int32_t, MinimalAllocationSizeForChunking, -1, "2097152: default, >0: size in B. Minimal size an allocation should have to use chunking.")
DECLARE_DEBUG_VARIABLE(int32_t, ForceOCL21FeaturesSupport, -1, "-1: default, 0: disable, 1:enable. Force support of OpenCL 2.0 and OpenCL 2.1 API features")
DECLARE_DEBUG_VARIABLE(int32_t, ForceSemaphoreDelayBetweenWaits, -1, "Specifies the minimum number of microseconds allowed for command streamer to wait before re-fetching the data. 0 - poll interval will be equal to the memory latency of the read completion")
DECLARE_DEBUG_VARIABLE(int32_t, ForceLocalMemoryAccessMode, -1, "-1: don't override, 0: default rules apply, 1: CPU can access local memory, 3: CPU never accesses local memory")
DECLARE_DEBUG_VARIABLE(int64_t, ForceSystemMemoryPlacement, 0, "0: default, >0: (bitmask) for given Graphics Allocation Type, force system memory placement")
DECLARE_DEBUG_VARIABLE(int64_t, ForceNonSystemMemoryPlacement, 0, "0: default, >0: (bitmask) for given Graphics Allocation Type, force non-system memory placement")
DECLARE_DEBUG_VARIABLE(int64_t, ForceMultiTileAllocPlacement, 0, "0: default, >0: (bitmask) for given Graphics Allocation Type, force allocation to be replicated across all tiles")
DECLARE_DEBUG_VARIABLE(int64_t, ForceSingleTileAllocPlacement, 0, "0: default, >0: (bitmask) for given Graphics Allocation Type, force allocation to NOT be replicated")
DECLARE_DEBUG_VARIABLE(int32_t, UpdateTaskCountFromWait, -1, " Do not update task count after each enqueue, but send update request while wait, -1: default(disabled), 0: disabled, 1: enabled on gpgpu engine with direct submission, 2: enabled on any direct submission, 3: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, EnableTimestampWaitForQueues, -1, "Wait on queues using timestamps, -1: default(disabled), 0: disabled, 1: enabled where UpdateTaskCountFromWait enabled, 2: enabled on gpgpu engine with direct submission, 3: enabled on any direct submission, 4: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, EnableTimestampWaitForEvents, -1, "Wait on events using timestamps, -1: default(disabled), 0: disabled, 1: enabled where UpdateTaskCountFromWait enabled, 2: enabled on gpgpu engine with direct submission, 3: enabled on any direct submission, 4: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, ForceHostPointerImport, -1, "-1: default, 0: disable, 1: enable, Forces the driver to import every host pointer coming into driver, WARNING this is not spec compliant.")
DECLARE_DEBUG_VARIABLE(int32_t, ProgramExtendedPipeControlPriorToNonPipelinedStateCommand, -1, "-1: default, 0: disable, 1: enable, Program additional extended version of PIPE CONTROL command before non pipelined state command")
DECLARE_DEBUG_VARIABLE(int32_t, EnableFrontEndTracking, -1, "-1: default: enabled, 0: disabled, 1: enabled. This flag creates multiple return point from List to Queue for Front End reconfiguration on Queue buffer for single List")
DECLARE_DEBUG_VARIABLE(int32_t, EnablePipelineSelectTracking, -1, "-1: default: enabled, 0: disabled, 1: enabled. This flag enables optimization that limits number of pipeline select dispatched by command lists")
DECLARE_DEBUG_VARIABLE(int32_t, EnableStateComputeModeTracking, -1, "-1: default: enabled, 0: disabled, 1: enabled. This flag enables tracking state compute mode changes in command lists")
DECLARE_DEBUG_VARIABLE(int32_t, EnableStateBaseAddressTracking, -1, "-1: default: enabled, 0: disabled, 1: enabled. This flag enables tracking state base address changes in command lists")
DECLARE_DEBUG_VARIABLE(int32_t, EventTimestampRefreshIntervalInMilliSec, -1, "-1: use driver default, This value sets the refresh interval for getting synchronized GPU and CPU timestamp")
DECLARE_DEBUG_VARIABLE(int32_t, ForceDummyBlitWa, -1, "-1: default, 0: disabled, 1: enabled, Forces a workaround with dummy blits, driver adds an extra blit before command MI_ARB_CHECK on bcs")
