Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-15 13:01:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

Unify surface state programming logic related to implicit scaling

Browse Source 
OCL image surface state programming for Xe Hp core is now reusing logic
of EncodeSurfaceState helper

Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski
2022-01-24 15:42:16 +00:00
committed by Compute-Runtime-Automation
parent 52d636394c
commit 5e238dc7f1
9 changed files with 52 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
opencl/source/xe_hp_core/image_xe_hp_core.cpp
View File

@ -17,25 +17,12 @@ static auto gfxCore = IGFX_XE_HP_CORE;
template <>
void ImageHw<Family>::appendSurfaceStateParams(Family::RENDER_SURFACE_STATE *surfaceState, uint32_t rootDeviceIndex, bool useGlobalAtomics) {
auto imageCtxType = this->context->peekContextType();
bool enableMultiGpuPartialWrites = (imageCtxType != ContextType::CONTEXT_TYPE_SPECIALIZED) && (context->containsMultipleSubDevices(rootDeviceIndex));
bool enableMultiGpuAtomics = enableMultiGpuPartialWrites;
if (DebugManager.flags.EnableMultiGpuAtomicsOptimization.get()) {
enableMultiGpuAtomics &= useGlobalAtomics;
}
surfaceState->setDisableSupportForMultiGpuAtomics(!enableMultiGpuAtomics);
surfaceState->setDisableSupportForMultiGpuPartialWrites(!enableMultiGpuPartialWrites);
if (DebugManager.flags.ForceMultiGpuAtomics.get() != -1) {
surfaceState->setDisableSupportForMultiGpuAtomics(!!DebugManager.flags.ForceMultiGpuAtomics.get());
}
if (DebugManager.flags.ForceMultiGpuPartialWrites.get() != -1) {
surfaceState->setDisableSupportForMultiGpuPartialWrites(!!DebugManager.flags.ForceMultiGpuPartialWrites.get());
}
EncodeSurfaceStateArgs args{};
args.outMemory = surfaceState;
args.useGlobalAtomics = useGlobalAtomics;
args.areMultipleSubDevicesInContext = context->containsMultipleSubDevices(rootDeviceIndex);
args.implicitScaling = args.areMultipleSubDevicesInContext;
EncodeSurfaceState<Family>::encodeImplicitScalingParams(args);
}
} // namespace NEO
#include "opencl/source/mem_obj/image_tgllp_and_later.inl"

38
opencl/test/unit_test/xe_hp_core/xehp/test_image_xe_hp_sdv.inl
View File

@ -18,7 +18,6 @@
#include "opencl/test/unit_test/mocks/mock_platform.h"
using XeHpSdvImageTests = ::testing::Test;
using isXePlatform = IsWithinGfxCore<IGFX_XE_HP_CORE, IGFX_XE_HPC_CORE>;
XEHPTEST_F(XeHpSdvImageTests, givenContextTypeDefaultWhenImageIsWritableAndOnlyOneTileIsAvailableThenRemainFlagsToTrue) {
DebugManagerStateRestore restorer;
@ -247,27 +246,30 @@ XEHPTEST_F(XeHpSdvImageTests, givenContextTypeSpecializedWhenImageIsWritableThen
}
struct MultiGpuGlobalAtomicsImageTest : public XeHpSdvImageTests,
public ::testing::WithParamInterface<std::tuple<unsigned int, unsigned int, ContextType, bool, bool>> {
public ::testing::WithParamInterface<std::tuple<unsigned int, unsigned int, bool, bool>> {
};
XEHPTEST_P(MultiGpuGlobalAtomicsImageTest, givenAppendSurfaceStateParamCalledThenDisableSupportForMultiGpuAtomicsIsSetCorrectly) {
unsigned int numAvailableDevices, memFlags;
ContextType contextType;
using RENDER_SURFACE_STATE = typename FamilyType::RENDER_SURFACE_STATE;
unsigned int numAvailableSubDevices, memFlags;
bool useGlobalAtomics, enableMultiGpuAtomicsOptimization;
std::tie(numAvailableDevices, memFlags, contextType, useGlobalAtomics, enableMultiGpuAtomicsOptimization) = GetParam();
std::tie(numAvailableSubDevices, memFlags, useGlobalAtomics, enableMultiGpuAtomicsOptimization) = GetParam();
DebugManagerStateRestore restorer;
DebugManager.flags.EnableMultiGpuAtomicsOptimization.set(enableMultiGpuAtomicsOptimization);
DebugManager.flags.CreateMultipleSubDevices.set(numAvailableDevices);
initPlatform();
if (numAvailableDevices == 1) {
EXPECT_EQ(0u, platform()->getClDevice(0)->getNumGenericSubDevices());
} else {
EXPECT_EQ(numAvailableDevices, platform()->getClDevice(0)->getNumGenericSubDevices());
UltClDeviceFactory deviceFactory{1, 2};
ClDeviceVector deviceVector;
for (auto i = 0u; i < numAvailableSubDevices; i++) {
deviceVector.push_back(deviceFactory.subDevices[i]);
}
using RENDER_SURFACE_STATE = typename FamilyType::RENDER_SURFACE_STATE;
MockContext context(platform()->getClDevice(0));
context.contextType = contextType;
if (deviceVector.empty()) {
deviceVector.push_back(deviceFactory.rootDevices[0]);
}
MockContext context(deviceVector);
cl_int retVal = CL_SUCCESS;
cl_image_format imageFormat = {};
@ -294,23 +296,21 @@ XEHPTEST_P(MultiGpuGlobalAtomicsImageTest, givenAppendSurfaceStateParamCalledThe
surfaceState.setDisableSupportForMultiGpuPartialWrites(false);
imageHw->appendSurfaceStateParams(&surfaceState, context.getDevice(0)->getRootDeviceIndex(), useGlobalAtomics);
bool enableGlobalAtomics = (contextType != ContextType::CONTEXT_TYPE_SPECIALIZED) && (numAvailableDevices > 1);
bool enableGlobalAtomics = numAvailableSubDevices != 1u;
if (enableMultiGpuAtomicsOptimization) {
enableGlobalAtomics &= useGlobalAtomics;
}
EXPECT_EQ(!enableGlobalAtomics, surfaceState.getDisableSupportForMultiGpuAtomics());
}
static unsigned int numAvailableDevicesForMultiGpuGlobalAtomicsImageTest[] = {1, 2};
static unsigned int numAvailableSubDevicesForMultiGpuGlobalAtomicsImageTest[] = {0, 1, 2};
static unsigned int memFlags[] = {CL_MEM_READ_ONLY, CL_MEM_READ_WRITE};
static ContextType contextTypes[] = {ContextType::CONTEXT_TYPE_DEFAULT, ContextType::CONTEXT_TYPE_SPECIALIZED, ContextType::CONTEXT_TYPE_UNRESTRICTIVE};
INSTANTIATE_TEST_CASE_P(MultiGpuGlobalAtomicsImageTest,
MultiGpuGlobalAtomicsImageTest,
::testing::Combine(
::testing::ValuesIn(numAvailableDevicesForMultiGpuGlobalAtomicsImageTest),
::testing::ValuesIn(numAvailableSubDevicesForMultiGpuGlobalAtomicsImageTest),
::testing::ValuesIn(memFlags),
::testing::ValuesIn(contextTypes),
::testing::Bool(),
::testing::Bool()));

1
shared/source/command_container/command_encoder.h
View File

@ -257,6 +257,7 @@ struct EncodeSurfaceState {
static void encodeBuffer(EncodeSurfaceStateArgs &args);
static void encodeExtraBufferParams(EncodeSurfaceStateArgs &args);
static void encodeImplicitScalingParams(const EncodeSurfaceStateArgs &args);
static void encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo);
static void appendBufferSurfaceState(EncodeSurfaceStateArgs &args);

3
shared/source/command_container/command_encoder.inl
View File

@ -478,6 +478,9 @@ template <typename Family>
void EncodeSurfaceState<Family>::appendParamsForImageFromBuffer(R_SURFACE_STATE *surfaceState) {
}
template <typename Family>
void EncodeSurfaceState<Family>::encodeImplicitScalingParams(const EncodeSurfaceStateArgs &args) {}
template <typename Family>
void *EncodeDispatchKernel<Family>::getInterfaceDescriptor(CommandContainer &container, uint32_t &iddOffset) {

20
shared/source/command_container/command_encoder_xehp_and_later.inl
View File

@ -649,25 +649,7 @@ void EncodeSurfaceState<Family>::encodeExtraBufferParams(EncodeSurfaceStateArgs
encodeExtraCacheSettings(surfaceState, *args.gmmHelper->getHardwareInfo());
if constexpr (Family::isUsingMultiGpuProgrammingInSurfaceState) {
bool enablePartialWrites = args.implicitScaling;
bool enableMultiGpuAtomics = enablePartialWrites;
if (DebugManager.flags.EnableMultiGpuAtomicsOptimization.get()) {
enableMultiGpuAtomics = args.useGlobalAtomics && (enablePartialWrites || args.areMultipleSubDevicesInContext);
}
surfaceState->setDisableSupportForMultiGpuAtomics(!enableMultiGpuAtomics);
surfaceState->setDisableSupportForMultiGpuPartialWrites(!enablePartialWrites);
if (DebugManager.flags.ForceMultiGpuAtomics.get() != -1) {
surfaceState->setDisableSupportForMultiGpuAtomics(!!DebugManager.flags.ForceMultiGpuAtomics.get());
}
if (DebugManager.flags.ForceMultiGpuPartialWrites.get() != -1) {
surfaceState->setDisableSupportForMultiGpuPartialWrites(!!DebugManager.flags.ForceMultiGpuPartialWrites.get());
}
}
encodeImplicitScalingParams(args);
if (EncodeSurfaceState<Family>::isAuxModeEnabled(surfaceState, gmm)) {
auto resourceFormat = gmm->gmmResourceInfo->getResourceFormat();

22
shared/source/xe_hp_core/command_encoder_xe_hp_core.cpp
View File

@ -31,6 +31,28 @@ template <>
inline void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo) {
}
template <>
void EncodeSurfaceState<Family>::encodeImplicitScalingParams(const EncodeSurfaceStateArgs &args) {
auto surfaceState = reinterpret_cast<R_SURFACE_STATE *>(args.outMemory);
bool enablePartialWrites = args.implicitScaling;
bool enableMultiGpuAtomics = enablePartialWrites;
if (DebugManager.flags.EnableMultiGpuAtomicsOptimization.get()) {
enableMultiGpuAtomics = args.useGlobalAtomics && (enablePartialWrites || args.areMultipleSubDevicesInContext);
}
surfaceState->setDisableSupportForMultiGpuAtomics(!enableMultiGpuAtomics);
surfaceState->setDisableSupportForMultiGpuPartialWrites(!enablePartialWrites);
if (DebugManager.flags.ForceMultiGpuAtomics.get() != -1) {
surfaceState->setDisableSupportForMultiGpuAtomics(!!DebugManager.flags.ForceMultiGpuAtomics.get());
}
if (DebugManager.flags.ForceMultiGpuPartialWrites.get() != -1) {
surfaceState->setDisableSupportForMultiGpuPartialWrites(!!DebugManager.flags.ForceMultiGpuPartialWrites.get());
}
}
template <>
void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTOR_DATA *pInterfaceDescriptor, const HardwareInfo &hwInfo, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize, SlmPolicy slmPolicy) {
}

1
shared/source/xe_hp_core/hw_cmds_base.h
View File

@ -31,7 +31,6 @@ struct XeHpCore {
static constexpr bool isUsingMediaSamplerDopClockGate = true;
static constexpr bool supportsSampler = true;
static constexpr bool isUsingGenericMediaStateClear = true;
static constexpr bool isUsingMultiGpuProgrammingInSurfaceState = true;
struct DataPortBindlessSurfaceExtendedMessageDescriptor {
union {

1
shared/source/xe_hpc_core/hw_cmds_base.h
View File

@ -37,7 +37,6 @@ struct XE_HPC_CORE {
static constexpr bool isUsingMediaSamplerDopClockGate = false;
static constexpr bool supportsSampler = false;
static constexpr bool isUsingGenericMediaStateClear = true;
static constexpr bool isUsingMultiGpuProgrammingInSurfaceState = false;
static bool isXlA0(const HardwareInfo &hwInfo) {
auto revId = hwInfo.platform.usRevId & pvcSteppingBits;

1
shared/source/xe_hpg_core/hw_cmds_base.h
View File

@ -31,7 +31,6 @@ struct XE_HPG_CORE {
static constexpr bool isUsingMediaSamplerDopClockGate = false;
static constexpr bool supportsSampler = true;
static constexpr bool isUsingGenericMediaStateClear = true;
static constexpr bool isUsingMultiGpuProgrammingInSurfaceState = false;
struct DataPortBindlessSurfaceExtendedMessageDescriptor {
union {