Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Disable host ptr tracking for GEN12LP 

Related-To: NEO-3661

Change-Id: I1774aad8ef32f294192a9b679cb9f66e4091198c
Signed-off-by: Pawel Wilma <pawel.wilma@intel.com>
This commit is contained in:
Pawel Wilma 2019-12-04 13:42:41 +01:00 committed by sys_ocldev
parent cd198fac0a
commit 632134db30
24 changed files with 68 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
core/helpers/hw_info.h
View File

@ -51,6 +51,7 @@ struct RuntimeCapabilityTable {
bool supportCacheFlushAfterWalker;
bool supportsImages;
bool supportsDeviceEnqueue;
bool hostPtrTrackingEnabled;
};
struct HardwareCapabilities {

1
core/memory_manager/graphics_allocation.h
View File

@ -56,6 +56,7 @@ class GraphicsAllocation : public IDNode<GraphicsAllocation> {
INTERNAL_HOST_MEMORY,
KERNEL_ISA,
LINEAR_STREAM,
MAP_ALLOCATION,
MCS,
PIPE,
PREEMPTION,

3
runtime/gen11/hw_info_ehl.inl
View File

@ -64,7 +64,8 @@ const RuntimeCapabilityTable EHL::capabilityTable{
false, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable EHL::workaroundTable = {};

3
runtime/gen11/hw_info_icllp.inl
View File

@ -66,7 +66,8 @@ const RuntimeCapabilityTable ICLLP::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable ICLLP::workaroundTable = {};

3
runtime/gen11/hw_info_lkf.inl
View File

@ -65,7 +65,8 @@ const RuntimeCapabilityTable LKF::capabilityTable{
false, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable LKF::workaroundTable = {};

3
runtime/gen12lp/hw_info_tgllp.inl
View File

@ -66,7 +66,8 @@ const RuntimeCapabilityTable TGLLP::capabilityTable{
false, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
false // hostPtrTrackingEnabled
};
WorkaroundTable TGLLP::workaroundTable = {};

3
runtime/gen8/hw_info_bdw.inl
View File

@ -70,7 +70,8 @@ const RuntimeCapabilityTable BDW::capabilityTable{
false, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable BDW::workaroundTable = {};

3
runtime/gen9/hw_info_bxt.inl
View File

@ -67,7 +67,8 @@ const RuntimeCapabilityTable BXT::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
false // supportsDeviceEnqueue
false, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable BXT::workaroundTable = {};

3
runtime/gen9/hw_info_cfl.inl
View File

@ -62,7 +62,8 @@ const RuntimeCapabilityTable CFL::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable CFL::workaroundTable = {};

3
runtime/gen9/hw_info_glk.inl
View File

@ -62,7 +62,8 @@ const RuntimeCapabilityTable GLK::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
false // supportsDeviceEnqueue
false, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable GLK::workaroundTable = {};

3
runtime/gen9/hw_info_kbl.inl
View File

@ -62,7 +62,8 @@ const RuntimeCapabilityTable KBL::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable KBL::workaroundTable = {};

3
runtime/gen9/hw_info_skl.inl
View File

@ -70,7 +70,8 @@ const RuntimeCapabilityTable SKL::capabilityTable{
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true // supportsDeviceEnqueue
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable SKL::workaroundTable = {};
FeatureTable SKL::featureTable = {};

2
runtime/mem_obj/buffer.cpp
View File

@ -283,7 +283,7 @@ Buffer *Buffer::create(Context *context,
if (memoryProperties.flags.useHostPtr) {
if (!zeroCopyAllowed && !isHostPtrSVM) {
AllocationProperties properties{rootDeviceIndex, false, size, GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR, false};
AllocationProperties properties{rootDeviceIndex, false, size, GraphicsAllocation::AllocationType::MAP_ALLOCATION, false};
properties.flags.flushL3RequiredForRead = properties.flags.flushL3RequiredForWrite = true;
mapAllocation = memoryManager->allocateGraphicsMemoryWithProperties(properties, hostPtr);
}

4
runtime/mem_obj/image.cpp
View File

@ -281,7 +281,7 @@ Image *Image::create(Context *context,
zeroCopy = true;
}
if (memory) {
AllocationProperties properties{rootDeviceIndex, false, hostPtrMinSize, GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR, false};
AllocationProperties properties{rootDeviceIndex, false, hostPtrMinSize, GraphicsAllocation::AllocationType::MAP_ALLOCATION, false};
properties.flags.flushL3RequiredForRead = properties.flags.flushL3RequiredForWrite = true;
mapAllocation = memoryManager->allocateGraphicsMemoryWithProperties(properties, hostPtr);
}
@ -380,7 +380,7 @@ Image *Image::create(Context *context,
} else {
errcodeRet = cmdQ->enqueueWriteImage(image, CL_TRUE, &copyOrigin[0], &copyRegion[0],
hostPtrRowPitch, hostPtrSlicePitch,
hostPtr, nullptr, 0, nullptr, nullptr);
hostPtr, mapAllocation, 0, nullptr, nullptr);
}
} else {
image->transferData(memory->getUnderlyingBuffer(), imgInfo.rowPitch, imgInfo.slicePitch,

3
runtime/mem_obj/mem_obj.cpp
View File

@ -332,7 +332,8 @@ void *MemObj::getBasePtrForMap(uint32_t rootDeviceIndex) {
} else {
auto memory = memoryManager->allocateSystemMemory(getSize(), MemoryConstants::pageSize);
setAllocatedMapPtr(memory);
AllocationProperties properties{rootDeviceIndex, false, getSize(), GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR, false};
AllocationProperties properties{rootDeviceIndex, false, getSize(), GraphicsAllocation::AllocationType::MAP_ALLOCATION, false};
auto allocation = memoryManager->allocateGraphicsMemoryWithProperties(properties, memory);
setMapAllocation(allocation);
return getAllocatedMapPtr();

11
runtime/memory_manager/memory_manager.cpp
View File

@ -260,6 +260,7 @@ bool MemoryManager::getAllocationData(AllocationData &allocationData, const Allo
case GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR:
case GraphicsAllocation::AllocationType::GLOBAL_SURFACE:
case GraphicsAllocation::AllocationType::IMAGE:
case GraphicsAllocation::AllocationType::MAP_ALLOCATION:
case GraphicsAllocation::AllocationType::PIPE:
case GraphicsAllocation::AllocationType::SHARED_BUFFER:
case GraphicsAllocation::AllocationType::SHARED_IMAGE:
@ -278,6 +279,7 @@ bool MemoryManager::getAllocationData(AllocationData &allocationData, const Allo
case GraphicsAllocation::AllocationType::DEVICE_QUEUE_BUFFER:
case GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR:
case GraphicsAllocation::AllocationType::FILL_PATTERN:
case GraphicsAllocation::AllocationType::MAP_ALLOCATION:
case GraphicsAllocation::AllocationType::MCS:
case GraphicsAllocation::AllocationType::PREEMPTION:
case GraphicsAllocation::AllocationType::PROFILING_TAG_BUFFER:
@ -340,7 +342,7 @@ GraphicsAllocation *MemoryManager::allocateGraphicsMemory(const AllocationData &
return allocateGraphicsMemoryForImage(allocationData);
}
if (allocationData.type == GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR &&
(!peekExecutionEnvironment().isFullRangeSvm() || !DebugManager.flags.EnableHostPtrTracking.get())) {
(!peekExecutionEnvironment().isFullRangeSvm() || !isHostPointerTrackingEnabled())) {
auto allocation = allocateGraphicsMemoryForNonSvmHostPtr(allocationData);
if (allocation) {
allocation->setFlushL3Required(allocationData.flags.flushL3);
@ -485,4 +487,11 @@ void *MemoryManager::getReservedMemory(size_t size, size_t alignment) {
return reservedMemory;
}
bool MemoryManager::isHostPointerTrackingEnabled() {
if (DebugManager.flags.EnableHostPtrTracking.get() != -1) {
return !!DebugManager.flags.EnableHostPtrTracking.get();
}
return (peekExecutionEnvironment().getHardwareInfo()->capabilityTable.hostPtrTrackingEnabled | is32bit);
}
} // namespace NEO

2
runtime/memory_manager/memory_manager.h
View File

@ -140,6 +140,8 @@ class MemoryManager {
::alignedFree(ptr);
}
MOCKABLE_VIRTUAL bool isHostPointerTrackingEnabled();
const ExecutionEnvironment &peekExecutionEnvironment() const { return executionEnvironment; }
OsContext *createAndRegisterOsContext(CommandStreamReceiver *commandStreamReceiver, aub_stream::EngineType engineType,

2
runtime/os_interface/debug_variables_base.inl
View File

@ -87,7 +87,7 @@ DECLARE_DEBUG_VARIABLE(bool, DisableConcurrentBlockExecution, false, "disables c
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(bool, DisableZeroCopyForBuffers, false, "When active all buffer allocations will not share memory with CPU.")
DECLARE_DEBUG_VARIABLE(bool, EnableHostPtrTracking, true, "Enable host ptr tracking")
DECLARE_DEBUG_VARIABLE(int32_t, EnableHostPtrTracking, -1, "Enable host ptr tracking: -1 - default platform setting, 0 - disabled, 1 - enabled")
DECLARE_DEBUG_VARIABLE(bool, DisableDcFlushInEpilogue, false, "Disable DC flush in epilogue")
/*FEATURE FLAGS*/

2
runtime/os_interface/windows/wddm_memory_manager.cpp
View File

@ -143,7 +143,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(co
wddmAllocation->setDefaultGmm(gmm);
if (!createWddmAllocation(wddmAllocation.get(), wddmAllocation->getAlignedCpuPtr())) {
if (!createWddmAllocation(wddmAllocation.get(), nullptr)) {
delete gmm;
return nullptr;
}

4
unit_tests/command_stream/command_stream_receiver_tests.cpp
View File

@ -419,7 +419,7 @@ TEST_F(CreateAllocationForHostSurfaceTest, givenReadOnlyHostPointerWhenAllocatio
size_t size = sizeof(memory);
HostPtrSurface surface(const_cast<char *>(memory), size, true);
if (device->isFullRangeSvm()) {
if (device->isFullRangeSvm() && gmockMemoryManager->isHostPointerTrackingEnabled()) {
EXPECT_CALL(*gmockMemoryManager, populateOsHandles(::testing::_))
.Times(1)
.WillOnce(::testing::Return(MemoryManager::AllocationStatus::InvalidHostPointer));
@ -446,7 +446,7 @@ TEST_F(CreateAllocationForHostSurfaceTest, givenReadOnlyHostPointerWhenAllocatio
size_t size = sizeof(memory);
HostPtrSurface surface(const_cast<char *>(memory), size, false);
if (device->isFullRangeSvm()) {
if (device->isFullRangeSvm() && gmockMemoryManager->isHostPointerTrackingEnabled()) {
EXPECT_CALL(*gmockMemoryManager, populateOsHandles(::testing::_))
.Times(1)
.WillOnce(::testing::Return(MemoryManager::AllocationStatus::InvalidHostPointer));

2
unit_tests/mem_obj/buffer_tests.cpp
View File

@ -293,7 +293,7 @@ TEST(Buffer, givenHostPtrPassedToBufferCreateWhenMemUseHostPtrFlagisSetAndBuffer
auto mapAllocation = buffer->getMapAllocation();
EXPECT_NE(nullptr, mapAllocation);
EXPECT_EQ(offsetedPtr, mapAllocation->getUnderlyingBuffer());
EXPECT_EQ(GraphicsAllocation::AllocationType::EXTERNAL_HOST_PTR, mapAllocation->getAllocationType());
EXPECT_EQ(GraphicsAllocation::AllocationType::MAP_ALLOCATION, mapAllocation->getAllocationType());
alignedFree(ptr);
}

24
unit_tests/memory_manager/memory_manager_tests.cpp
View File

@ -1183,7 +1183,7 @@ TEST_P(OsAgnosticMemoryManagerWithParams, givenFullGpuAddressSpaceWhenAllocateGr
bool requiresL3Flush = GetParam();
MockExecutionEnvironment executionEnvironment(*platformDevices);
executionEnvironment.setHwInfo(platformDevices[0]);
if (!executionEnvironment.isFullRangeSvm()) {
if (!executionEnvironment.isFullRangeSvm() || !executionEnvironment.getHardwareInfo()->capabilityTable.hostPtrTrackingEnabled) {
return;
}
OsAgnosticMemoryManager memoryManager(executionEnvironment);
@ -1201,7 +1201,7 @@ TEST_P(OsAgnosticMemoryManagerWithParams, givenFullGpuAddressSpaceWhenAllocateGr
TEST_P(OsAgnosticMemoryManagerWithParams, givenDisabledHostPtrTrackingWhenAllocateGraphicsMemoryForHostPtrIsCalledThenAllocationWithoutFragmentsIsCreated) {
DebugManagerStateRestore restore;
DebugManager.flags.EnableHostPtrTracking.set(false);
DebugManager.flags.EnableHostPtrTracking.set(0);
bool requiresL3Flush = GetParam();
MockExecutionEnvironment executionEnvironment(*platformDevices);
@ -1945,3 +1945,23 @@ TEST(MemoryManagerTest, whenMemoryManagerReturnsNullptrThenAllocateGlobalsSurfac
allocation = allocateGlobalsSurface(&context, &device, 1024, false, true, nullptr);
EXPECT_EQ(nullptr, allocation);
}
HWTEST_F(MemoryAllocatorTest, givenMemoryManagerWhenEnableHostPtrTrackingFlagIsSetTo0ThenHostPointerTrackingIsDisabled) {
DebugManagerStateRestore dbgRestore;
DebugManager.flags.EnableHostPtrTracking.set(0);
EXPECT_FALSE(memoryManager->isHostPointerTrackingEnabled());
}
HWTEST_F(MemoryAllocatorTest, givenMemoryManagerWhenEnableHostPtrTrackingFlagIsNotSetTo1ThenHostPointerTrackingIsEnabled) {
DebugManagerStateRestore dbgRestore;
DebugManager.flags.EnableHostPtrTracking.set(1);
EXPECT_TRUE(memoryManager->isHostPointerTrackingEnabled());
}
HWTEST_F(MemoryAllocatorTest, givenMemoryManagerWhenEnableHostPtrTrackingFlagIsSetNotSetThenHostPointerTrackingDependsOnCapabilityTable) {
if (is32bit) {
EXPECT_TRUE(memoryManager->isHostPointerTrackingEnabled());
} else {
EXPECT_EQ(executionEnvironment->getHardwareInfo()->capabilityTable.hostPtrTrackingEnabled, memoryManager->isHostPointerTrackingEnabled());
}
}

2
unit_tests/scenarios/windows/enqueue_read_write_buffer_scenarios_windows_tests.cpp
View File

@ -108,7 +108,7 @@ HWTEST_F(EnqueueBufferWindowsTest, givenMisalignedHostPtrWhenEnqueueReadBufferCa
if (kernel->getKernelInfo().kernelArgInfo[1].kernelArgPatchInfoVector[0].size == sizeof(uint64_t)) {
auto pKernelArg = (uint64_t *)(kernel->getCrossThreadData() +
kernel->getKernelInfo().kernelArgInfo[1].kernelArgPatchInfoVector[0].crossthreadOffset);
EXPECT_EQ(reinterpret_cast<uint64_t>(alignDown(misalignedPtr, 4)), *pKernelArg);
EXPECT_EQ(alignDown(gpuVa, 4), static_cast<uint64_t>(*pKernelArg));
EXPECT_EQ(*pKernelArg, surfaceStateDst.getSurfaceBaseAddress());
} else if (kernel->getKernelInfo().kernelArgInfo[1].kernelArgPatchInfoVector[0].size == sizeof(uint32_t)) {

2
unit_tests/test_files/igdrcl.config
View File

@ -108,7 +108,7 @@ RenderCompressedBuffersEnabled = -1
AUBDumpAllocsOnEnqueueReadOnly = 0
AUBDumpForceAllToLocalMemory = 0
EnableCacheFlushAfterWalker = -1
EnableHostPtrTracking = 1
EnableHostPtrTracking = -1
DisableDcFlushInEpilogue = 0
OverrideInvalidEngineWithDefault = 0
EnableFormatQuery = 0