Dont inject commands between blits

Signed-off-by: Bartosz Dunajski <[email protected]>

Author: BartoszDunajski

level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_2.cpp

@@ -846,7 +846,6 @@ using BlitBlockCopyPlatforms = IsWithinProducts<IGFX_SKYLAKE, IGFX_TIGERLAKE_LP>
 HWTEST2_F(CommandListCreate, givenCopyCommandListWhenCopyRegionWithinMaxBlitSizeThenOneBlitCommandHasBeenSpown, BlitBlockCopyPlatforms) {
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
     using XY_COPY_BLT = typename GfxFamily::XY_COPY_BLT;
-    using MI_ARB_CHECK = typename GfxFamily::MI_ARB_CHECK;
 
     auto commandList = std::make_unique<WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>>();
     commandList->initialize(device, NEO::EngineGroupType::Copy);
@@ -872,9 +871,8 @@ HWTEST2_F(CommandListCreate, givenCopyCommandListWhenCopyRegionWithinMaxBlitSize
         cmdList, ptrOffset(commandList->commandContainer.getCommandStream()->getCpuBase(), 0), commandList->commandContainer.getCommandStream()->getUsed()));
     auto itor = find<XY_COPY_BLT *>(cmdList.begin(), cmdList.end());
     EXPECT_NE(cmdList.end(), itor);
-    itor = find<MI_ARB_CHECK *>(cmdList.begin(), cmdList.end());
-    EXPECT_NE(cmdList.end(), itor);
     itor++;
+    itor = find<XY_COPY_BLT *>(itor, cmdList.end());
     EXPECT_EQ(cmdList.end(), itor);
 }
 
@@ -1094,10 +1092,15 @@ HWTEST_F(CommandListCreate, GivenCommandListWhenUnalignedPtrThenLeftMiddleAndRig
 
     ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
         cmdList, ptrOffset(commandList->commandContainer.getCommandStream()->getCpuBase(), 0), commandList->commandContainer.getCommandStream()->getUsed()));
+
     auto itor = find<XY_COPY_BLT *>(cmdList.begin(), cmdList.end());
-    for (uint32_t i = 0; i < 3u; i++, itor++) {
-        EXPECT_NE(itor++, cmdList.end());
-    }
+    EXPECT_NE(cmdList.end(), itor);
+
+    itor = find<XY_COPY_BLT *>(++itor, cmdList.end());
+    EXPECT_NE(cmdList.end(), itor);
+
+    itor = find<XY_COPY_BLT *>(++itor, cmdList.end());
+    EXPECT_NE(cmdList.end(), itor);
 }
 } // namespace ult
 } // namespace L0

opencl/test/unit_test/command_stream/command_stream_receiver_hw_1_tests.cpp

@@ -612,7 +612,7 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenEstimatingCommandsSizeThenCal
     const uint32_t numberOfBlitOperations = 4;
 
     auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_END);
-    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
 
     auto expectedAlignedSize = baseSize + MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getHardwareInfo());
@@ -641,7 +641,7 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenDirectsubmissionEnabledEstima
     const uint32_t numberOfBlitOperations = 4;
 
     auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_START);
-    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
 
     auto expectedAlignedSize = baseSize + MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getHardwareInfo());
@@ -668,7 +668,7 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenEstimatingCommandsSizeForWrit
     const Vec3<size_t> bltSize = {(3 * max2DBlitSize), 4, 2};
     const size_t numberOfBlts = 3 * bltSize.y * bltSize.z;
     const size_t numberOfBlitOperations = 4 * bltSize.y * bltSize.z;
-    const size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    const size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
 
     auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_END);
     auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
@@ -697,7 +697,7 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenDirectSubmissionEnabledEstima
     const Vec3<size_t> bltSize = {(3 * max2DBlitSize), 4, 2};
     const size_t numberOfBlts = 3 * bltSize.y * bltSize.z;
     const size_t numberOfBlitOperations = 4 * bltSize.y * bltSize.z;
-    const size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    const size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
 
     auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_START);
     auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
@@ -722,7 +722,7 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenDirectSubmissionEnabledEstima
 }
 
 HWTEST_F(BcsTests, givenTimestampPacketWriteRequestWhenEstimatingSizeForCommandsThenAddMiFlushDw) {
-    constexpr size_t expectedBaseSize = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr size_t expectedBaseSize = sizeof(typename FamilyType::XY_COPY_BLT);
 
     auto expectedSizeWithTimestampPacketWrite = expectedBaseSize + EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite();
     auto expectedSizeWithoutTimestampPacketWrite = expectedBaseSize;
@@ -737,7 +737,7 @@ HWTEST_F(BcsTests, givenTimestampPacketWriteRequestWhenEstimatingSizeForCommands
 }
 
 HWTEST_F(BcsTests, givenTimestampPacketWriteRequestWhenEstimatingSizeForCommandsWithProfilingThenAddMiFlushDw) {
-    size_t expectedBaseSize = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK) + EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite();
+    size_t expectedBaseSize = sizeof(typename FamilyType::XY_COPY_BLT) + EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite();
 
     auto expectedSizeWithTimestampPacketWriteAndProfiling = expectedBaseSize + 4 * sizeof(typename FamilyType::MI_STORE_REGISTER_MEM);
 
@@ -760,7 +760,7 @@ HWTEST_F(BcsTests, givenBltSizeAndCsrDependenciesWhenEstimatingCommandSizeThenAd
     csrDependencies.push_back(&timestamp0);
     csrDependencies.push_back(&timestamp1);
 
-    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     size_t expectedSize = (cmdsSizePerBlit * numberOfBlts) +
                           TimestampPacketHelper::getRequiredCmdStreamSize<FamilyType>(csrDependencies);
 
@@ -822,10 +822,6 @@ HWTEST_F(BcsTests, givenBltSizeWithLeftoverWhenDispatchedThenProgramAllRequiredC
         EXPECT_EQ(expectedHeight, bltCmd->getTransferHeight());
         EXPECT_EQ(expectedWidth, bltCmd->getDestinationPitch());
         EXPECT_EQ(expectedWidth, bltCmd->getSourcePitch());
-
-        auto miArbCheckCmd = genCmdCast<typename FamilyType::MI_ARB_CHECK *>(*(cmdIterator++));
-        EXPECT_NE(nullptr, miArbCheckCmd);
-        EXPECT_TRUE(memcmp(&FamilyType::cmdInitArbCheck, miArbCheckCmd, sizeof(typename FamilyType::MI_ARB_CHECK)) == 0);
     }
 
     if (UnitTestHelper<FamilyType>::isAdditionalSynchronizationRequired(pDevice->getHardwareInfo())) {
@@ -1045,10 +1041,6 @@ HWTEST_P(BcsDetaliedTestsWithParams, givenBltSizeWithLeftoverWhenDispatchedThenP
         EXPECT_EQ(srcAddr, expectedSrcAddr);
 
         offset += (expectedWidth * expectedHeight);
-
-        auto miArbCheckCmd = genCmdCast<typename FamilyType::MI_ARB_CHECK *>(*(cmdIterator++));
-        EXPECT_NE(nullptr, miArbCheckCmd);
-        EXPECT_TRUE(memcmp(&FamilyType::cmdInitArbCheck, miArbCheckCmd, sizeof(typename FamilyType::MI_ARB_CHECK)) == 0);
     }
 }
 
@@ -1128,10 +1120,6 @@ HWTEST_P(BcsDetaliedTestsWithParams, givenBltSizeWithLeftoverWhenDispatchedThenP
         EXPECT_EQ(srcAddr, bltCmd->getSourceBaseAddress());
 
         offset += (expectedWidth * expectedHeight);
-
-        auto miArbCheckCmd = genCmdCast<typename FamilyType::MI_ARB_CHECK *>(*(cmdIterator++));
-        EXPECT_NE(nullptr, miArbCheckCmd);
-        EXPECT_TRUE(memcmp(&FamilyType::cmdInitArbCheck, miArbCheckCmd, sizeof(typename FamilyType::MI_ARB_CHECK)) == 0);
     }
 }
 
@@ -1206,10 +1194,6 @@ HWTEST_P(BcsDetaliedTestsWithParams, givenBltSizeWithLeftoverWhenDispatchedThenP
         EXPECT_EQ(srcAddr, bltCmd->getSourceBaseAddress());
 
         offset += (expectedWidth * expectedHeight);
-
-        auto miArbCheckCmd = genCmdCast<typename FamilyType::MI_ARB_CHECK *>(*(cmdIterator++));
-        EXPECT_NE(nullptr, miArbCheckCmd);
-        EXPECT_TRUE(memcmp(&FamilyType::cmdInitArbCheck, miArbCheckCmd, sizeof(typename FamilyType::MI_ARB_CHECK)) == 0);
     }
 }
 

opencl/test/unit_test/command_stream/command_stream_receiver_hw_2_tests.cpp

@@ -31,7 +31,7 @@ using namespace NEO;
 
 HWTEST_F(BcsTests, givenBltSizeWhenEstimatingCommandSizeThenAddAllRequiredCommands) {
     constexpr auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
-    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     size_t notAlignedBltSize = (3 * max2DBlitSize) + 1;
     size_t alignedBltSize = (3 * max2DBlitSize);
     uint32_t alignedNumberOfBlts = 3;
@@ -55,7 +55,7 @@ HWTEST_F(BcsTests, givenBltSizeWhenEstimatingCommandSizeThenAddAllRequiredComman
 
 HWTEST_F(BcsTests, givenDebugCapabilityWhenEstimatingCommandSizeThenAddAllRequiredCommands) {
     constexpr auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
-    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     const size_t debugCommandsSize = (EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + EncodeSempahore<FamilyType>::getSizeMiSemaphoreWait()) * 2;
 
     constexpr uint32_t numberOfBlts = 3;
@@ -79,7 +79,7 @@ HWTEST_F(BcsTests, givenDebugCapabilityWhenEstimatingCommandSizeThenAddAllRequir
 
 HWTEST_F(BcsTests, givenBltSizeWhenEstimatingCommandSizeForReadBufferRectThenAddAllRequiredCommands) {
     constexpr auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
-    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     Vec3<size_t> notAlignedBltSize = {(3 * max2DBlitSize) + 1, 4, 2};
     Vec3<size_t> alignedBltSize = {(3 * max2DBlitSize), 4, 2};
     size_t alignedNumberOfBlts = 3 * alignedBltSize.y * alignedBltSize.z;
@@ -104,7 +104,7 @@ HWTEST_F(BcsTests, givenBltWithBigCopySizeWhenEstimatingCommandSizeForReadBuffer
     auto maxWidthToCopy = static_cast<size_t>(BlitCommandsHelper<FamilyType>::getMaxBlitWidth(rootDeviceEnvironment));
     auto maxHeightToCopy = static_cast<size_t>(BlitCommandsHelper<FamilyType>::getMaxBlitHeight(rootDeviceEnvironment));
 
-    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    constexpr auto cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT);
     Vec3<size_t> alignedBltSize = {(3 * maxWidthToCopy), (4 * maxHeightToCopy), 2};
     Vec3<size_t> notAlignedBltSize = {(3 * maxWidthToCopy + 1), (4 * maxHeightToCopy), 2};
 

shared/source/helpers/blit_commands_helper_base.inl

@@ -42,30 +42,23 @@ uint64_t BlitCommandsHelper<GfxFamily>::getMaxBlitHeight(const RootDeviceEnviron
 
 template <typename GfxFamily>
 void BlitCommandsHelper<GfxFamily>::dispatchPostBlitCommand(LinearStream &linearStream) {
-    switch (DebugManager.flags.PostBlitCommand.get()) {
-    case 1:
-        EncodeMiFlushDW<GfxFamily>::programMiFlushDw(linearStream, 0, 0, false, false);
-        break;
-    case 2:
-        break;
-    default: {
+    if (DebugManager.flags.PostBlitCommand.get() == 0) {
         auto miArbCheckStream = linearStream.getSpaceForCmd<typename GfxFamily::MI_ARB_CHECK>();
         *miArbCheckStream = GfxFamily::cmdInitArbCheck;
-        break;
-    }
+    } else if (DebugManager.flags.PostBlitCommand.get() == 1) {
+        EncodeMiFlushDW<GfxFamily>::programMiFlushDw(linearStream, 0, 0, false, false);
     }
 }
 
 template <typename GfxFamily>
 size_t BlitCommandsHelper<GfxFamily>::estimatePostBlitCommandSize() {
-    switch (DebugManager.flags.PostBlitCommand.get()) {
-    case 1:
-        return sizeof(typename GfxFamily::MI_FLUSH_DW);
-    case 2:
-        return 0;
-    default:
+    if (DebugManager.flags.PostBlitCommand.get() == 0) {
         return sizeof(typename GfxFamily::MI_ARB_CHECK);
+    } else if (DebugManager.flags.PostBlitCommand.get() == 1) {
+        return sizeof(typename GfxFamily::MI_FLUSH_DW);
     }
+
+    return 0;
 }
 
 template <typename GfxFamily>

shared/test/unit_test/helpers/blit_commands_helper_tests.cpp

@@ -96,7 +96,8 @@ HWTEST_F(BlitTests, givenDebugVariableWhenEstimatingPostBlitsCommandSizeThenRetu
 
     DebugManagerStateRestore restore{};
 
-    EXPECT_EQ(arbCheckSize, BlitCommandsHelper<FamilyType>::estimatePostBlitCommandSize());
+    EXPECT_EQ(0u, BlitCommandsHelper<FamilyType>::estimatePostBlitCommandSize());
+
     DebugManager.flags.PostBlitCommand.set(0);
     EXPECT_EQ(arbCheckSize, BlitCommandsHelper<FamilyType>::estimatePostBlitCommandSize());
 
@@ -117,9 +118,7 @@ HWTEST_F(BlitTests, givenDebugVariableWhenDispatchingPostBlitsCommandThenUseCorr
 
     // -1: default
     BlitCommandsHelper<FamilyType>::dispatchPostBlitCommand(linearStream);
-    CmdParse<FamilyType>::parseCommandBuffer(commands, linearStream.getCpuBase(), linearStream.getUsed());
-    auto arbCheck = find<MI_ARB_CHECK *>(commands.begin(), commands.end());
-    EXPECT_NE(commands.end(), arbCheck);
+    EXPECT_EQ(0u, linearStream.getUsed());
 
     // 0: MI_ARB_CHECK
     memset(streamBuffer, 0, sizeof(streamBuffer));
@@ -128,7 +127,7 @@ HWTEST_F(BlitTests, givenDebugVariableWhenDispatchingPostBlitsCommandThenUseCorr
     DebugManager.flags.PostBlitCommand.set(0);
     BlitCommandsHelper<FamilyType>::dispatchPostBlitCommand(linearStream);
     CmdParse<FamilyType>::parseCommandBuffer(commands, linearStream.getCpuBase(), linearStream.getUsed());
-    arbCheck = find<MI_ARB_CHECK *>(commands.begin(), commands.end());
+    auto arbCheck = find<MI_ARB_CHECK *>(commands.begin(), commands.end());
     EXPECT_NE(commands.end(), arbCheck);
 
     // 1: MI_FLUSH_DW