performance: Remove global fence from command stream on BMG

Related-To: NEO-14642

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2025-04-30 14:08:12 +00:00
committed by Compute-Runtime-Automation
parent 7db6402df9
commit 304fba1eba
6 changed files with 29 additions and 23 deletions

View File

@@ -20,7 +20,7 @@ bool ProductHelperHw<gfxProduct>::isBlitterForImagesSupported() const {
template <>
bool ProductHelperHw<gfxProduct>::isGlobalFenceInCommandStreamRequired(const HardwareInfo &hwInfo) const {
return !hwInfo.capabilityTable.isIntegratedDevice;
return false;
}
template <>

View File

@@ -3075,7 +3075,7 @@ HWTEST2_F(CommandStreamReceiverHwTest, givenDeviceToHostCopyWhenFenceIsRequiredT
fenceExpected &= getHelper<ProductHelper>().isDeviceToHostCopySignalingFenceRequired();
size_t expectedFenceCount = fenceExpected ? 1 : 0;
if (!pDevice->getHardwareInfo().capabilityTable.isIntegratedDevice) {
if (getHelper<ProductHelper>().isGlobalFenceInCommandStreamRequired(pDevice->getHardwareInfo())) {
expectedFenceCount += 2;
}

View File

@@ -828,7 +828,7 @@ HWTEST_F(DirectSubmissionTest,
EXPECT_EQ(0u, NEO::IoFunctions::mockFcloseCalled);
size_t expectedSize = Dispatcher::getSizePreemption() +
directSubmission.getSizeSemaphoreSection(false);
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
expectedSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -910,7 +910,7 @@ HWTEST_F(DirectSubmissionTest,
directSubmission.getDiagnosticModeSection();
expectedSize += expectedExecCount * (directSubmission.getSizeDispatch(false, false, directSubmission.dispatchMonitorFenceRequired(false)) - directSubmission.getSizeNewResourceHandler());
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
expectedSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -953,7 +953,7 @@ HWTEST_F(DirectSubmissionTest,
}
size_t cmdOffset = 0;
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
cmdOffset = directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -1019,7 +1019,7 @@ HWTEST_F(DirectSubmissionTest,
EXPECT_EQ(expectedDispatch, directSubmission.getSizeDispatch(false, false, directSubmission.dispatchMonitorFenceRequired(false)) - directSubmission.getSizeNewResourceHandler());
expectedSize += expectedExecCount * expectedDispatch;
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
expectedSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {

View File

@@ -49,7 +49,7 @@ struct DirectSubmissionDispatchMiMemFenceTest : public DirectSubmissionDispatchB
DirectSubmissionDispatchBufferTest::SetUp();
auto &productHelper = pDevice->getProductHelper();
miMemFenceSupported = pDevice->getHardwareInfo().capabilityTable.isIntegratedDevice ? false : productHelper.isGlobalFenceInDirectSubmissionRequired(pDevice->getHardwareInfo());
miMemFenceSupported = productHelper.isGlobalFenceInDirectSubmissionRequired(pDevice->getHardwareInfo());
auto &compilerProductHelper = pDevice->getCompilerProductHelper();
heaplessStateInit = compilerProductHelper.isHeaplessStateInitEnabled(compilerProductHelper.isHeaplessModeEnabled(*defaultHwInfo));
@@ -109,9 +109,18 @@ struct DirectSubmissionDispatchMiMemFenceTest : public DirectSubmissionDispatchB
EXPECT_EQ(expectedFenceCount, fenceCount);
EXPECT_EQ(expectedSysMemFenceCount, sysMemFenceCount);
} else {
EXPECT_EQ(-1, systemMemoryFenceId);
if (directSubmission.globalFenceAllocation) {
if (expectedSysMemFenceCount > 0) {
EXPECT_NE(-1, systemMemoryFenceId);
} else {
EXPECT_EQ(-1, systemMemoryFenceId);
}
EXPECT_EQ(expectedSysMemFenceCount, sysMemFenceCount);
} else {
EXPECT_EQ(-1, systemMemoryFenceId);
EXPECT_EQ(0u, sysMemFenceCount);
}
EXPECT_EQ(0u, fenceCount);
EXPECT_EQ(0u, sysMemFenceCount);
}
}
@@ -261,7 +270,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DirectSubmissionDispatchBufferTest,
directSubmission.getSizeSemaphoreSection(false) +
sizeof(MI_LOAD_REGISTER_IMM) +
sizeof(MI_LOAD_REGISTER_MEM);
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -640,7 +649,7 @@ HWTEST_F(DirectSubmissionDispatchBufferTest,
EXPECT_EQ(1u, directSubmission.submitCount);
size_t submitSize = RenderDispatcher<FamilyType>::getSizePreemption() +
directSubmission.getSizeSemaphoreSection(false);
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -694,7 +703,7 @@ HWTEST_F(DirectSubmissionDispatchBufferTest,
EXPECT_EQ(1u, directSubmission.handleResidencyCount);
size_t submitSize = directSubmission.getSizeDispatch(false, false, directSubmission.dispatchMonitorFenceRequired(false)) - directSubmission.getSizeNewResourceHandler();
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -720,7 +729,7 @@ HWTEST_F(DirectSubmissionDispatchBufferTest,
EXPECT_EQ(1u, directSubmission.submitCount);
size_t submitSize = RenderDispatcher<FamilyType>::getSizePreemption() +
directSubmission.getSizeSemaphoreSection(false);
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -771,7 +780,7 @@ HWTEST_F(DirectSubmissionDispatchBufferTest,
EXPECT_EQ(1u, directSubmission.handleResidencyCount);
size_t submitSize = directSubmission.getSizeDispatch(false, false, directSubmission.dispatchMonitorFenceRequired(false)) - directSubmission.getSizeNewResourceHandler();
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}
if (directSubmission.isRelaxedOrderingEnabled()) {
@@ -864,7 +873,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DirectSubmissionDispatchBufferTest,
directSubmission.getSizeSemaphoreSection(false) +
sizeof(MI_LOAD_REGISTER_IMM) +
sizeof(MI_LOAD_REGISTER_MEM);
if (directSubmission.miMemFenceRequired && !heaplessStateInit) {
if (directSubmission.globalFenceAllocation && !heaplessStateInit) {
submitSize += directSubmission.getSizeSystemMemoryFenceAddress();
}

View File

@@ -1368,7 +1368,7 @@ HWTEST_TEMPLATED_F(WddmCommandStreamMockGdiTest, givenDirectSubmissionEnabledOnB
auto &compilerProductHelper = device->getCompilerProductHelper();
auto heaplessStateInit = compilerProductHelper.isHeaplessStateInitEnabled(compilerProductHelper.isHeaplessModeEnabled(*defaultHwInfo));
if (directSubmission->miMemFenceRequired && !heaplessStateInit) {
if (directSubmission->globalFenceAllocation && !heaplessStateInit) {
expectedSize += directSubmission->getSizeSystemMemoryFenceAddress();
}
if (directSubmission->isRelaxedOrderingEnabled()) {

View File

@@ -520,17 +520,14 @@ XE2_HPG_CORETEST_F(GfxCoreHelperTestsXe2HpgCore, givenGfxCoreHelperWhenAskedIfFe
}
XE2_HPG_CORETEST_F(GfxCoreHelperTestsXe2HpgCore, givenDefaultMemorySynchronizationCommandsWhenGettingSizeForAdditionalSynchronizationThenCorrectValueIsReturned) {
using MI_MEM_FENCE = typename FamilyType::MI_MEM_FENCE;
EXPECT_EQ(!pDevice->getHardwareInfo().capabilityTable.isIntegratedDevice * sizeof(MI_MEM_FENCE), MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getRootDeviceEnvironment()));
EXPECT_EQ(0u, MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getRootDeviceEnvironment()));
}
XE2_HPG_CORETEST_F(GfxCoreHelperTestsXe2HpgCore, givenDebugMemorySynchronizationCommandsWhenGettingSizeForAdditionalSynchronizationThenCorrectValueIsReturned) {
DebugManagerStateRestore restorer;
debugManager.flags.DisablePipeControlPrecedingPostSyncCommand.set(1);
using MI_MEM_FENCE = typename FamilyType::MI_MEM_FENCE;
EXPECT_EQ(!pDevice->getHardwareInfo().capabilityTable.isIntegratedDevice * 2 * sizeof(MI_MEM_FENCE), MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getRootDeviceEnvironment()));
EXPECT_EQ(0u, MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getRootDeviceEnvironment()));
}
XE2_HPG_CORETEST_F(GfxCoreHelperTestsXe2HpgCore, givenDontProgramGlobalFenceAsMiMemFenceCommandInCommandStreamWhenGettingSizeForAdditionalSynchronizationThenCorrectValueIsReturned) {
@@ -680,9 +677,9 @@ XE2_HPG_CORETEST_F(ProductHelperTestXe2HpgCore, givenProductHelperWhenCallUseGem
EXPECT_TRUE(productHelper.useGemCreateExtInAllocateMemoryByKMD());
}
XE2_HPG_CORETEST_F(ProductHelperTestXe2HpgCore, givenProductHelperWhenAskingForGlobalFenceSupportThenReturnTrue) {
XE2_HPG_CORETEST_F(ProductHelperTestXe2HpgCore, givenProductHelperWhenAskingForGlobalFenceSupportThenReturnFalse) {
const auto &productHelper = getHelper<ProductHelper>();
EXPECT_EQ(productHelper.isGlobalFenceInCommandStreamRequired(*defaultHwInfo), !defaultHwInfo->capabilityTable.isIntegratedDevice);
EXPECT_FALSE(productHelper.isGlobalFenceInCommandStreamRequired(*defaultHwInfo));
}
XE2_HPG_CORETEST_F(ProductHelperTestXe2HpgCore, givenProductHelperWhenAskingForCooperativeEngineSupportThenReturnTrue) {