intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-11 16:45:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Limit cooperative dispatch support to platforms supporting cooperative engine

Browse Source 
By default, cooperative dispatch is not supported
for platforms not supporting cooperative engine.

Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>
This commit is contained in:
Filip Hazubski
2022-11-20 22:36:09 +00:00
committed by Compute-Runtime-Automation
parent 6c7ca96f28
commit c124bfbc6c
11 changed files with 193 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -6047,7 +6047,9 @@ cl_int CL_API_CALL clEnqueueNDCountKernelINTEL(cl_command_queue commandQueue,
auto &gfxCoreHelper = device.getGfxCoreHelper();
auto engineGroupType = gfxCoreHelper.getEngineGroupType(pCommandQueue->getGpgpuEngine().getEngineType(),
pCommandQueue->getGpgpuEngine().getEngineUsage(), hardwareInfo);
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo) &&
(DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.get() == false) &&
(DebugManager.flags.OverrideMaxWorkGroupCount.get() == -1)) {
retVal = CL_INVALID_COMMAND_QUEUE;
return retVal;
}

90
opencl/test/unit_test/command_queue/enqueue_kernel_1_tests.cpp
View File

@@ -10,6 +10,7 @@
#include "shared/source/helpers/preamble.h"
#include "shared/test/common/helpers/debug_manager_state_restore.h"
#include "shared/test/common/helpers/kernel_binary_helper.h"
#include "shared/test/common/helpers/raii_hw_helper.h"
#include "shared/test/common/helpers/unit_test_helper.h"
#include "shared/test/common/mocks/mock_csr.h"
#include "shared/test/common/mocks/mock_submissions_aggregator.h"
@@ -175,6 +176,9 @@ TEST(EnqueueMultiDeviceKernelTest, givenMultiDeviceKernelWhenSetArgDeviceUSMThen
}
TEST_F(EnqueueKernelTest, givenKernelWhenNotAllArgsAreSetButSetKernelArgIsCalledTwiceThenClEnqueueNDRangeKernelReturnsError) {
DebugManagerStateRestore stateRestore;
DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
const size_t n = 512;
size_t globalWorkSize[3] = {n, 1, 1};
size_t localWorkSize[3] = {256, 1, 1};
@@ -310,13 +314,6 @@ TEST_F(EnqueueKernelTest, GivenNullKernelWhenEnqueuingNDCountKernelINTELThenInva
using clEnqueueNDCountKernelTests = api_tests;
TEST_F(clEnqueueNDCountKernelTests, GivenQueueIncapableWhenEnqueuingNDCountKernelINTELThenInvalidOperationIsReturned) {
auto &gfxCoreHelper = GfxCoreHelper::get(::defaultHwInfo->platform.eRenderCoreFamily);
auto engineGroupType = gfxCoreHelper.getEngineGroupType(pCommandQueue->getGpgpuEngine().getEngineType(),
pCommandQueue->getGpgpuEngine().getEngineUsage(), *::defaultHwInfo);
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, *::defaultHwInfo)) {
GTEST_SKIP();
}
cl_uint workDim = 1;
size_t globalWorkOffset[3] = {0, 0, 0};
size_t workgroupCount[3] = {1, 1, 1};
@@ -338,6 +335,9 @@ TEST_F(clEnqueueNDCountKernelTests, GivenQueueIncapableWhenEnqueuingNDCountKerne
}
TEST_F(EnqueueKernelTest, givenKernelWhenAllArgsAreSetThenClEnqueueNDCountKernelINTELReturnsSuccess) {
DebugManagerStateRestore stateRestore;
DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
const size_t n = 512;
size_t workgroupCount[3] = {2, 1, 1};
size_t localWorkSize[3] = {64, 1, 1};
@@ -345,9 +345,11 @@ TEST_F(EnqueueKernelTest, givenKernelWhenAllArgsAreSetThenClEnqueueNDCountKernel
CommandQueue *pCmdQ2 = createCommandQueue(pClDevice);
GfxCoreHelper &gfxCoreHelper = GfxCoreHelper::get(hardwareInfo.platform.eRenderCoreFamily);
auto &productHelper = *ProductHelper::get(hardwareInfo.platform.eProductFamily);
auto engineGroupType = gfxCoreHelper.getEngineGroupType(pCmdQ2->getGpgpuEngine().getEngineType(),
pCmdQ2->getGpgpuEngine().getEngineUsage(), hardwareInfo);
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
if (productHelper.isCooperativeEngineSupported(hardwareInfo) &&
!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
pCmdQ2->getGpgpuEngine().osContext = pCmdQ2->getDevice().getEngine(aub_stream::ENGINE_CCS, EngineUsage::LowPriority).osContext;
}
@@ -386,6 +388,9 @@ TEST_F(EnqueueKernelTest, givenKernelWhenAllArgsAreSetThenClEnqueueNDCountKernel
}
TEST_F(EnqueueKernelTest, givenKernelWhenNotAllArgsAreSetButSetKernelArgIsCalledTwiceThenClEnqueueNDCountKernelINTELReturnsError) {
DebugManagerStateRestore stateRestore;
DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
const size_t n = 512;
size_t workgroupCount[3] = {2, 1, 1};
size_t localWorkSize[3] = {256, 1, 1};
@@ -393,9 +398,11 @@ TEST_F(EnqueueKernelTest, givenKernelWhenNotAllArgsAreSetButSetKernelArgIsCalled
CommandQueue *pCmdQ2 = createCommandQueue(pClDevice);
GfxCoreHelper &gfxCoreHelper = GfxCoreHelper::get(hardwareInfo.platform.eRenderCoreFamily);
auto &productHelper = *ProductHelper::get(hardwareInfo.platform.eProductFamily);
auto engineGroupType = gfxCoreHelper.getEngineGroupType(pCmdQ2->getGpgpuEngine().getEngineType(),
pCmdQ2->getGpgpuEngine().getEngineUsage(), hardwareInfo);
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
if (productHelper.isCooperativeEngineSupported(hardwareInfo) &&
!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
pCmdQ2->getGpgpuEngine().osContext = pCmdQ2->getDevice().getEngine(aub_stream::ENGINE_CCS, EngineUsage::LowPriority).osContext;
}
@@ -434,6 +441,9 @@ TEST_F(EnqueueKernelTest, givenKernelWhenNotAllArgsAreSetButSetKernelArgIsCalled
}
TEST_F(EnqueueKernelTest, givenKernelWhenSetKernelArgIsCalledForEachArgButAtLeastFailsThenClEnqueueNDCountKernelINTELReturnsError) {
DebugManagerStateRestore stateRestore;
DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
const size_t n = 512;
size_t workgroupCount[3] = {2, 1, 1};
size_t localWorkSize[3] = {256, 1, 1};
@@ -441,9 +451,11 @@ TEST_F(EnqueueKernelTest, givenKernelWhenSetKernelArgIsCalledForEachArgButAtLeas
CommandQueue *pCmdQ2 = createCommandQueue(pClDevice);
GfxCoreHelper &gfxCoreHelper = GfxCoreHelper::get(hardwareInfo.platform.eRenderCoreFamily);
auto &productHelper = *ProductHelper::get(hardwareInfo.platform.eProductFamily);
auto engineGroupType = gfxCoreHelper.getEngineGroupType(pCmdQ2->getGpgpuEngine().getEngineType(),
pCmdQ2->getGpgpuEngine().getEngineUsage(), hardwareInfo);
if (!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
if (productHelper.isCooperativeEngineSupported(hardwareInfo) &&
!gfxCoreHelper.isCooperativeDispatchSupported(engineGroupType, hardwareInfo)) {
pCmdQ2->getGpgpuEngine().osContext = pCmdQ2->getDevice().getEngine(aub_stream::ENGINE_CCS, EngineUsage::LowPriority).osContext;
}
@@ -498,6 +510,64 @@ TEST_F(EnqueueKernelTest, GivenInvalidEventListCountWhenEnqueuingNDCountKernelIN
EXPECT_EQ(CL_INVALID_EVENT_WAIT_LIST, retVal);
}
template <typename FamilyType>
struct MockGfxCoreHelperCooperativeDispatchSupportedOverride : GfxCoreHelperHw<FamilyType> {
bool isCooperativeDispatchSupported(const EngineGroupType engineGroupType, const HardwareInfo &hwInfo) const override {
return isCooperativeDispatchSupportedValue;
}
bool isCooperativeDispatchSupportedValue = false;
};
HWTEST_F(EnqueueKernelTest, GivenCommandQueueNotSupportingCooperativeDispatchWhenCallingClEnqueueNDCountKernelINTELWithConcurrenteKernelThenCorrectReturnValueIsReturned) {
MockGfxCoreHelperCooperativeDispatchSupportedOverride<FamilyType> gfxCoreHelper{};
VariableBackup<GfxCoreHelper *> gfxCoreHelperFactoryBackup{&NEO::gfxCoreHelperFactory[static_cast<size_t>(hardwareInfo.platform.eRenderCoreFamily)]};
gfxCoreHelperFactoryBackup = &gfxCoreHelper;
auto pMockCommandQueueHw = std::make_unique<MockCommandQueue>(context, pClDevice, nullptr, false);
MockKernelWithInternals mockKernel(*pClDevice, context);
mockKernel.kernelInfo.kernelDescriptor.kernelAttributes.flags.usesSyncBuffer = true;
mockKernel.mockMultiDeviceKernel->setKernelExecutionType(CL_KERNEL_EXEC_INFO_CONCURRENT_TYPE_INTEL);
auto getClEnqueueNDCountKernelINTELResult = [&pMockCommandQueueHw, &mockKernel]() -> cl_int {
size_t workgroupCount[3] = {1, 1, 1};
size_t localWorkSize[3] = {1, 1, 1};
return clEnqueueNDCountKernelINTEL(
pMockCommandQueueHw.get(),
mockKernel.mockMultiDeviceKernel,
1,
nullptr,
workgroupCount,
localWorkSize,
0,
nullptr,
nullptr);
};
for (auto useDebugFlagForceTheoreticalMaxWgCount : ::testing::Bool()) {
DebugManagerStateRestore stateRestore;
if (useDebugFlagForceTheoreticalMaxWgCount) {
DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
}
for (auto useDebugFlagOverrideMaxWgCount : ::testing::Bool()) {
if (useDebugFlagOverrideMaxWgCount) {
DebugManager.flags.OverrideMaxWorkGroupCount.set(1);
}
for (auto isCooperativeDispatchSupported : ::testing::Bool()) {
gfxCoreHelper.isCooperativeDispatchSupportedValue = isCooperativeDispatchSupported;
if (isCooperativeDispatchSupported || useDebugFlagForceTheoreticalMaxWgCount || useDebugFlagOverrideMaxWgCount) {
EXPECT_EQ(CL_SUCCESS, getClEnqueueNDCountKernelINTELResult());
} else {
EXPECT_EQ(CL_INVALID_COMMAND_QUEUE, getClEnqueueNDCountKernelINTELResult());
}
}
}
}
}
HWTEST_F(EnqueueKernelTest, WhenEnqueingKernelThenTaskLevelIsIncremented) {
auto taskLevelBefore = pCmdQ->taskLevel;
callOneWorkItemNDRKernel();

4
opencl/test/unit_test/command_queue/sync_buffer_handler_tests.cpp
View File

@@ -6,6 +6,7 @@
*/
#include "shared/source/program/sync_buffer_handler.h"
#include "shared/test/common/helpers/debug_manager_state_restore.h"
#include "shared/test/common/mocks/ult_device_factory.h"
#include "shared/test/common/test_macros/hw_test.h"
@@ -65,6 +66,8 @@ class SyncBufferHandlerTest : public SyncBufferEnqueueHandlerTest {
template <typename FamilyType>
void setUpT() {
NEO::DebugManager.flags.ForceTheoreticalMaxWorkGroupCount.set(true);
SyncBufferEnqueueHandlerTest::SetUp();
kernelInternals = std::make_unique<MockKernelWithInternals>(*pClDevice, context);
kernelInternals->kernelInfo.kernelDescriptor.kernelAttributes.bufferAddressingMode = KernelDescriptor::Stateless;
@@ -112,6 +115,7 @@ class SyncBufferHandlerTest : public SyncBufferEnqueueHandlerTest {
std::unique_ptr<MockKernelWithInternals> kernelInternals;
MockKernel *kernel;
MockCommandQueue *commandQueue;
DebugManagerStateRestore restorer;
};
HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenAllocateSyncBufferPatchAndConcurrentKernelWhenEnqueuingKernelThenSyncBufferIsUsed) {