intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-30 09:58:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: improve implicit scaling interface

Browse Source 
Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2024-01-19 11:54:53 +00:00
committed by Compute-Runtime-Automation
parent 93e941fd73
commit b266f1f3cc
8 changed files with 40 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_2.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2020-2023 Intel Corporation
* Copyright (C) 2020-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -1591,7 +1591,14 @@ HWTEST2_F(MultiTileCommandListAppendLaunchKernelXeHpCoreTest, givenCooperativeKe
itorWalker = find<typename FamilyType::DefaultWalkerType *>(cmdList.begin(), cmdList.end());
cmd = genCmdCast<typename FamilyType::DefaultWalkerType *>(*itorWalker);
EXPECT_TRUE(cmd->getWorkloadPartitionEnable());
EXPECT_EQ(16u, cmd->getPartitionSize());
const auto &gfxCoreHelper = device->getGfxCoreHelper();
if (gfxCoreHelper.singleTileExecImplicitScalingRequired(true)) {
EXPECT_EQ(16u, cmd->getPartitionSize());
} else {
EXPECT_EQ(4u, cmd->getPartitionSize());
}
}
HWTEST2_F(MultiTileCommandListAppendLaunchKernelXeHpCoreTest,

4
opencl/source/kernel/kernel.cpp
View File

@@ -1264,7 +1264,9 @@ bool Kernel::hasRunFinished(TimestampPacketContainer *timestampContainer) {
}
bool Kernel::isSingleSubdevicePreferred() const {
return this->singleSubdevicePreferredInCurrentEnqueue || this->usesSyncBuffer();
auto &gfxCoreHelper = this->getGfxCoreHelper();
return this->singleSubdevicePreferredInCurrentEnqueue || gfxCoreHelper.singleTileExecImplicitScalingRequired(this->usesSyncBuffer());
}
void Kernel::setInlineSamplers() {

4
opencl/test/unit_test/kernel/kernel_tests.cpp
View File

@@ -380,6 +380,8 @@ TEST_F(KernelTests, GivenInvalidParamNameWhenGettingWorkGroupInfoThenInvalidValu
}
TEST_F(KernelTests, WhenIsSingleSubdevicePreferredIsCalledThenCorrectValuesAreReturned) {
auto &helper = pClDevice->getGfxCoreHelper();
std::unique_ptr<MockKernel> kernel{MockKernel::create<MockKernel>(pClDevice->getDevice(), pProgram)};
for (auto usesSyncBuffer : ::testing::Bool()) {
kernel->getAllocatedKernelInfo()->kernelDescriptor.kernelAttributes.flags.usesSyncBuffer = usesSyncBuffer;
@@ -387,7 +389,7 @@ TEST_F(KernelTests, WhenIsSingleSubdevicePreferredIsCalledThenCorrectValuesAreRe
kernel->singleSubdevicePreferredInCurrentEnqueue = singleSubdevicePreferredInCurrentEnqueue;
EXPECT_EQ(usesSyncBuffer, kernel->usesSyncBuffer());
auto expectedSingleSubdevicePreferredInCurrentEnqueue = singleSubdevicePreferredInCurrentEnqueue || usesSyncBuffer;
auto expectedSingleSubdevicePreferredInCurrentEnqueue = singleSubdevicePreferredInCurrentEnqueue || helper.singleTileExecImplicitScalingRequired(usesSyncBuffer);
EXPECT_EQ(expectedSingleSubdevicePreferredInCurrentEnqueue, kernel->isSingleSubdevicePreferred());
}
}

2
shared/source/command_container/command_encoder_xehp_and_later.inl
View File

@@ -403,7 +403,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
!(container.getFlushTaskUsedForImmediate() || container.isUsingPrimaryBuffer()),
!args.isKernelDispatchedFromImmediateCmdList,
args.dcFlushEnable,
args.isCooperative,
gfxCoreHelper.singleTileExecImplicitScalingRequired(args.isCooperative),
workPartitionAllocationGpuVa,
hwInfo);
} else {

4
shared/source/helpers/gfx_core_helper.h
View File

@@ -182,6 +182,8 @@ class GfxCoreHelper {
virtual bool is48ResourceNeededForCmdBuffer() const = 0;
virtual uint32_t getKernelPrivateMemSize(const KernelDescriptor &kernelDescriptor) const = 0;
virtual bool singleTileExecImplicitScalingRequired(bool cooperativeKernel) const = 0;
virtual ~GfxCoreHelper() = default;
protected:
@@ -402,6 +404,8 @@ class GfxCoreHelperHw : public GfxCoreHelper {
uint32_t getKernelPrivateMemSize(const KernelDescriptor &kernelDescriptor) const override;
bool singleTileExecImplicitScalingRequired(bool cooperativeKernel) const override;
~GfxCoreHelperHw() override = default;
protected:

5
shared/source/helpers/gfx_core_helper_base.inl
View File

@@ -740,4 +740,9 @@ bool GfxCoreHelperHw<GfxFamily>::is48ResourceNeededForCmdBuffer() const {
return true;
}
template <typename GfxFamily>
bool GfxCoreHelperHw<GfxFamily>::singleTileExecImplicitScalingRequired(bool cooperativeKernel) const {
return cooperativeKernel;
}
} // namespace NEO

9
shared/test/unit_test/encoders/test_encode_dispatch_kernel_xehp_and_later.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2021-2023 Intel Corporation
* Copyright (C) 2021-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -1084,7 +1084,14 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesImplicitScaling, givenCooperativ
auto partitionWalkerCmd = genCmdCast<DefaultWalkerType *>(*itor);
EXPECT_EQ(DefaultWalkerType::PARTITION_TYPE::PARTITION_TYPE_X, partitionWalkerCmd->getPartitionType());
const auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
uint32_t expectedPartitionSize = dims[0];
if (!gfxCoreHelper.singleTileExecImplicitScalingRequired(isCooperative)) {
expectedPartitionSize /= dispatchArgs.partitionCount;
}
EXPECT_EQ(expectedPartitionSize, partitionWalkerCmd->getPartitionSize());
}

7
shared/test/unit_test/helpers/gfx_core_helper_tests.cpp
View File

@@ -1657,6 +1657,13 @@ TEST_F(GfxCoreHelperTest, whenOnlyPerThreadPrivateMemorySizeIsDefinedThenItIsRet
EXPECT_EQ(0x100u, getHelper<GfxCoreHelper>().getKernelPrivateMemSize(kernelDescriptor));
}
HWTEST_F(GfxCoreHelperTest, givenCooperativeKernelWhenAskingForSingleTileDispatchThenReturnTrue) {
auto &helper = getHelper<GfxCoreHelper>();
EXPECT_TRUE(helper.singleTileExecImplicitScalingRequired(true));
EXPECT_FALSE(helper.singleTileExecImplicitScalingRequired(false));
}
HWTEST2_F(GfxCoreHelperTest, whenPrivateScratchSizeIsDefinedThenItIsReturnedAsKernelPrivateMemorySize, IsAtLeastXeHpCore) {
KernelDescriptor kernelDescriptor{};
kernelDescriptor.kernelAttributes.perHwThreadPrivateMemorySize = 0x100u;