intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-08 22:12:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Store SLM sizes per root device in Kernel

Browse Source 
reduce usages of getDefaultKernelInfo

Related-To: NEO-5001
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski
2020-12-18 14:47:42 +00:00
committed by Compute-Runtime-Automation
parent 09bdd2ad09
commit be7ae13911
15 changed files with 48 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
opencl/source/api/api.cpp
View File

@@ -1845,7 +1845,7 @@ cl_int CL_API_CALL clSetKernelArg(cl_kernel kernel,
retVal = CL_INVALID_KERNEL;
break;
}
if (pKernel->getDefaultKernelInfo().kernelArgInfo.size() <= argIndex) {
if (pKernel->getKernelArguments().size() <= argIndex) {
retVal = CL_INVALID_ARG_INDEX;
break;
}

2
opencl/source/helpers/dispatch_info.cpp
View File

@@ -11,7 +11,7 @@
namespace NEO {
bool DispatchInfo::usesSlm() const {
return (kernel == nullptr) ? false : kernel->slmTotalSize > 0;
return (kernel == nullptr) ? false : kernel->getSlmTotalSize(pClDevice->getRootDeviceIndex()) > 0;
}
bool DispatchInfo::usesStatelessPrintfSurface() const {

6
opencl/source/helpers/hardware_commands_helper_base.inl
View File

@@ -180,8 +180,10 @@ size_t HardwareCommandsHelper<GfxFamily>::sendInterfaceDescriptorData(
interfaceDescriptor.setDenormMode(INTERFACE_DESCRIPTOR_DATA::DENORM_MODE_SETBYKERNEL);
auto slmTotalSize = kernel.getSlmTotalSize(rootDeviceIndex);
setGrfInfo(&interfaceDescriptor, kernel, sizeCrossThreadData, sizePerThreadData, rootDeviceIndex);
EncodeDispatchKernel<GfxFamily>::appendAdditionalIDDFields(&interfaceDescriptor, hardwareInfo, threadsPerThreadGroup, kernel.slmTotalSize, SlmPolicy::SlmPolicyNone);
EncodeDispatchKernel<GfxFamily>::appendAdditionalIDDFields(&interfaceDescriptor, hardwareInfo, threadsPerThreadGroup, slmTotalSize, SlmPolicy::SlmPolicyNone);
interfaceDescriptor.setBindingTablePointer(static_cast<uint32_t>(bindingTablePointer));
@@ -190,7 +192,7 @@ size_t HardwareCommandsHelper<GfxFamily>::sendInterfaceDescriptorData(
EncodeDispatchKernel<GfxFamily>::adjustBindingTablePrefetch(interfaceDescriptor, numSamplers, bindingTablePrefetchSize);
auto programmableIDSLMSize =
static_cast<SHARED_LOCAL_MEMORY_SIZE>(HwHelperHw<GfxFamily>::get().computeSlmValues(hardwareInfo, kernel.slmTotalSize));
static_cast<SHARED_LOCAL_MEMORY_SIZE>(HwHelperHw<GfxFamily>::get().computeSlmValues(hardwareInfo, slmTotalSize));
interfaceDescriptor.setSharedLocalMemorySize(programmableIDSLMSize);
EncodeDispatchKernel<GfxFamily>::programBarrierEnable(interfaceDescriptor,

4
opencl/source/helpers/task_information.cpp
View File

@@ -166,6 +166,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
printfHandler.get()->makeResident(commandStreamReceiver);
}
makeTimestampPacketsResident(commandStreamReceiver);
auto rootDeviceIndex = commandQueue.getDevice().getRootDeviceIndex();
if (executionModelKernel) {
uint32_t taskCount = commandStreamReceiver.peekTaskCount() + 1;
@@ -195,7 +196,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
scheduler.makeResident(commandStreamReceiver);
// Update SLM usage
slmUsed |= scheduler.slmTotalSize > 0;
slmUsed |= scheduler.getSlmTotalSize(rootDeviceIndex) > 0;
this->kernel->getProgram()->getBlockKernelManager()->makeInternalAllocationsResident(commandStreamReceiver);
}
@@ -210,7 +211,6 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
commandQueue.getGpgpuCommandStreamReceiver(), bcsCsr);
}
auto rootDeviceIndex = commandQueue.getDevice().getRootDeviceIndex();
const auto &kernelDescriptor = kernel->getKernelInfo(rootDeviceIndex).kernelDescriptor;
auto memoryCompressionState = commandStreamReceiver.getMemoryCompressionState(kernel->isAuxTranslationRequired());

27
opencl/source/kernel/kernel.cpp
View File

@@ -65,8 +65,7 @@ class Surface;
uint32_t Kernel::dummyPatchLocation = 0xbaddf00d;
Kernel::Kernel(Program *programArg, const KernelInfoContainer &kernelInfosArg, bool schedulerKernel)
: slmTotalSize(kernelInfosArg[programArg->getDevices()[0]->getRootDeviceIndex()]->workloadInfo.slmStaticSize),
isParentKernel(kernelInfosArg[programArg->getDevices()[0]->getRootDeviceIndex()]->kernelDescriptor.kernelAttributes.flags.usesDeviceSideEnqueue),
: isParentKernel(kernelInfosArg[programArg->getDevices()[0]->getRootDeviceIndex()]->kernelDescriptor.kernelAttributes.flags.usesDeviceSideEnqueue),
isSchedulerKernel(schedulerKernel),
executionEnvironment(programArg->getExecutionEnvironment()),
program(programArg),
@@ -78,7 +77,9 @@ Kernel::Kernel(Program *programArg, const KernelInfoContainer &kernelInfosArg, b
program->retainForKernel();
imageTransformer.reset(new ImageTransformer);
for (const auto &pClDevice : deviceVector) {
kernelDeviceInfos[pClDevice->getRootDeviceIndex()].maxKernelWorkGroupSize = static_cast<uint32_t>(pClDevice->getSharedDeviceInfo().maxWorkGroupSize);
auto rootDeviceIndex = pClDevice->getRootDeviceIndex();
kernelDeviceInfos[rootDeviceIndex].maxKernelWorkGroupSize = static_cast<uint32_t>(pClDevice->getSharedDeviceInfo().maxWorkGroupSize);
kernelDeviceInfos[rootDeviceIndex].slmTotalSize = kernelInfosArg[rootDeviceIndex]->workloadInfo.slmStaticSize;
}
}
@@ -100,7 +101,7 @@ Kernel::~Kernel() {
}
for (uint32_t i = 0; i < patchedArgumentsNum; i++) {
if (getDefaultKernelInfo().kernelArgInfo.at(i).isSampler) {
if (SAMPLER_OBJ == getKernelArguments()[i].type) {
auto sampler = castToObject<Sampler>(kernelArguments.at(i).object);
if (sampler) {
sampler->decRefInternal();
@@ -372,6 +373,8 @@ cl_int Kernel::initialize() {
if (program->isKernelDebugEnabled() && kernelInfo.patchInfo.pAllocateSystemThreadSurface) {
debugEnabled = true;
}
auto numArgs = kernelInfo.kernelArgInfo.size();
kernelDeviceInfo.slmSizes.resize(numArgs);
isDeviceInitialized.set(rootDeviceIndex);
}
@@ -384,13 +387,11 @@ cl_int Kernel::initialize() {
auto &defaultKernelInfo = getDefaultKernelInfo();
auto numArgs = defaultKernelInfo.kernelArgInfo.size();
kernelArguments.resize(numArgs);
slmSizes.resize(numArgs);
kernelArgHandlers.resize(numArgs);
kernelArgRequiresCacheFlush.resize(numArgs);
for (uint32_t i = 0; i < numArgs; ++i) {
storeKernelArg(i, NONE_OBJ, nullptr, nullptr, 0);
slmSizes[i] = 0;
// set the argument handler
auto &argInfo = defaultKernelInfo.kernelArgInfo[i];
@@ -483,7 +484,6 @@ cl_int Kernel::getInfo(cl_kernel_info paramName, size_t paramValueSize,
const _cl_context *ctxt;
cl_uint refCount = 0;
uint64_t nonCannonizedGpuAddress = 0llu;
auto defaultRootDeviceIndex = getDevices()[0]->getRootDeviceIndex();
auto &defaultKernelInfo = getKernelInfo(defaultRootDeviceIndex);
switch (paramName) {
@@ -1112,7 +1112,7 @@ uint32_t Kernel::getMaxWorkGroupCount(const cl_uint workDim, const size_t *local
availableThreadCount,
dssCount,
dssCount * KB * hardwareInfo.capabilityTable.slmSize,
hwHelper.alignSlmSize(slmTotalSize),
hwHelper.alignSlmSize(kernelDeviceInfos[rootDeviceIndex].slmTotalSize),
static_cast<uint32_t>(hwHelper.getMaxBarrierRegisterPerSlice()),
hwHelper.getBarriersCountFromHasBarriers(barrierCount),
workDim,
@@ -1280,7 +1280,7 @@ cl_int Kernel::setArgLocal(uint32_t argIndex,
storeKernelArg(argIndex, SLM_OBJ, nullptr, argVal, argSize);
slmSizes[argIndex] = argSize;
kernelDeviceInfos[rootDeviceIndex].slmSizes[argIndex] = argSize;
// Extract our current slmOffset
auto slmOffset = *ptrOffset(crossThreadData,
@@ -1291,7 +1291,7 @@ cl_int Kernel::setArgLocal(uint32_t argIndex,
// Update all slm offsets after this argIndex
++argIndex;
while (argIndex < slmSizes.size()) {
while (argIndex < kernelDeviceInfos[rootDeviceIndex].slmSizes.size()) {
const auto &kernelArgInfo = defaultKernelInfo.kernelArgInfo[argIndex];
auto slmAlignment = kernelArgInfo.slmAlignment;
@@ -1306,11 +1306,11 @@ cl_int Kernel::setArgLocal(uint32_t argIndex,
*patchLocation = slmOffset;
}
slmOffset += static_cast<uint32_t>(slmSizes[argIndex]);
slmOffset += static_cast<uint32_t>(kernelDeviceInfos[rootDeviceIndex].slmSizes[argIndex]);
++argIndex;
}
slmTotalSize = defaultKernelInfo.workloadInfo.slmStaticSize + alignUp(slmOffset, KB);
kernelDeviceInfos[rootDeviceIndex].slmTotalSize = defaultKernelInfo.workloadInfo.slmStaticSize + alignUp(slmOffset, KB);
return CL_SUCCESS;
}
@@ -2679,4 +2679,7 @@ void Kernel::setWorkDim(uint32_t rootDeviceIndex, uint32_t workDim) {
uint32_t Kernel::getMaxKernelWorkGroupSize(uint32_t rootDeviceIndex) const {
return kernelDeviceInfos[rootDeviceIndex].maxKernelWorkGroupSize;
}
uint32_t Kernel::getSlmTotalSize(uint32_t rootDeviceIndex) const {
return kernelDeviceInfos[rootDeviceIndex].slmTotalSize;
}
} // namespace NEO

9
opencl/source/kernel/kernel.h
View File

@@ -182,7 +182,7 @@ class Kernel : public BaseObject<_cl_kernel> {
}
size_t getKernelArgsNumber() const {
return getDefaultKernelInfo().kernelArgInfo.size();
return kernelArguments.size();
}
bool requiresSshForBuffers(uint32_t rootDeviceIndex) const {
@@ -308,11 +308,8 @@ class Kernel : public BaseObject<_cl_kernel> {
static uint32_t dummyPatchLocation;
std::vector<size_t> slmSizes;
uint32_t allBufferArgsStateful = CL_TRUE;
uint32_t slmTotalSize;
bool isBuiltIn = false;
const bool isParentKernel;
const bool isSchedulerKernel;
@@ -406,6 +403,7 @@ class Kernel : public BaseObject<_cl_kernel> {
void setNumWorkGroupsValues(uint32_t rootDeviceIndex, uint32_t numWorkGroupsX, uint32_t numWorkGroupsY, uint32_t numWorkGroupsZ);
void setWorkDim(uint32_t rootDeviceIndex, uint32_t workDim);
uint32_t getMaxKernelWorkGroupSize(uint32_t rootDeviceIndex) const;
uint32_t getSlmTotalSize(uint32_t rootDeviceIndex) const;
protected:
struct ObjectCounts {
@@ -574,6 +572,9 @@ class Kernel : public BaseObject<_cl_kernel> {
size_t numberOfBindingTableStates = 0u;
size_t localBindingTableOffset = 0u;
std::vector<size_t> slmSizes;
uint32_t slmTotalSize = 0u;
std::unique_ptr<char[]> pSshLocal;
uint32_t sshLocalSize = 0u;
char *crossThreadData = nullptr;

2
opencl/source/program/kernel_info.cpp
View File

@@ -138,7 +138,7 @@ WorkSizeInfo::WorkSizeInfo(const DispatchInfo &dispatchInfo) {
this->maxWorkGroupSize = dispatchInfo.getKernel()->getMaxKernelWorkGroupSize(rootDeviceIndex);
this->hasBarriers = kernelInfo.kernelDescriptor.kernelAttributes.usesBarriers();
this->simdSize = static_cast<uint32_t>(kernelInfo.getMaxSimdSize());
this->slmTotalSize = static_cast<uint32_t>(dispatchInfo.getKernel()->slmTotalSize);
this->slmTotalSize = static_cast<uint32_t>(dispatchInfo.getKernel()->getSlmTotalSize(rootDeviceIndex));
this->coreFamily = device.getHardwareInfo().platform.eRenderCoreFamily;
this->numThreadsPerSubSlice = static_cast<uint32_t>(device.getSharedDeviceInfo().maxNumEUsPerSubSlice) *
device.getSharedDeviceInfo().numThreadsPerEU;

2
opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_2_tests.cpp
View File

@@ -380,7 +380,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenKernelWithSlmWhenPreviousSLML
commandStreamReceiver->lastSentL3Config = L3Config;
commandStreamReceiver->lastSentThreadArbitrationPolicy = kernel.mockKernel->getThreadArbitrationPolicy();
((MockKernel *)kernel)->setTotalSLMSize(1024);
((MockKernel *)kernel)->setTotalSLMSize(rootDeviceIndex, 1024);
cmdList.clear();
commandQueue.enqueueKernel(kernel, 1, nullptr, &GWS, nullptr, 0, nullptr, nullptr);

4
opencl/test/unit_test/command_stream/command_stream_receiver_hw_tests.inl
View File

@@ -43,7 +43,7 @@ void CommandStreamReceiverHwTest<GfxFamily>::givenKernelWithSlmWhenPreviousNOSLM
commandStreamReceiver->isPreambleSent = true;
commandStreamReceiver->lastSentL3Config = 0;
static_cast<MockKernel *>(kernel)->setTotalSLMSize(1024);
static_cast<MockKernel *>(kernel)->setTotalSLMSize(rootDeviceIndex, 1024);
cmdList.clear();
commandQueue.enqueueKernel(kernel, 1, nullptr, &GWS, nullptr, 0, nullptr, nullptr);
@@ -89,7 +89,7 @@ void CommandStreamReceiverHwTest<GfxFamily>::givenBlockedKernelWithSlmWhenPrevio
commandStreamReceiver->isPreambleSent = true;
commandStreamReceiver->lastSentL3Config = 0;
static_cast<MockKernel *>(kernel)->setTotalSLMSize(1024);
static_cast<MockKernel *>(kernel)->setTotalSLMSize(rootDeviceIndex, 1024);
commandQueue.enqueueKernel(kernel, 1, nullptr, &GWS, nullptr, 1, &blockingEvent, nullptr);

2
opencl/test/unit_test/helpers/dispatch_info_builder_tests.cpp
View File

@@ -68,7 +68,7 @@ class DispatchInfoBuilderFixture : public ContextFixture, public ClDeviceFixture
pKernel->setCrossThreadData(pCrossThreadData, sizeof(pCrossThreadData));
pKernel->setKernelArgHandler(0, &Kernel::setArgBuffer);
pKernel->slmTotalSize = 128;
pKernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize = 128;
pKernel->isBuiltIn = true;
}

4
opencl/test/unit_test/helpers/dispatch_info_tests.cpp
View File

@@ -40,7 +40,7 @@ class DispatchInfoFixture : public ContextFixture, public ClDeviceFixture {
pProgram = new MockProgram(pContext, false, toClDeviceVector(*pClDevice));
pKernel = new MockKernel(pProgram, MockKernel::toKernelInfoContainer(*pKernelInfo, rootDeviceIndex));
pKernel->slmTotalSize = 128;
pKernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize = 128;
}
void TearDown() override {
delete pKernel;
@@ -56,7 +56,7 @@ class DispatchInfoFixture : public ContextFixture, public ClDeviceFixture {
SPatchMediaVFEState *pMediaVFEstate = nullptr;
SPatchAllocateStatelessPrintfSurface *pPrintfSurface = nullptr;
MockProgram *pProgram = nullptr;
Kernel *pKernel = nullptr;
MockKernel *pKernel = nullptr;
};
typedef Test<DispatchInfoFixture> DispatchInfoTest;

13
opencl/test/unit_test/kernel/clone_kernel_tests.cpp
View File

@@ -6,6 +6,7 @@
*/
#include "shared/source/memory_manager/unified_memory_manager.h"
#include "shared/test/unit_test/utilities/base_object_utils.h"
#include "opencl/source/accelerators/intel_accelerator.h"
#include "opencl/source/accelerators/intel_motion_estimation.h"
@@ -162,7 +163,7 @@ TEST_F(CloneKernelTest, GivenArgLocalWhenCloningKernelThenKernelInfoIsCorrect) {
EXPECT_EQ(pSourceKernel->getPatchedArgumentsNum(), pClonedKernel->getPatchedArgumentsNum());
EXPECT_EQ(pSourceKernel->getKernelArgInfo(0).isPatched, pClonedKernel->getKernelArgInfo(0).isPatched);
EXPECT_EQ(alignUp(slmSize, 1024), pClonedKernel->slmTotalSize);
EXPECT_EQ(alignUp(slmSize, 1024), pClonedKernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize);
}
TEST_F(CloneKernelTest, GivenArgBufferWhenCloningKernelThenKernelInfoIsCorrect) {
@@ -336,10 +337,10 @@ TEST_F(CloneKernelTest, GivenArgAcceleratorWhenCloningKernelThenKernelInfoIsCorr
}
TEST_F(CloneKernelTest, GivenArgSamplerWhenCloningKernelThenKernelInfoIsCorrect) {
std::unique_ptr<Sampler> sampler(new MockSampler(pContext,
true,
(cl_addressing_mode)CL_ADDRESS_MIRRORED_REPEAT,
(cl_filter_mode)CL_FILTER_NEAREST));
auto sampler = clUniquePtr<Sampler>(new MockSampler(pContext,
true,
(cl_addressing_mode)CL_ADDRESS_MIRRORED_REPEAT,
(cl_filter_mode)CL_FILTER_NEAREST));
uint32_t objectId = SAMPLER_OBJECT_ID_SHIFT + pKernelInfo->kernelArgInfo[0].offsetHeap;
@@ -381,6 +382,8 @@ TEST_F(CloneKernelTest, GivenArgSamplerWhenCloningKernelThenKernelInfoIsCorrect)
auto pNormalizedCoords = ptrOffset(crossThreadData, argInfo.offsetSamplerNormalizedCoords);
EXPECT_EQ(GetNormCoordsEnum(sampler->normalizedCoordinates), *pNormalizedCoords);
EXPECT_EQ(3, sampler->getRefInternalCount());
}
HWCMDTEST_F(IGFX_GEN8_CORE, CloneKernelTest, GivenArgDeviceQueueWhenCloningKernelThenKernelInfoIsCorrect) {

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

@@ -83,7 +83,7 @@ TEST_F(KernelSlmArgTest, WhenSettingSizeThenAlignmentOfHigherSlmArgsIsUpdated) {
slmOffset = ptrOffset(crossThreadData, 0x30);
EXPECT_EQ(0x400u, *slmOffset);
EXPECT_EQ(5 * KB, pKernel->slmTotalSize);
EXPECT_EQ(5 * KB, pKernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize);
}
TEST_F(KernelSlmArgTest, GivenReverseOrderWhenSettingSizeThenAlignmentOfHigherSlmArgsIsUpdated) {
@@ -100,5 +100,5 @@ TEST_F(KernelSlmArgTest, GivenReverseOrderWhenSettingSizeThenAlignmentOfHigherSl
slmOffset = ptrOffset(crossThreadData, 0x30);
EXPECT_EQ(0x400u, *slmOffset);
EXPECT_EQ(5 * KB, pKernel->slmTotalSize);
EXPECT_EQ(5 * KB, pKernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize);
}

2
opencl/test/unit_test/kernel/kernel_tests.cpp
View File

@@ -2288,7 +2288,7 @@ TEST_F(KernelCrossThreadTests, GivenSlmStatisSizeWhenCreatingKernelThenSlmTotalS
MockKernel *kernel = new MockKernel(program.get(), MockKernel::toKernelInfoContainer(*pKernelInfo, rootDeviceIndex));
EXPECT_EQ(1024u, kernel->slmTotalSize);
EXPECT_EQ(1024u, kernel->kernelDeviceInfos[rootDeviceIndex].slmTotalSize);
delete kernel;
}

4
opencl/test/unit_test/mocks/mock_kernel.h
View File

@@ -211,8 +211,8 @@ class MockKernel : public Kernel {
}
}
void setTotalSLMSize(uint32_t size) {
slmTotalSize = size;
void setTotalSLMSize(uint32_t rootDeviceIndex, uint32_t size) {
kernelDeviceInfos[rootDeviceIndex].slmTotalSize = size;
}
void setKernelArguments(std::vector<SimpleKernelArgInfo> kernelArguments) {