intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 06:49:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: Optimize intra-module kernel ISA allocations

Browse Source 
So far, there is a separate page allocated for each kernel's ISA within
`KernelImmutableData::initialize()`. Apparently the ISA blocks are often
much smaller than a 64k page, which leads to poor memory utilization and
was even observed to cause the device OOM error if a single module has
several keys.

Improve the situation by reusing the parent allocation (owned by the
module instance) for modules, which kernel ISAs can fit together within
a single 64k page. This improves the memory utilization on a single
module level.

Related-To: NEO-7788
Signed-off-by: Maciej Bielski <maciej.bielski@intel.com>
This commit is contained in:
Maciej Bielski
2023-08-09 13:39:39 +00:00
committed by Compute-Runtime-Automation
parent 1b7e178b25
commit c348831470
22 changed files with 649 additions and 177 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -77,7 +77,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
{
auto alloc = args.dispatchInterface->getIsaAllocation();
UNRECOVERABLE_IF(nullptr == alloc);
auto offset = alloc->getGpuAddressToPatch();
auto offset = alloc->getGpuAddressToPatch() + args.dispatchInterface->getIsaOffsetInParentAllocation();
idd.setKernelStartPointer(offset);
idd.setKernelStartPointerHigh(0u);
}

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

@@ -87,7 +87,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
{
auto alloc = args.dispatchInterface->getIsaAllocation();
UNRECOVERABLE_IF(nullptr == alloc);
auto offset = alloc->getGpuAddressToPatch();
auto offset = alloc->getGpuAddressToPatch() + args.dispatchInterface->getIsaOffsetInParentAllocation();
if (!localIdsGenerationByRuntime) {
offset += kernelDescriptor.entryPoints.skipPerThreadDataLoad;
}

5
shared/source/helpers/gfx_core_helper.h
View File

@@ -53,6 +53,7 @@ class GfxCoreHelper {
static std::unique_ptr<GfxCoreHelper> create(const GFXCORE_FAMILY gfxCoreFamily);
virtual size_t getMaxBarrierRegisterPerSlice() const = 0;
virtual size_t getPaddingForISAAllocation() const = 0;
virtual size_t getKernelIsaPointerAlignment() const = 0;
virtual uint32_t getComputeUnitsUsedForScratch(const RootDeviceEnvironment &rootDeviceEnvironment) const = 0;
virtual uint32_t getPitchAlignmentForImage(const RootDeviceEnvironment &rootDeviceEnvironment) const = 0;
virtual void adjustDefaultEngineType(HardwareInfo *pHwInfo, const ProductHelper &productHelper) = 0;
@@ -215,6 +216,10 @@ class GfxCoreHelperHw : public GfxCoreHelper {
size_t getPaddingForISAAllocation() const override;
size_t getKernelIsaPointerAlignment() const override {
return static_cast<size_t>(GfxFamily::cmdInitInterfaceDescriptorData.KERNELSTARTPOINTER_ALIGN_SIZE);
}
uint32_t getComputeUnitsUsedForScratch(const RootDeviceEnvironment &rootDeviceEnvironment) const override;
uint32_t getPitchAlignmentForImage(const RootDeviceEnvironment &rootDeviceEnvironment) const override;

1
shared/source/kernel/dispatch_kernel_encoder_interface.h
View File

@@ -40,6 +40,7 @@ struct DispatchKernelEncoderI {
virtual uint32_t getSurfaceStateHeapDataSize() const = 0;
virtual GraphicsAllocation *getIsaAllocation() const = 0;
virtual uint64_t getIsaOffsetInParentAllocation() const = 0;
virtual const uint8_t *getDynamicStateHeapData() const = 0;
virtual uint32_t getRequiredWorkgroupOrder() const = 0;

39
shared/test/unit_test/encoders/test_encode_dispatch_kernel.cpp
View File

@@ -997,6 +997,45 @@ HWTEST2_F(EncodeDispatchKernelTest, givenBindlessKernelWhenDispatchingKernelThen
EXPECT_NE(usedAfter, usedBefore);
}
HWTEST2_F(EncodeDispatchKernelTest, givenKernelsSharingISAParentAllocationsWhenProgrammingWalkerThenKernelStartPointerHasProperOffset, IsBeforeXeHpCore) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
auto dispatchInterface = std::make_unique<MockDispatchKernelEncoder>();
dispatchInterface->getIsaOffsetInParentAllocationResult = 8 << INTERFACE_DESCRIPTOR_DATA::KERNELSTARTPOINTER_BIT_SHIFT;
uint32_t dims[] = {2, 1, 1};
bool requiresUncachedMocs = false;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, requiresUncachedMocs);
EncodeDispatchKernel<FamilyType>::encode(*cmdContainer.get(), dispatchArgs);
auto idd = static_cast<INTERFACE_DESCRIPTOR_DATA *>(cmdContainer->getIddBlock());
EXPECT_EQ(idd->getKernelStartPointer(), dispatchInterface->getIsaAllocation()->getGpuAddressToPatch() + dispatchInterface->getIsaOffsetInParentAllocation());
}
HWTEST_F(EncodeDispatchKernelTest, givenKernelStartPointerAlignmentInInterfaceDescriptorWhenHelperGetterUsedThenCorrectValueReturned) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::KERNELSTARTPOINTER_ALIGN_SIZE, pDevice->getGfxCoreHelper().getKernelIsaPointerAlignment());
}
HWTEST2_F(EncodeDispatchKernelTest, givenKernelsSharingISAParentAllocationsWhenProgrammingWalkerThenKernelStartPointerHasProperOffset, IsAtLeastXeHpCore) {
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
auto dispatchInterface = std::make_unique<MockDispatchKernelEncoder>();
dispatchInterface->getIsaOffsetInParentAllocationResult = 8 << INTERFACE_DESCRIPTOR_DATA::KERNELSTARTPOINTER_BIT_SHIFT;
uint32_t dims[] = {2, 1, 1};
bool requiresUncachedMocs = false;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, requiresUncachedMocs);
EncodeDispatchKernel<FamilyType>::encode(*cmdContainer.get(), dispatchArgs);
GenCmdList commands;
CmdParse<FamilyType>::parseCommandBuffer(commands, ptrOffset(cmdContainer->getCommandStream()->getCpuBase(), 0), cmdContainer->getCommandStream()->getUsed());
auto itor = find<WALKER_TYPE *>(commands.begin(), commands.end());
ASSERT_NE(itor, commands.end());
auto walkerCmd = genCmdCast<WALKER_TYPE *>(*itor);
EXPECT_EQ(walkerCmd->getInterfaceDescriptor().getKernelStartPointer(), dispatchInterface->getIsaAllocation()->getGpuAddressToPatch() + dispatchInterface->getIsaOffsetInParentAllocation());
}
HWTEST_F(EncodeDispatchKernelTest, givenNonBindlessOrStatelessArgWhenDispatchingKernelThenSurfaceStateOffsetInCrossThreadDataIsNotPatched) {
using BINDING_TABLE_STATE = typename FamilyType::BINDING_TABLE_STATE;
using DataPortBindlessSurfaceExtendedMessageDescriptor = typename FamilyType::DataPortBindlessSurfaceExtendedMessageDescriptor;

1
shared/test/unit_test/mocks/mock_dispatch_kernel_encoder_interface.h
View File

@@ -55,5 +55,6 @@ struct MockDispatchKernelEncoder : public DispatchKernelEncoderI {
ADDMETHOD_CONST_NOBASE(getDynamicStateHeapData, const uint8_t *, nullptr, ());
ADDMETHOD_CONST_NOBASE(requiresGenerationOfLocalIdsByRuntime, bool, true, ());
ADDMETHOD_CONST_NOBASE(getSlmPolicy, SlmPolicy, SlmPolicy::SlmPolicyNone, ());
ADDMETHOD_CONST_NOBASE(getIsaOffsetInParentAllocation, uint64_t, 0lu, ());
};
} // namespace NEO