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

Fail build program on PVC with stateful accesses

Browse Source 
Related-To: NEO-6075

After this change driver will fail clBuildProgram/zeModuleCreate api calls
whenever stateful access is discovered on PVC.
This is required since in this case allocation greater than 4GB
will not work.
If user still wants to use stateful addressing mode,
-cl-opt-smaller-than-4GB-buffers-only / -ze-opt-smaller-than-4GB-buffers-only
build option should be passed as build option, but then user can not use
bufers greater than 4GB.


Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
This commit is contained in:
Kamil Kopryk
2022-02-02 00:21:00 +00:00
committed by Compute-Runtime-Automation
parent 4e31612c31
commit 9466113cef
13 changed files with 419 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
level_zero/core/source/module/module_imp.cpp
View File

@ -16,6 +16,7 @@
#include "shared/source/device_binary_format/elf/elf.h"
#include "shared/source/device_binary_format/elf/elf_encoder.h"
#include "shared/source/device_binary_format/elf/ocl_elf.h"
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/helpers/api_specific_config.h"
#include "shared/source/helpers/constants.h"
#include "shared/source/helpers/kernel_helpers.h"
@ -43,6 +44,7 @@ namespace BuildOptions {
NEO::ConstStringRef optDisable = "-ze-opt-disable";
NEO::ConstStringRef optLevel = "-ze-opt-level";
NEO::ConstStringRef greaterThan4GbRequired = "-ze-opt-greater-than-4GB-buffer-required";
NEO::ConstStringRef smallerThan4GbBuffersOnly = "-ze-opt-smaller-than-4GB-buffers-only";
NEO::ConstStringRef hasBufferOffsetArg = "-ze-intel-has-buffer-offset-arg";
NEO::ConstStringRef debugKernelEnable = "-ze-kernel-debug-enable";
} // namespace BuildOptions
@ -124,8 +126,10 @@ std::string ModuleTranslationUnit::generateCompilerOptions(const char *buildOpti
internalOptions = NEO::CompilerOptions::concatenate(internalOptions, BuildOptions::debugKernelEnable);
}
if (NEO::DebugManager.flags.DisableStatelessToStatefulOptimization.get() ||
device->getNEODevice()->areSharedSystemAllocationsAllowed()) {
auto disableStatelessToStatefulOptimization = NEO::DebugManager.flags.DisableStatelessToStatefulOptimization.get();
auto isForceToStatelessNeeded = NEO::AddressingModeHelper::forceToStatelessNeeded(options, BuildOptions::smallerThan4GbBuffersOnly.str(), device->getHwInfo());
if (disableStatelessToStatefulOptimization || isForceToStatelessNeeded) {
internalOptions = NEO::CompilerOptions::concatenate(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired);
}
@ -523,6 +527,14 @@ bool ModuleImp::initialize(const ze_module_desc_t *desc, NEO::Device *neoDevice)
this->updateBuildLog(neoDevice);
verifyDebugCapabilities();
auto containsStatefulAccess = NEO::AddressingModeHelper::containsStatefulAccess(translationUnit->programInfo.kernelInfos);
auto isForceToStatelessNeeded = NEO::AddressingModeHelper::forceToStatelessNeeded(translationUnit->options, BuildOptions::smallerThan4GbBuffersOnly.str(), device->getHwInfo());
auto isUserKernel = (type == ModuleType::User);
if (containsStatefulAccess && isForceToStatelessNeeded && isUserKernel) {
success = false;
}
if (false == success) {
return false;
}

174
level_zero/core/test/unit_tests/sources/module/test_module.cpp
View File

@ -10,6 +10,8 @@
#include "shared/source/device_binary_format/debug_zebin.h"
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/gmm_helper/gmm_helper.h"
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/helpers/compiler_hw_info_config.h"
#include "shared/source/kernel/implicit_args.h"
#include "shared/source/program/kernel_info.h"
#include "shared/test/common/helpers/debug_manager_state_restore.h"
@ -1783,28 +1785,19 @@ HWTEST_F(ModuleTranslationUnitTest, WhenBuildOptionsAreNullThenReuseExistingOpti
EXPECT_NE(pMockCompilerInterface->inputInternalOptions.find("cl-intel-greater-than-4GB-buffer-required"), std::string::npos);
}
HWTEST_F(ModuleTranslationUnitTest, givenSystemSharedAllocationAllowedWhenBuildingModuleThen4GbBuffersAreRequired) {
HWTEST_F(ModuleTranslationUnitTest, givenForceToStatelessRequiredWhenBuildingModuleThen4GbBuffersAreRequired) {
auto mockCompilerInterface = new MockCompilerInterface;
auto &rootDeviceEnvironment = neoDevice->executionEnvironment->rootDeviceEnvironments[neoDevice->getRootDeviceIndex()];
rootDeviceEnvironment->compilerInterface.reset(mockCompilerInterface);
{
neoDevice->getRootDeviceEnvironment().getMutableHardwareInfo()->capabilityTable.sharedSystemMemCapabilities = 1;
MockModuleTranslationUnit moduleTu(device);
auto ret = moduleTu.buildFromSpirV("", 0U, nullptr, "", nullptr);
EXPECT_TRUE(ret);
MockModuleTranslationUnit moduleTu(device);
auto ret = moduleTu.buildFromSpirV("", 0U, nullptr, "", nullptr);
EXPECT_TRUE(ret);
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (compilerHwInfoConfig.isForceToStatelessRequired()) {
EXPECT_NE(mockCompilerInterface->inputInternalOptions.find("cl-intel-greater-than-4GB-buffer-required"), std::string::npos);
}
{
neoDevice->getRootDeviceEnvironment().getMutableHardwareInfo()->capabilityTable.sharedSystemMemCapabilities = 0;
MockModuleTranslationUnit moduleTu(device);
auto ret = moduleTu.buildFromSpirV("", 0U, nullptr, "", nullptr);
EXPECT_TRUE(ret);
} else {
EXPECT_EQ(mockCompilerInterface->inputInternalOptions.find("cl-intel-greater-than-4GB-buffer-required"), std::string::npos);
}
}
@ -2012,6 +2005,155 @@ TEST_F(ModuleTest, GivenInjectInternalBuildOptionsWhenBuildingBuiltinModuleThenI
EXPECT_FALSE(CompilerOptions::contains(cip->buildInternalOptions, "-abc"));
};
TEST_F(ModuleTest, givenForceToStatelessRequiredWhenGenerateCompilerOptionsThenOptionsAreCorrect) {
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (!compilerHwInfoConfig.isForceToStatelessRequired()) {
GTEST_SKIP();
}
DebugManagerStateRestore restorer;
auto module = std::make_unique<ModuleImp>(device, nullptr, ModuleType::User);
ASSERT_NE(nullptr, module);
auto moduleTranslationUnit = module->getTranslationUnit();
ASSERT_NE(nullptr, moduleTranslationUnit);
std::string buildOptions;
std::string internalBuildOptions;
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(-1);
buildOptions = "";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required"));
}
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(-1);
buildOptions = "-ze-opt-smaller-than-4GB-buffers-only";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::Not(testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required")));
}
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(0);
buildOptions = "";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required"));
}
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(0);
buildOptions = "-ze-opt-smaller-than-4GB-buffers-only";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required"));
}
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(1);
buildOptions = "";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::Not(testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required")));
}
{
NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(1);
buildOptions = "-ze-opt-smaller-than-4GB-buffers-only";
auto internalOptions = moduleTranslationUnit->generateCompilerOptions(buildOptions.c_str(), internalBuildOptions.c_str());
EXPECT_THAT(internalOptions, testing::Not(testing::HasSubstr("-cl-intel-greater-than-4GB-buffer-required")));
}
}
TEST_F(ModuleTest, whenContainsStatefulAccessIsCalledThenResultIsCorrect) {
class MyModuleImpl : public ModuleImp {
public:
using ModuleImp::ModuleImp;
};
std::vector<std::tuple<bool, SurfaceStateHeapOffset, CrossThreadDataOffset>> testParams = {
{false, undefined<SurfaceStateHeapOffset>, undefined<CrossThreadDataOffset>},
{true, 0x40, undefined<CrossThreadDataOffset>},
{true, undefined<SurfaceStateHeapOffset>, 0x40},
{true, 0x40, 0x40},
};
for (auto &[expectedResult, surfaceStateHeapOffset, crossThreadDataOffset] : testParams) {
auto module = std::make_unique<MyModuleImpl>(device, nullptr, ModuleType::User);
ASSERT_NE(nullptr, module);
auto moduleTranslationUnit = module->getTranslationUnit();
ASSERT_NE(nullptr, moduleTranslationUnit);
auto kernelInfo = std::make_unique<KernelInfo>();
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.clear();
auto argDescriptor = ArgDescriptor(ArgDescriptor::ArgTPointer);
argDescriptor.as<ArgDescPointer>().bindful = surfaceStateHeapOffset;
argDescriptor.as<ArgDescPointer>().bindless = crossThreadDataOffset;
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.push_back(argDescriptor);
moduleTranslationUnit->programInfo.kernelInfos.clear();
moduleTranslationUnit->programInfo.kernelInfos.push_back(kernelInfo.release());
EXPECT_EQ(expectedResult, NEO::AddressingModeHelper::containsStatefulAccess(moduleTranslationUnit->programInfo.kernelInfos));
}
}
using ModuleInitializeTest = Test<DeviceFixture>;
TEST_F(ModuleInitializeTest, whenModuleInitializeIsCalledThenCorrectResultIsReturned) {
class MockModuleImp : public ModuleImp {
public:
using ModuleImp::ModuleImp;
using ModuleImp::translationUnit;
void setAddressingMode(bool isStateful) {
auto kernelInfo = std::make_unique<KernelInfo>();
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.clear();
auto argDescriptor = ArgDescriptor(ArgDescriptor::ArgTPointer);
if (isStateful) {
argDescriptor.as<ArgDescPointer>().bindful = 0x40;
argDescriptor.as<ArgDescPointer>().bindless = 0x40;
} else {
argDescriptor.as<ArgDescPointer>().bindful = undefined<SurfaceStateHeapOffset>;
argDescriptor.as<ArgDescPointer>().bindless = undefined<CrossThreadDataOffset>;
}
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.push_back(argDescriptor);
kernelInfo->heapInfo.KernelHeapSize = 0x1;
kernelInfo->heapInfo.pKernelHeap = reinterpret_cast<void *>(0xffff);
this->translationUnit->programInfo.kernelInfos.clear();
this->translationUnit->programInfo.kernelInfos.push_back(kernelInfo.release());
}
};
class MyMockModuleTU : public MockModuleTU {
public:
using MockModuleTU::MockModuleTU;
bool createFromNativeBinary(const char *input, size_t inputSize) override { return true; }
};
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (!compilerHwInfoConfig.isForceToStatelessRequired()) {
GTEST_SKIP();
}
DebugManagerStateRestore restorer;
std::string testFile;
retrieveBinaryKernelFilenameNoRevision(testFile, "test_kernel_", ".bin");
size_t size = 0;
auto src = loadDataFromFile(testFile.c_str(), size);
ASSERT_NE(0u, size);
ASSERT_NE(nullptr, src);
ze_module_desc_t moduleDesc = {};
moduleDesc.format = ZE_MODULE_FORMAT_NATIVE;
moduleDesc.pInputModule = reinterpret_cast<const uint8_t *>(src.get());
moduleDesc.inputSize = size;
std::array<std::tuple<bool, bool, ModuleType, int32_t>, 4> testParams = {{{true, false, ModuleType::Builtin, -1},
{true, true, ModuleType::Builtin, 1},
{true, true, ModuleType::Builtin, 0},
{false, true, ModuleType::User, 0}}};
for (auto &[expectedResult, isStateful, moduleType, debugKey] : testParams) {
MockModuleImp module(device, nullptr, moduleType);
module.translationUnit = std::make_unique<MyMockModuleTU>(device);
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(debugKey);
module.setAddressingMode(isStateful);
EXPECT_EQ(expectedResult, module.initialize(&moduleDesc, device->getNEODevice()));
}
}
using ModuleDebugDataTest = Test<DeviceFixture>;
TEST_F(ModuleDebugDataTest, GivenDebugDataWithRelocationsWhenCreatingRelocatedDebugDataThenRelocationsAreApplied) {
auto cip = new NEO::MockCompilerInterfaceCaptureBuildOptions();

10
opencl/source/program/build.cpp
View File

@ -10,6 +10,8 @@
#include "shared/source/device/device.h"
#include "shared/source/device_binary_format/device_binary_formats.h"
#include "shared/source/execution_environment/execution_environment.h"
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/helpers/compiler_hw_info_config.h"
#include "shared/source/helpers/compiler_options_parser.h"
#include "shared/source/program/kernel_info.h"
#include "shared/source/source_level_debugger/source_level_debugger.h"
@ -175,6 +177,14 @@ cl_int Program::build(
phaseReached[clDevice->getRootDeviceIndex()] = BuildPhase::BinaryProcessing;
}
auto containsStatefulAccess = AddressingModeHelper::containsStatefulAccess(buildInfos[clDevices[0]->getRootDeviceIndex()].kernelInfoArray);
auto forceToStatelessNeeded = AddressingModeHelper::forceToStatelessNeeded(options, CompilerOptions::smallerThan4gbBuffersOnly.str(), clDevices[0]->getHardwareInfo());
auto isUserKernel = !isBuiltIn;
if (containsStatefulAccess && forceToStatelessNeeded && isUserKernel) {
retVal = CL_BUILD_PROGRAM_FAILURE;
}
if (retVal != CL_SUCCESS) {
break;
}

8
opencl/source/program/program.cpp
View File

@ -15,6 +15,7 @@
#include "shared/source/device_binary_format/elf/elf_encoder.h"
#include "shared/source/device_binary_format/elf/ocl_elf.h"
#include "shared/source/device_binary_format/patchtokens_decoder.h"
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/helpers/api_specific_config.h"
#include "shared/source/helpers/compiler_options_parser.h"
#include "shared/source/helpers/debug_helpers.h"
@ -71,9 +72,11 @@ std::string Program::getInternalOptions() const {
if (force32BitAddressess && !isBuiltIn) {
CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::arch32bit);
}
auto &hwInfo = pClDevice->getHardwareInfo();
auto disableStatelessToStatefulOptimization = DebugManager.flags.DisableStatelessToStatefulOptimization.get();
auto forceToStatelessNeeded = AddressingModeHelper::forceToStatelessNeeded(options, CompilerOptions::smallerThan4gbBuffersOnly.str(), hwInfo);
if ((isBuiltIn && is32bit) || pClDevice->areSharedSystemAllocationsAllowed() ||
DebugManager.flags.DisableStatelessToStatefulOptimization.get()) {
if ((isBuiltIn && is32bit) || forceToStatelessNeeded || disableStatelessToStatefulOptimization) {
CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::greaterThan4gbBuffersRequired);
}
@ -90,7 +93,6 @@ std::string Program::getInternalOptions() const {
CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::hasBufferOffsetArg);
}
auto &hwInfo = pClDevice->getHardwareInfo();
const auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isForceEmuInt32DivRemSPWARequired(hwInfo)) {
CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::forceEmuInt32DivRemSP);

1
opencl/test/unit_test/aub_tests/xe_hpc_core/um_stateless_compression_aub_tests_xe_hpc_core.cpp
View File

@ -139,6 +139,7 @@ class UmStatelessCompressionWithStatefulAccess : public ProgramFixture,
void SetUp() override {
DebugManager.flags.EnableStatelessCompressionWithUnifiedMemory.set(1);
DebugManager.flags.EnableSharedSystemUsmSupport.set(0);
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(1);
compareCompressedMemory = GetParam();
ProgramFixture::SetUp();

15
opencl/test/unit_test/built_ins/built_in_tests.cpp
View File

@ -1459,17 +1459,21 @@ TEST_F(BuiltInTests, GivenTypeSourceWhenCreatingProgramFromCodeThenValidPointerI
EXPECT_NE(nullptr, program.get());
}
TEST_F(BuiltInTests, givenCreateProgramFromSourceWhenDeviceSupportSharedSystemAllocationThenInternalOptionsDisableStosoFlag) {
TEST_F(BuiltInTests, givenCreateProgramFromSourceWhenDeviceHasForceToStatelessRequiredOr32BitThenInternalOptionsHasGreaterThan4gbBuffersRequiredFlag) {
auto builtinsLib = std::unique_ptr<BuiltinsLib>(new BuiltinsLib());
pClDevice->deviceInfo.sharedSystemMemCapabilities = CL_UNIFIED_SHARED_MEMORY_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_ATOMIC_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ATOMIC_ACCESS_INTEL;
pDevice->getRootDeviceEnvironment().getMutableHardwareInfo()->capabilityTable.sharedSystemMemCapabilities = 1;
const BuiltinCode bc = builtinsLib->getBuiltinCode(EBuiltInOps::CopyBufferToBuffer, BuiltinCode::ECodeType::Source, *pDevice);
EXPECT_NE(0u, bc.resource.size());
auto program = std::unique_ptr<Program>(BuiltinDispatchInfoBuilder::createProgramFromCode(bc, toClDeviceVector(*pClDevice)));
EXPECT_NE(nullptr, program.get());
auto builtinInternalOptions = program->getInternalOptions();
EXPECT_THAT(builtinInternalOptions, testing::HasSubstr(std::string(CompilerOptions::greaterThan4gbBuffersRequired)));
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (compilerHwInfoConfig.isForceToStatelessRequired() || is32bit) {
EXPECT_THAT(builtinInternalOptions, testing::HasSubstr(std::string(CompilerOptions::greaterThan4gbBuffersRequired)));
} else {
EXPECT_THAT(builtinInternalOptions, testing::Not(testing::HasSubstr(std::string(CompilerOptions::greaterThan4gbBuffersRequired))));
}
}
TEST_F(BuiltInTests, GivenTypeIntermediateWhenCreatingProgramFromCodeThenNullPointerIsReturned) {
@ -1517,8 +1521,9 @@ TEST_F(BuiltInTests, GivenForce32bitWhenCreatingProgramThenCorrectKernelIsCreate
auto it = builtinInternalOptions.find(NEO::CompilerOptions::arch32bit.data());
EXPECT_EQ(std::string::npos, it);
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
it = builtinInternalOptions.find(NEO::CompilerOptions::greaterThan4gbBuffersRequired.data());
if (is32bit || pDevice->areSharedSystemAllocationsAllowed()) {
if (is32bit || compilerHwInfoConfig.isForceToStatelessRequired()) {
EXPECT_NE(std::string::npos, it);
} else {
EXPECT_EQ(std::string::npos, it);

163
opencl/test/unit_test/program/program_tests.cpp
View File

@ -14,7 +14,9 @@
#include "shared/source/device_binary_format/elf/ocl_elf.h"
#include "shared/source/device_binary_format/patchtokens_decoder.h"
#include "shared/source/gmm_helper/gmm_helper.h"
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/compiler_hw_info_config.h"
#include "shared/source/helpers/hash.h"
#include "shared/source/helpers/hw_helper.h"
#include "shared/source/helpers/ptr_math.h"
@ -1104,7 +1106,10 @@ TEST_F(ProgramFromSourceTest, GivenFlagsWhenCompilingProgramThenBuildOptionsHave
// Check build options that were applied
EXPECT_TRUE(CompilerOptions::contains(cip->buildOptions, CompilerOptions::fastRelaxedMath)) << cip->buildOptions;
EXPECT_FALSE(CompilerOptions::contains(cip->buildInternalOptions, CompilerOptions::gtpinRera)) << cip->buildInternalOptions;
if (!pDevice->areSharedSystemAllocationsAllowed()) {
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (!compilerHwInfoConfig.isForceToStatelessRequired()) {
EXPECT_FALSE(CompilerOptions::contains(cip->buildInternalOptions, CompilerOptions::greaterThan4gbBuffersRequired)) << cip->buildInternalOptions;
}
EXPECT_TRUE(CompilerOptions::contains(cip->buildInternalOptions, pPlatform->getClDevice(0)->peekCompilerExtensions())) << cip->buildInternalOptions;
@ -1650,6 +1655,140 @@ TEST_F(ProgramTests, WhenProgramIsCreatedThenCorrectOclVersionIsInOptions) {
}
}
TEST_F(ProgramTests, whenForceToStatelessNeededIsCalledThenCorrectResultIsReturned) {
DebugManagerStateRestore restorer;
class MyMockProgram : public Program {
public:
using Program::options;
using Program::Program;
};
MyMockProgram program(pContext, false, toClDeviceVector(*pClDevice));
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(pClDevice->getHardwareInfo().platform.eProductFamily);
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(-1);
program.options = "";
EXPECT_EQ(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()), compilerHwInfoConfig.isForceToStatelessRequired());
}
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(-1);
program.options = "-cl-opt-smaller-than-4GB-buffers-only";
EXPECT_FALSE(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()));
}
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(0);
program.options = "";
EXPECT_EQ(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()), compilerHwInfoConfig.isForceToStatelessRequired());
}
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(0);
program.options = "-cl-opt-smaller-than-4GB-buffers-only";
EXPECT_EQ(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()), compilerHwInfoConfig.isForceToStatelessRequired());
}
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(1);
program.options = "";
EXPECT_FALSE(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()));
}
{
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(1);
program.options = "-cl-opt-smaller-than-4GB-buffers-only";
EXPECT_FALSE(AddressingModeHelper::forceToStatelessNeeded(program.options, NEO::CompilerOptions::smallerThan4gbBuffersOnly.str(), pClDevice->getHardwareInfo()));
}
}
TEST_F(ProgramTests, whenContainsStatefulAccessIsCalledThenReturnCorrectResult) {
std::vector<std::tuple<bool, SurfaceStateHeapOffset, CrossThreadDataOffset>> testParams = {
{false, undefined<SurfaceStateHeapOffset>, undefined<CrossThreadDataOffset>},
{true, 0x40, undefined<CrossThreadDataOffset>},
{true, undefined<SurfaceStateHeapOffset>, 0x40},
{true, 0x40, 0x40},
};
for (auto &[expectedResult, surfaceStateHeapOffset, crossThreadDataOffset] : testParams) {
MockProgram program(pContext, false, toClDeviceVector(*pClDevice));
auto kernelInfo = std::make_unique<KernelInfo>();
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.clear();
auto argDescriptor = ArgDescriptor(ArgDescriptor::ArgTPointer);
argDescriptor.as<ArgDescPointer>().bindful = surfaceStateHeapOffset;
argDescriptor.as<ArgDescPointer>().bindless = crossThreadDataOffset;
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.push_back(argDescriptor);
program.addKernelInfo(kernelInfo.release(), 0);
EXPECT_EQ(expectedResult, AddressingModeHelper::containsStatefulAccess(program.buildInfos[0].kernelInfoArray));
}
}
TEST_F(ProgramTests, givenStatefulAndStatelessAccessesWhenProgramBuildIsCalledThenCorrectResultIsReturned) {
DebugManagerStateRestore restorer;
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(pClDevice->getHardwareInfo().platform.eProductFamily);
class MyMockProgram : public Program {
public:
using Program::buildInfos;
using Program::createdFrom;
using Program::irBinary;
using Program::irBinarySize;
using Program::isBuiltIn;
using Program::options;
using Program::Program;
using Program::sourceCode;
void setAddressingMode(bool isStateful) {
auto kernelInfo = std::make_unique<KernelInfo>();
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.clear();
auto argDescriptor = ArgDescriptor(ArgDescriptor::ArgTPointer);
if (isStateful) {
argDescriptor.as<ArgDescPointer>().bindful = 0x40;
argDescriptor.as<ArgDescPointer>().bindless = 0x40;
} else {
argDescriptor.as<ArgDescPointer>().bindful = undefined<SurfaceStateHeapOffset>;
argDescriptor.as<ArgDescPointer>().bindless = undefined<CrossThreadDataOffset>;
}
kernelInfo->kernelDescriptor.payloadMappings.explicitArgs.push_back(argDescriptor);
this->buildInfos[0].kernelInfoArray.clear();
this->buildInfos[0].kernelInfoArray.push_back(kernelInfo.release());
}
cl_int processGenBinary(const ClDevice &clDevice) override {
return CL_SUCCESS;
}
};
std::array<std::tuple<int, bool, int32_t>, 3> testParams = {{{CL_SUCCESS, false, -1},
{CL_SUCCESS, true, 1},
{CL_BUILD_PROGRAM_FAILURE, true, 0}}};
for (auto &[result, isStatefulAccess, debuyKey] : testParams) {
if (!compilerHwInfoConfig.isForceToStatelessRequired()) {
result = CL_SUCCESS;
}
MyMockProgram program(pContext, false, toClDeviceVector(*pClDevice));
program.isBuiltIn = false;
program.sourceCode = "test_kernel";
program.createdFrom = Program::CreatedFrom::SOURCE;
program.setAddressingMode(isStatefulAccess);
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(debuyKey);
EXPECT_EQ(result, program.build(toClDeviceVector(*pClDevice), nullptr, false));
}
{
MyMockProgram programWithBuiltIn(pContext, true, toClDeviceVector(*pClDevice));
programWithBuiltIn.isBuiltIn = true;
programWithBuiltIn.irBinary.reset(new char[16]);
programWithBuiltIn.irBinarySize = 16;
programWithBuiltIn.setAddressingMode(true);
DebugManager.flags.UseSmallerThan4gbBuffersOnly.set(0);
EXPECT_EQ(CL_SUCCESS, programWithBuiltIn.build(toClDeviceVector(*pClDevice), nullptr, false));
}
}
TEST_F(ProgramTests, GivenForcedClVersionWhenProgramIsCreatedThenCorrectOclOptionIsPresent) {
std::pair<unsigned int, std::string> testedValues[] = {
{0, "-ocl-version=120"},
@ -1694,14 +1833,6 @@ TEST_F(ProgramTests, WhenCreatingProgramThenBindlessIsEnabledOnlyIfDebugFlagIsEn
}
}
TEST_F(ProgramTests, givenDeviceThatSupportsSharedSystemMemoryAllocationWhenProgramIsCompiledThenItForcesStatelessCompilation) {
pClDevice->deviceInfo.sharedSystemMemCapabilities = CL_UNIFIED_SHARED_MEMORY_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_ATOMIC_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ATOMIC_ACCESS_INTEL;
pClDevice->getRootDeviceEnvironment().getMutableHardwareInfo()->capabilityTable.sharedSystemMemCapabilities = 1;
MockProgram program(pContext, false, toClDeviceVector(*pClDevice));
auto internalOptions = program.getInternalOptions();
EXPECT_TRUE(CompilerOptions::contains(internalOptions.c_str(), CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
}
TEST_F(ProgramTests, GivenForce32BitAddressessWhenProgramIsCreatedThenGreaterThan4gbBuffersRequiredIsCorrectlySet) {
DebugManagerStateRestore dbgRestorer;
cl_int retVal = CL_DEVICE_NOT_FOUND;
@ -1710,7 +1841,8 @@ TEST_F(ProgramTests, GivenForce32BitAddressessWhenProgramIsCreatedThenGreaterTha
const_cast<DeviceInfo *>(&pDevice->getDeviceInfo())->force32BitAddressess = true;
MockProgram program(pContext, false, toClDeviceVector(*pClDevice));
auto internalOptions = program.getInternalOptions();
if (pDevice->areSharedSystemAllocationsAllowed()) {
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (compilerHwInfoConfig.isForceToStatelessRequired()) {
EXPECT_TRUE(CompilerOptions::contains(internalOptions, CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
} else {
EXPECT_FALSE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
@ -1725,7 +1857,8 @@ TEST_F(ProgramTests, Given32bitSupportWhenProgramIsCreatedThenGreaterThan4gbBuff
DebugManager.flags.DisableStatelessToStatefulOptimization.set(false);
std::unique_ptr<MockProgram> program{Program::createBuiltInFromSource<MockProgram>("", pContext, pContext->getDevices(), nullptr)};
auto internalOptions = program->getInternalOptions();
if ((false == pDevice->areSharedSystemAllocationsAllowed()) && (false == is32bit)) {
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if ((false == compilerHwInfoConfig.isForceToStatelessRequired()) && (false == is32bit)) {
EXPECT_FALSE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
} else {
EXPECT_TRUE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
@ -1752,13 +1885,15 @@ TEST_F(ProgramTests, Force32BitAddressessWhenProgramIsCreatedThenGreaterThan4gbB
const_cast<DeviceInfo *>(&pDevice->getDeviceInfo())->force32BitAddressess = true;
std::unique_ptr<MockProgram> program{Program::createBuiltInFromSource<MockProgram>("", pContext, pContext->getDevices(), nullptr)};
auto internalOptions = program->getInternalOptions();
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(defaultHwInfo->platform.eProductFamily);
if (is32bit) {
EXPECT_TRUE(CompilerOptions::contains(internalOptions, CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
} else {
if (false == pDevice->areSharedSystemAllocationsAllowed()) {
EXPECT_FALSE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
} else {
if (compilerHwInfoConfig.isForceToStatelessRequired()) {
EXPECT_TRUE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
} else {
EXPECT_FALSE(CompilerOptions::contains(internalOptions, NEO::CompilerOptions::greaterThan4gbBuffersRequired)) << internalOptions;
}
}
}

1
opencl/test/unit_test/test_files/igdrcl.config
View File

@ -111,6 +111,7 @@ ForceCsrFlushing = 0
ForceCsrReprogramming = 0
OmitTimestampPacketDependencies = 0
DisableStatelessToStatefulOptimization = 0
UseSmallerThan4gbBuffersOnly = -1
DisableConcurrentBlockExecution = 0
UseNoRingFlushesKmdMode = 1
DisableZeroCopyForUseHostPtr = 0

1
shared/source/compiler_interface/compiler_options/compiler_options_base.h
View File

@ -16,6 +16,7 @@
namespace NEO {
namespace CompilerOptions {
static constexpr ConstStringRef greaterThan4gbBuffersRequired = "-cl-intel-greater-than-4GB-buffer-required";
static constexpr ConstStringRef smallerThan4gbBuffersOnly = "-cl-opt-smaller-than-4GB-buffers-only";
static constexpr ConstStringRef hasBufferOffsetArg = "-cl-intel-has-buffer-offset-arg";
static constexpr ConstStringRef kernelDebugEnable = "-cl-kernel-debug-enable";
static constexpr ConstStringRef arch32bit = "-m32";

1
shared/source/debug_settings/debug_variables_base.inl
View File

@ -238,6 +238,7 @@ DECLARE_DEBUG_VARIABLE(bool, DisableStatelessToStatefulOptimization, false, "Dis
DECLARE_DEBUG_VARIABLE(bool, DisableConcurrentBlockExecution, false, "disables concurrent block kernel execution")
DECLARE_DEBUG_VARIABLE(bool, UseNoRingFlushesKmdMode, true, "Windows only, passes flag to KMD that informs KMD to not emit any ring buffer flushes.")
DECLARE_DEBUG_VARIABLE(bool, DisableZeroCopyForUseHostPtr, false, "When active all buffer allocations created with CL_MEM_USE_HOST_PTR flag will not share memory with CPU.")
DECLARE_DEBUG_VARIABLE(int32_t, UseSmallerThan4gbBuffersOnly, -1, " -1: default, 0: disabled, 1: enabled. When enabled, the driver will not force stateless accesses on devices with default stateless addressing mode")
DECLARE_DEBUG_VARIABLE(int32_t, EnableHostPtrTracking, -1, "Enable host ptr tracking: -1 - default platform setting, 0 - disabled, 1 - enabled")
DECLARE_DEBUG_VARIABLE(int32_t, MaxHwThreadsPercent, 0, "If not zero then maximum number of used HW threads is capped to max * MaxHwThreadsPercent / 100")
DECLARE_DEBUG_VARIABLE(int32_t, MinHwThreadsUnoccupied, 0, "If not zero then maximum number of used HW threads is reduced by MinHwThreadsUnoccupied")

2
shared/source/helpers/CMakeLists.txt
View File

@ -7,6 +7,8 @@
set(NEO_CORE_HELPERS
${CMAKE_CURRENT_SOURCE_DIR}/CMakeLists.txt
${CMAKE_CURRENT_SOURCE_DIR}/abort.h
${CMAKE_CURRENT_SOURCE_DIR}/addressing_mode_helper.h
${CMAKE_CURRENT_SOURCE_DIR}/addressing_mode_helper.cpp
${CMAKE_CURRENT_SOURCE_DIR}/address_patch.h
${CMAKE_CURRENT_SOURCE_DIR}/affinity_mask.h
${CMAKE_CURRENT_SOURCE_DIR}/aligned_memory.h

45
shared/source/helpers/addressing_mode_helper.cpp Normal file
View File

@ -0,0 +1,45 @@
/*
* Copyright (C) 2021-2022 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/helpers/addressing_mode_helper.h"
#include "shared/source/compiler_interface/compiler_options/compiler_options_base.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/helpers/compiler_hw_info_config.h"
#include "shared/source/program/kernel_info.h"
namespace NEO::AddressingModeHelper {
bool forceToStatelessNeeded(const std::string &options, const std::string &smallerThan4GbBuffersOnlyOption, const HardwareInfo &hwInfo) {
auto preferStateful = false;
if (NEO::CompilerOptions::contains(options, smallerThan4GbBuffersOnlyOption)) {
preferStateful = true;
}
if (NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.get() != -1) {
preferStateful = static_cast<bool>(NEO::DebugManager.flags.UseSmallerThan4gbBuffersOnly.get());
}
const auto &compilerHwInfoConfig = *CompilerHwInfoConfig::get(hwInfo.platform.eProductFamily);
auto forceStateless = !preferStateful && compilerHwInfoConfig.isForceToStatelessRequired();
return forceStateless;
}
bool containsStatefulAccess(const std::vector<KernelInfo *> &kernelInfos) {
for (const auto &kernelInfo : kernelInfos) {
for (const auto &arg : kernelInfo->kernelDescriptor.payloadMappings.explicitArgs) {
auto isStatefulAccess = arg.is<NEO::ArgDescriptor::ArgTPointer>() &&
(NEO::isValidOffset(arg.as<NEO::ArgDescPointer>().bindless) ||
NEO::isValidOffset(arg.as<NEO::ArgDescPointer>().bindful));
if (isStatefulAccess) {
return true;
}
}
}
return false;
}
} // namespace NEO::AddressingModeHelper

22
shared/source/helpers/addressing_mode_helper.h Normal file
View File

@ -0,0 +1,22 @@
/*
* Copyright (C) 2021-2022 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include <string>
#include <vector>
namespace NEO {
struct KernelInfo;
struct HardwareInfo;
namespace AddressingModeHelper {
bool forceToStatelessNeeded(const std::string &options, const std::string &smallerThan4GbBuffersOnlyOption, const HardwareInfo &hwInfo);
bool containsStatefulAccess(const std::vector<KernelInfo *> &kernelInfos);
} // namespace AddressingModeHelper
} // namespace NEO