intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 23:03:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: Add heapless mode programming in ocl

Browse Source 
Related-To: NEO-7621
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
This commit is contained in:
Kamil Kopryk
2023-11-23 13:58:58 +00:00
committed by Compute-Runtime-Automation
parent c35b13ccae
commit ce7298d512
82 changed files with 1927 additions and 1224 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
shared/source/xe_hpg_core/command_encoder_xe_hpg_core.cpp
View File

@@ -26,15 +26,17 @@ using Family = NEO::XeHpgCoreFamily;
namespace NEO {
template <>
void EncodeDispatchKernel<Family>::adjustTimestampPacket(WALKER_TYPE &walkerCmd, const HardwareInfo &hwInfo) {
template <typename WalkerType>
void EncodeDispatchKernel<Family>::adjustTimestampPacket(WalkerType &walkerCmd, const HardwareInfo &hwInfo) {
auto &postSyncData = walkerCmd.getPostSync();
postSyncData.setDataportSubsliceCacheFlush(true);
}
template <>
void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTOR_DATA *pInterfaceDescriptor, const RootDeviceEnvironment &rootDeviceEnvironment, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize, SlmPolicy slmPolicy) {
using PREFERRED_SLM_ALLOCATION_SIZE = typename Family::INTERFACE_DESCRIPTOR_DATA::PREFERRED_SLM_ALLOCATION_SIZE;
template <typename InterfaceDescriptorType>
void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(InterfaceDescriptorType *pInterfaceDescriptor, const RootDeviceEnvironment &rootDeviceEnvironment, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize, SlmPolicy slmPolicy) {
using PREFERRED_SLM_ALLOCATION_SIZE = typename InterfaceDescriptorType::PREFERRED_SLM_ALLOCATION_SIZE;
auto &hwInfo = *rootDeviceEnvironment.getHardwareInfo();
const uint32_t threadsPerDssCount = hwInfo.gtSystemInfo.ThreadCount / hwInfo.gtSystemInfo.DualSubSliceCount;
const uint32_t workGroupCountPerDss = threadsPerDssCount / threadsPerThreadGroup;
@@ -92,7 +94,8 @@ void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTO
}
template <>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WALKER_TYPE &walkerCmd) {
template <typename WalkerType, typename InterfaceDescriptorType>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(InterfaceDescriptorType &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WalkerType &walkerCmd) {
const auto &productHelper = device.getProductHelper();
if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
if (interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup() == 1) {
@@ -108,6 +111,7 @@ void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCR
}
}
template <>
template <>
void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, uint32_t value, const HardwareInfo &hwInfo) {
using BARRIERS = INTERFACE_DESCRIPTOR_DATA::NUMBER_OF_BARRIERS;
@@ -118,7 +122,8 @@ void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DAT
}
template <>
void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WALKER_TYPE &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
template <typename WalkerType>
void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
auto *releaseHelper = rootDeviceEnvironment.getReleaseHelper();
bool l3PrefetchDisable = releaseHelper->isPrefetchDisablingRequired();
int32_t overrideL3PrefetchDisable = DebugManager.flags.ForceL3PrefetchForComputeWalker.get();
@@ -217,6 +222,13 @@ size_t EncodeMiFlushDW<Family>::getWaSize(const EncodeDummyBlitWaArgs &waArgs) {
template void flushGpuCache<Family>(LinearStream *commandStream, const Range<L3Range> &ranges, uint64_t postSyncAddress, const HardwareInfo &hwInfo);
template struct EncodeDispatchKernel<Family>;
template void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields<Family::WALKER_TYPE>(const RootDeviceEnvironment &rootDeviceEnvironment, Family::WALKER_TYPE &walkerCmd, const EncodeWalkerArgs &walkerArgs);
template void EncodeDispatchKernel<Family>::adjustTimestampPacket<Family::WALKER_TYPE>(Family::WALKER_TYPE &walkerCmd, const HardwareInfo &hwInfo);
template void EncodeDispatchKernel<Family>::setGrfInfo<Family::INTERFACE_DESCRIPTOR_DATA>(Family::INTERFACE_DESCRIPTOR_DATA *pInterfaceDescriptor, uint32_t numGrf, const size_t &sizeCrossThreadData, const size_t &sizePerThreadData, const RootDeviceEnvironment &rootDeviceEnvironment);
template void EncodeDispatchKernel<Family>::appendAdditionalIDDFields<Family::INTERFACE_DESCRIPTOR_DATA>(Family::INTERFACE_DESCRIPTOR_DATA *pInterfaceDescriptor, const RootDeviceEnvironment &rootDeviceEnvironment, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize, SlmPolicy slmPolicy);
template void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData<Family::WALKER_TYPE, Family::INTERFACE_DESCRIPTOR_DATA>(Family::INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, Family::WALKER_TYPE &walkerCmd);
template void EncodeDispatchKernel<Family>::setupPostSyncMocs<Family::WALKER_TYPE>(Family::WALKER_TYPE &walkerCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool dcFlush);
template struct EncodeStates<Family>;
template struct EncodeMath<Family>;
template struct EncodeMathMMIO<Family>;

20
shared/source/xe_hpg_core/hw_cmds_xe_hpg_core_base.h
View File

@@ -149,6 +149,26 @@ struct XeHpgCoreFamily : public XeHpgCore {
static constexpr bool supportsCmdSet(GFXCORE_FAMILY cmdSetBaseFamily) {
return cmdSetBaseFamily == IGFX_XE_HP_CORE;
}
template <typename WalkerType = WALKER_TYPE>
static WalkerType getInitGpuWalker() {
return cmdInitGpgpuWalker;
}
template <typename WalkerType = WALKER_TYPE>
static constexpr size_t getInterfaceDescriptorSize() {
return sizeof(INTERFACE_DESCRIPTOR_DATA);
}
template <typename InterfaceDescriptorType>
static InterfaceDescriptorType getInitInterfaceDescriptor() {
return cmdInitInterfaceDescriptorData;
}
template <typename WalkerType>
static constexpr bool isHeaplessMode() {
return false;
}
};
} // namespace NEO

5
shared/source/xe_hpg_core/implicit_scaling_xe_hpg_core.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2021-2022 Intel Corporation
* Copyright (C) 2021-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -17,4 +17,7 @@ template <>
bool ImplicitScalingDispatch<Family>::pipeControlStallRequired = true;
template struct ImplicitScalingDispatch<Family>;
template void ImplicitScalingDispatch<Family>::dispatchCommands<Family::WALKER_TYPE>(LinearStream &commandStream, Family::WALKER_TYPE &walkerCmd, void **outWalkerPtr, const DeviceBitfield &devices, uint32_t &partitionCount, bool useSecondaryBatchBuffer, bool apiSelfCleanup, bool usesImages, bool dcFlush, bool forceExecutionOnSingleTile, uint64_t workPartitionAllocationGpuVa, const HardwareInfo &hwInfo);
template size_t ImplicitScalingDispatch<Family>::getSize<Family::WALKER_TYPE>(bool apiSelfCleanup, bool preferStaticPartitioning, const DeviceBitfield &devices, const Vec3<size_t> &groupStart, const Vec3<size_t> &groupCount);
} // namespace NEO