Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

feature: Enable eu thread scheduling mode override via IDD 

Enables eu thread scheduling mode override via the
INTERFACE_DESCRIPTOR_DATA and COMPUTE_WALKER.

Defines encodeEuSchedulingPolicy operating on INTERFACE_DESCRIPTOR_DATA
or Xe3 platform.

Adds tests verifying that thread scheduling mode can be overridden via
COMPUTE_WALKER.

Related-To: NEO-13771, HSD-18041256338, HSD-14013056398

Signed-off-by: Vysochyn, Illia <illia.vysochyn@intel.com>
This commit is contained in:
Vysochyn, Illia 2025-02-25 14:11:01 +00:00 committed by Compute-Runtime-Automation
parent a9da0c4ed5
commit 1eafbc30f8
5 changed files with 87 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
shared/source/command_container/command_encoder_from_xe_hpg_core_to_xe2_hpg.inl
View File

@ -1,5 +1,5 @@
/*
* Copyright (C) 2024 Intel Corporation
* Copyright (C) 2024-2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@ -22,4 +22,9 @@ bool EncodeEnableRayTracing<Family>::is48bResourceNeededForRayTracing() {
return true;
}
template <typename Family>
template <typename InterfaceDescriptorType>
void EncodeDispatchKernel<Family>::encodeEuSchedulingPolicy(InterfaceDescriptorType *pInterfaceDescriptor, const KernelDescriptor &kernelDesc, int32_t defaultPipelinedThreadArbitrationPolicy) {
}
} // namespace NEO

5
shared/source/command_container/command_encoder_from_xe_hpg_core_to_xe3_core.inl
View File

@ -37,11 +37,6 @@ template <bool isHeapless>
void EncodeDispatchKernel<Family>::setScratchAddress(uint64_t &scratchAddress, uint32_t requiredScratchSlot0Size, uint32_t requiredScratchSlot1Size, IndirectHeap *ssh, CommandStreamReceiver &submissionCsr) {
}
template <typename Family>
template <typename InterfaceDescriptorType>
void EncodeDispatchKernel<Family>::encodeEuSchedulingPolicy(InterfaceDescriptorType *pInterfaceDescriptor, const KernelDescriptor &kernelDesc, int32_t defaultPipelinedThreadArbitrationPolicy) {
}
template <typename Family>
bool EncodeDispatchKernel<Family>::singleTileExecImplicitScalingRequired(bool cooperativeKernel) {
return cooperativeKernel;

30
shared/source/xe3_core/command_encoder_xe3_core.cpp
View File

@ -18,6 +18,8 @@
#include "shared/source/helpers/cache_policy.h"
#include "shared/source/helpers/constants.h"
#include "shared/source/kernel/grf_config.h"
#include "shared/source/kernel/kernel_descriptor.h"
#include "shared/source/os_interface/product_helper.h"
#include "shared/source/release_helper/release_helper.h"
#include "shared/source/xe3_core/hw_cmds_base.h"
@ -128,6 +130,34 @@ void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDevice
}
}
template <typename Family>
template <typename InterfaceDescriptorType>
void EncodeDispatchKernel<Family>::encodeEuSchedulingPolicy(InterfaceDescriptorType *pInterfaceDescriptor, const KernelDescriptor &kernelDesc, int32_t defaultPipelinedThreadArbitrationPolicy) {
using INTERFACE_DESCRIPTOR_DATA = typename Family::INTERFACE_DESCRIPTOR_DATA;
if constexpr (std::is_same_v<InterfaceDescriptorType, INTERFACE_DESCRIPTOR_DATA>) {
auto pipelinedThreadArbitrationPolicy = kernelDesc.kernelAttributes.threadArbitrationPolicy;
if (pipelinedThreadArbitrationPolicy == ThreadArbitrationPolicy::NotPresent) {
pipelinedThreadArbitrationPolicy = static_cast<ThreadArbitrationPolicy>(defaultPipelinedThreadArbitrationPolicy);
}
switch (pipelinedThreadArbitrationPolicy) {
case ThreadArbitrationPolicy::RoundRobin:
pInterfaceDescriptor->setEuThreadSchedulingModeOverride(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_ROUND_ROBIN);
break;
case ThreadArbitrationPolicy::AgeBased:
pInterfaceDescriptor->setEuThreadSchedulingModeOverride(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_OLDEST_FIRST);
break;
case ThreadArbitrationPolicy::RoundRobinAfterDependency:
pInterfaceDescriptor->setEuThreadSchedulingModeOverride(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_STALL_BASED_ROUND_ROBIN);
break;
default:
pInterfaceDescriptor->setEuThreadSchedulingModeOverride(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_HW_DEFAULT);
}
}
}
template <>
bool EncodeEnableRayTracing<Family>::is48bResourceNeededForRayTracing() {
if (debugManager.flags.Enable64bAddressingForRayTracing.get() != -1) {

15
shared/test/unit_test/command_container/command_encoder_tests.cpp
View File

@ -893,7 +893,20 @@ HWTEST2_F(CommandEncoderTests, whenAskingForImplicitScalingValuesThenAlwaysRetur
EXPECT_FALSE(ImplicitScalingDispatch<FamilyType>::platformSupportsImplicitScaling(*rootExecEnv));
}
HWTEST2_F(CommandEncoderTests, givenInterfaceDescriptorWhenEncodeEuSchedulingPolicyIsCalledThenNothingIsChanged, HeapfulSupportedMatch) {
HWTEST2_F(CommandEncoderTests, givenInterfaceDescriptorWhenEncodeEuSchedulingPolicyIsCalledThenChanged, IsAtLeastXe3Core) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
INTERFACE_DESCRIPTOR_DATA idd = FamilyType::cmdInitInterfaceDescriptorData;
KernelDescriptor kernelDescriptor;
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::AgeBased;
int32_t defaultPipelinedThreadArbitrationPolicy = ThreadArbitrationPolicy::RoundRobin;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(idd.getEuThreadSchedulingModeOverride(), INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_OLDEST_FIRST);
}
HWTEST2_F(CommandEncoderTests, givenInterfaceDescriptorWhenEncodeEuSchedulingPolicyIsCalledThenNothingIsChanged, IsAtMostXe2HpgCore) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;

37
shared/test/unit_test/xe3_core/command_encoder_xe3_core_tests.cpp
View File

@ -195,3 +195,40 @@ XE3_CORETEST_F(CommandEncodeStatesXe3Test, givenHeapSharingEnabledWhenRetrieving
itorCmd = find<_3DSTATE_BINDING_TABLE_POOL_ALLOC *>(commands.begin(), commands.end());
EXPECT_EQ(commands.end(), itorCmd);
}
XE3_CORETEST_F(Xe3CoreCommandEncoderTest, givenPipelinedEuThreadArbitrationPolicyWhenEncodeEuSchedulingPolicyIsCalledThenIddContainsCorrectEuSchedulingPolicy) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
INTERFACE_DESCRIPTOR_DATA idd = FamilyType::cmdInitInterfaceDescriptorData;
KernelDescriptor kernelDescriptor;
int32_t defaultPipelinedThreadArbitrationPolicy = ThreadArbitrationPolicy::NotPresent;
{
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::NotPresent;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_HW_DEFAULT, idd.getEuThreadSchedulingModeOverride());
}
defaultPipelinedThreadArbitrationPolicy = ThreadArbitrationPolicy::RoundRobin;
{
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::NotPresent;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_ROUND_ROBIN, idd.getEuThreadSchedulingModeOverride());
}
{
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::AgeBased;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_OLDEST_FIRST, idd.getEuThreadSchedulingModeOverride());
}
{
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::RoundRobin;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_ROUND_ROBIN, idd.getEuThreadSchedulingModeOverride());
}
{
kernelDescriptor.kernelAttributes.threadArbitrationPolicy = ThreadArbitrationPolicy::RoundRobinAfterDependency;
EncodeDispatchKernel<FamilyType>::encodeEuSchedulingPolicy(&idd, kernelDescriptor, defaultPipelinedThreadArbitrationPolicy);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::EU_THREAD_SCHEDULING_MODE_OVERRIDE::EU_THREAD_SCHEDULING_MODE_OVERRIDE_STALL_BASED_ROUND_ROBIN, idd.getEuThreadSchedulingModeOverride());
}
}