source/command_container/command_encoder_xehp_and_later.inl

37 template <typename Family>
38 void EncodeDispatchKernel<Family>::encode(CommandContainer &container,
42 …using SHARED_LOCAL_MEMORY_SIZE = typename Family::INTERFACE_DESCRIPTOR_DATA::SHARED_LOCAL_MEMORY_S…
43     using STATE_BASE_ADDRESS = typename Family::STATE_BASE_ADDRESS;
44     using MI_BATCH_BUFFER_END = typename Family::MI_BATCH_BUFFER_END;
45     using INLINE_DATA = typename Family::INLINE_DATA;
67         *bbEnd = Family::cmdInitBatchBufferEnd;
73 …if (PreambleHelper<Family>::isSpecialPipelineSelectModeChanged(container.lastPipelineSelectModeReq…
75         EncodeComputeMode<Family>::adjustPipelineSelect(container, kernelDescriptor);
78     WALKER_TYPE walkerCmd = Family::cmdInitGpgpuWalker;
82 …bool inlineDataProgramming = EncodeDispatchKernel<Family>::inlineDataProgrammingRequired(kernelDes…
97     EncodeDispatchKernel<Family>::programBarrierEnable(idd,
102         HwHelperHw<Family>::get().computeSlmValues(hwInfo, dispatchInterface->getSlmTotalSize()));
116 …bindingTablePointer = static_cast<uint32_t>(EncodeSurfaceState<Family>::pushBindingTableAndSurface…
125     PreemptionHelper::programInterfaceDescriptorDataPreemption<Family>(&idd, preemptionMode);
135         samplerStateOffset = EncodeStates<Family>::copySamplerState(
147 …EncodeDispatchKernel<Family>::adjustBindingTablePrefetch(idd, samplerCount, bindingTableStateCount…
192 …    EncodeIndirectParams<Family>::encode(container, gpuPtr, dispatchInterface, implicitArgsGpuPtr);
211         args.dcFlushEnable = MemorySynchronizationCommands<Family>::isDcFlushAllowed(true, hwInfo);
212         MemorySynchronizationCommands<Family>::addPipeControl(*container.getCommandStream(), args);
217 …   EncodeStateBaseAddress<Family>::encode(container, sbaCmd, statelessMocsIndex, useGlobalAtomics);
225     EncodeDispatchKernel<Family>::encodeThreadData(walkerCmd,
238     using POSTSYNC_DATA = typename Family::POSTSYNC_DATA;
253         if (MemorySynchronizationCommands<Family>::isDcFlushAllowed(true, hwInfo)) {
259         EncodeDispatchKernel<Family>::adjustTimestampPacket(walkerCmd, hwInfo);
264     EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, hwInfo);
266     EncodeDispatchKernel<Family>::appendAdditionalIDDFields(&idd, hwInfo, threadsPerThreadGroup,
270 …EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(hwInfo, walkerCmd, isCooperative ? Kern…
272     PreemptionHelper::applyPreemptionWaCmdsBegin<Family>(listCmdBufferStream, *device);
280         ImplicitScalingDispatch<Family>::dispatchCommands(*listCmdBufferStream,
295     PreemptionHelper::applyPreemptionWaCmdsEnd<Family>(listCmdBufferStream, *device);
298 template <typename Family>
299 inline void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const HardwareInfo &hwInfo, …
302 template <typename Family>
303 bool EncodeDispatchKernel<Family>::isRuntimeLocalIdsGenerationRequired(uint32_t activeChannels,
379 template <typename Family>
380 void EncodeDispatchKernel<Family>::encodeThreadData(WALKER_TYPE &walkerCmd,
442 template <typename Family>
443 size_t EncodeDispatchKernel<Family>::estimateEncodeDispatchKernelCmdsSize(Device *device, const Vec…
447     totalSize += PreemptionHelper::getPreemptionWaCsSize<Family>(*device);
448     totalSize += EncodeStates<Family>::getAdjustStateComputeModeSize();
449     totalSize += EncodeIndirectParams<Family>::getCmdsSizeForIndirectParams();
450     totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupCountIndirect();
451     totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupSizeIndirect();
454 …totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetWorkDimIndirect(dispatchInterface->get…
456             totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupCountIndirect();
457             totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupSizeIndirect();
458 …totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetWorkDimIndirect(dispatchInterface->get…
465 …totalSize += ImplicitScalingDispatch<Family>::getSize(true, staticPartitioning, device->getDeviceB…
471 template <typename Family>
472 void EncodeStateBaseAddress<Family>::setIohAddressForDebugger(NEO::Debugger::SbaAddresses &sbaAddre…
475 template <typename Family>
476 void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd…
479     EncodeStateBaseAddress<Family>::encode(container, sbaCmd, statelessMocsIndex, false);
482 template <typename Family>
483 void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd…
488     StateBaseAddressHelper<Family>::programStateBaseAddress(
519         auto cmd = Family::cmdInitStateBindingTablePoolAlloc;
525         *(typename Family::_3DSTATE_BINDING_TABLE_POOL_ALLOC *)buffer = cmd;
529 template <typename Family>
530 size_t EncodeStateBaseAddress<Family>::getRequiredSizeForStateBaseAddress(Device &device, CommandCo…
534     size_t size = sizeof(typename Family::STATE_BASE_ADDRESS);
536         size += sizeof(typename Family::STATE_BASE_ADDRESS);
540         size += sizeof(typename Family::_3DSTATE_BINDING_TABLE_POOL_ALLOC);
546 template <typename Family>
547 void EncodeComputeMode<Family>::programComputeModeCommand(LinearStream &csr, StateComputeModeProper…
548     using STATE_COMPUTE_MODE = typename Family::STATE_COMPUTE_MODE;
551     STATE_COMPUTE_MODE stateComputeMode = Family::cmdInitStateComputeMode;
558         maskBits |= Family::stateComputeModeForceNonCoherentMask;
563         maskBits |= Family::stateComputeModeLargeGrfModeMask;
568         maskBits |= Family::stateComputeModeForceDisableSupportMultiGpuAtomics;
573         maskBits |= Family::stateComputeModeForceDisableSupportMultiGpuPartialWrites;
582 template <typename Family>
583 void EncodeComputeMode<Family>::adjustPipelineSelect(CommandContainer &container, const NEO::Kernel…
584     using PIPELINE_SELECT = typename Family::PIPELINE_SELECT;
585     auto pipelineSelectCmd = Family::cmdInitPipelineSelect;
588 …PreambleHelper<Family>::appendProgramPipelineSelect(&pipelineSelectCmd, isSpecialModeSelected, con…
596 template <typename Family>
597 inline void EncodeMediaInterfaceDescriptorLoad<Family>::encode(CommandContainer &container) {
600 template <typename Family>
601 void EncodeMiFlushDW<Family>::appendMiFlushDw(MI_FLUSH_DW *miFlushDwCmd, const HardwareInfo &hwInfo…
606 template <typename Family>
607 void EncodeMiFlushDW<Family>::programMiFlushDwWA(LinearStream &commandStream) {
609     *miFlushDwCmd = Family::cmdInitMiFlushDw;
612 template <typename Family>
613 size_t EncodeMiFlushDW<Family>::getMiFlushDwWaSize() {
614     return sizeof(typename Family::MI_FLUSH_DW);
617 template <typename Family>
618 bool EncodeSurfaceState<Family>::doBindingTablePrefetch() {
622 template <typename Family>
623 void EncodeSurfaceState<Family>::encodeExtraBufferParams(EncodeSurfaceStateArgs &args) {
663     if (EncodeSurfaceState<Family>::isAuxModeEnabled(surfaceState, gmm)) {
683 template <typename Family>
684 inline void EncodeSurfaceState<Family>::setCoherencyType(R_SURFACE_STATE *surfaceState, COHERENCY_T…
688 template <typename Family>
689 void EncodeSempahore<Family>::programMiSemaphoreWait(MI_SEMAPHORE_WAIT *cmd,
694     MI_SEMAPHORE_WAIT localCmd = Family::cmdInitMiSemaphoreWait;
704 template <typename Family>
705 inline void EncodeWA<Family>::encodeAdditionalPipelineSelect(Device &device, LinearStream &stream, …
707 template <typename Family>
708 inline size_t EncodeWA<Family>::getAdditionalPipelineSelectSize(Device &device) {
712 template <typename Family>
713 inline void EncodeStoreMemory<Family>::programStoreDataImm(MI_STORE_DATA_IMM *cmdBuffer,
719     MI_STORE_DATA_IMM storeDataImmediate = Family::cmdInitStoreDataImm;