intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 14:55:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

Introduce RegisteredMethodDispatcher.

Browse Source 
- Inject dispatch methods per DispatchInfo
- Each DispatchInfo in MultiDispatchInfo can have different behaviour
- Implement AuxTranslation programming with new approach

Change-Id: Ie28de0c72a77b8e91509a5b9b8740d72fedf4ad6
Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2019-07-09 14:24:33 +02:00
committed by sys_ocldev
parent a72109d209
commit ddb0d82e83
15 changed files with 241 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
unit_tests/built_ins/built_in_tests.cpp
View File

@@ -246,8 +246,9 @@ TEST_F(BuiltInTests, BuiltinDispatchInfoBuilderCopyBufferToBuffer) {
delete dstPtr;
}
TEST_F(BuiltInTests, givenInputBufferWhenBuildingNonAuxDispatchInfoForAuxTranslationThenPickAndSetupCorrectKernels) {
BuiltinDispatchInfoBuilder &builder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
HWTEST_F(BuiltInTests, givenInputBufferWhenBuildingNonAuxDispatchInfoForAuxTranslationThenPickAndSetupCorrectKernels) {
BuiltinDispatchInfoBuilder &baseBuilder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MemObjsForAuxTranslation memObjsForAuxTranslation;
MultiDispatchInfo multiDispatchInfo;
@@ -265,7 +266,7 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingNonAuxDispatchInfoForAuxTransla
memObjsForAuxTranslation.insert(&buffer);
}
EXPECT_TRUE(builder.buildDispatchInfos(multiDispatchInfo, builtinOpsParams));
EXPECT_TRUE(builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams));
EXPECT_EQ(3u, multiDispatchInfo.size());
for (auto &dispatchInfo : multiDispatchInfo) {
@@ -294,8 +295,9 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingNonAuxDispatchInfoForAuxTransla
EXPECT_NE(builtinKernels[1], builtinKernels[2]);
}
TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxDispatchInfoForAuxTranslationThenPickAndSetupCorrectKernels) {
BuiltinDispatchInfoBuilder &builder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
HWTEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxDispatchInfoForAuxTranslationThenPickAndSetupCorrectKernels) {
BuiltinDispatchInfoBuilder &baseBuilder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MemObjsForAuxTranslation memObjsForAuxTranslation;
MultiDispatchInfo multiDispatchInfo;
@@ -313,7 +315,7 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxDispatchInfoForAuxTranslatio
memObjsForAuxTranslation.insert(&buffer);
}
EXPECT_TRUE(builder.buildDispatchInfos(multiDispatchInfo, builtinOpsParams));
EXPECT_TRUE(builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams));
EXPECT_EQ(3u, multiDispatchInfo.size());
for (auto &dispatchInfo : multiDispatchInfo) {
@@ -342,8 +344,9 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxDispatchInfoForAuxTranslatio
EXPECT_NE(builtinKernels[1], builtinKernels[2]);
}
TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxTranslationDispatchThenPickDifferentKernelsDependingOnRequest) {
BuiltinDispatchInfoBuilder &builder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
HWTEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxTranslationDispatchThenPickDifferentKernelsDependingOnRequest) {
BuiltinDispatchInfoBuilder &baseBuilder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MemObjsForAuxTranslation memObjsForAuxTranslation;
MockBuffer mockBuffer[3];
@@ -358,10 +361,10 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxTranslationDispatchThenPickD
}
builtinOpsParams.auxTranslationDirection = AuxTranslationDirection::AuxToNonAux;
EXPECT_TRUE(builder.buildDispatchInfos(multiDispatchInfo, builtinOpsParams));
EXPECT_TRUE(builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams));
builtinOpsParams.auxTranslationDirection = AuxTranslationDirection::NonAuxToAux;
EXPECT_TRUE(builder.buildDispatchInfos(multiDispatchInfo, builtinOpsParams));
EXPECT_TRUE(builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams));
EXPECT_EQ(6u, multiDispatchInfo.size());
@@ -376,8 +379,9 @@ TEST_F(BuiltInTests, givenInputBufferWhenBuildingAuxTranslationDispatchThenPickD
EXPECT_NE(builtinKernels[2], builtinKernels[5]);
}
TEST_F(BuiltInTests, givenInvalidAuxTranslationDirectionWhenBuildingDispatchInfosThenAbort) {
BuiltinDispatchInfoBuilder &builder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
HWTEST_F(BuiltInTests, givenInvalidAuxTranslationDirectionWhenBuildingDispatchInfosThenAbort) {
BuiltinDispatchInfoBuilder &baseBuilder = pBuiltIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, *pContext, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MemObjsForAuxTranslation memObjsForAuxTranslation;
MockBuffer mockBuffer;
@@ -389,7 +393,7 @@ TEST_F(BuiltInTests, givenInvalidAuxTranslationDirectionWhenBuildingDispatchInfo
memObjsForAuxTranslation.insert(&mockBuffer);
builtinOpsParams.auxTranslationDirection = AuxTranslationDirection::None;
EXPECT_THROW(builder.buildDispatchInfos(multiDispatchInfo, builtinOpsParams), std::exception);
EXPECT_THROW(builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams), std::exception);
}
class MockAuxBuilInOp : public BuiltInOp<EBuiltInOps::AuxTranslation> {
@@ -411,7 +415,7 @@ TEST_F(BuiltInTests, whenAuxBuiltInIsConstructedThenResizeKernelInstancedTo5) {
EXPECT_EQ(5u, mockAuxBuiltInOp.convertToNonAuxKernel.size());
}
TEST_F(BuiltInTests, givenMoreBuffersForAuxTranslationThanKernelInstancesWhenDispatchingThenResize) {
HWTEST_F(BuiltInTests, givenMoreBuffersForAuxTranslationThanKernelInstancesWhenDispatchingThenResize) {
MockAuxBuilInOp mockAuxBuiltInOp(*pBuiltIns, *pContext, *pDevice);
EXPECT_EQ(5u, mockAuxBuiltInOp.convertToAuxKernel.size());
EXPECT_EQ(5u, mockAuxBuiltInOp.convertToNonAuxKernel.size());
@@ -428,7 +432,7 @@ TEST_F(BuiltInTests, givenMoreBuffersForAuxTranslationThanKernelInstancesWhenDis
memObjsForAuxTranslation.insert(&buffer);
}
EXPECT_TRUE(mockAuxBuiltInOp.buildDispatchInfos(multiDispatchInfo, builtinOpsParams));
EXPECT_TRUE(mockAuxBuiltInOp.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams));
EXPECT_EQ(7u, mockAuxBuiltInOp.convertToAuxKernel.size());
EXPECT_EQ(7u, mockAuxBuiltInOp.convertToNonAuxKernel.size());
}

86
unit_tests/command_queue/dispatch_walker_tests.cpp
View File

@@ -5,6 +5,7 @@
*
*/
#include "runtime/built_ins/aux_translation_builtin.h"
#include "runtime/command_queue/gpgpu_walker.h"
#include "runtime/command_queue/hardware_interface.h"
#include "runtime/event/perf_counter.h"
@@ -18,6 +19,7 @@
#include "unit_tests/fixtures/device_fixture.h"
#include "unit_tests/helpers/debug_manager_state_restore.h"
#include "unit_tests/helpers/hw_parse.h"
#include "unit_tests/mocks/mock_buffer.h"
#include "unit_tests/mocks/mock_command_queue.h"
#include "unit_tests/mocks/mock_graphics_allocation.h"
#include "unit_tests/mocks/mock_kernel.h"
@@ -1266,7 +1268,13 @@ TEST(DispatchWalker, calculateDispatchDim) {
}
}
HWTEST_F(DispatchWalkerTest, givenKernelWhenAuxTranslationRequiredThenPipeControlWithStallAndDCFlushAdded) {
HWTEST_F(DispatchWalkerTest, givenKernelWhenAuxToNonAuxWhenTranslationRequiredThenPipeControlWithStallAndDCFlushAdded) {
MockContext context;
auto executionEnvironment = pDevice->getExecutionEnvironment();
auto builtIns = executionEnvironment->getBuiltIns();
BuiltinDispatchInfoBuilder &baseBuilder = builtIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, context, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MockKernel kernel(program.get(), kernelInfo, *pDevice);
kernelInfo.workloadInfo.workDimOffset = 0;
ASSERT_EQ(CL_SUCCESS, kernel.initialize());
@@ -1274,11 +1282,18 @@ HWTEST_F(DispatchWalkerTest, givenKernelWhenAuxTranslationRequiredThenPipeContro
auto &cmdStream = pCmdQ->getCS(0);
void *buffer = cmdStream.getCpuBase();
kernel.auxTranslationRequired = true;
MockBuffer mockBuffer[2];
MockMultiDispatchInfo multiDispatchInfo(&kernel);
DispatchInfo di1(&kernel, 1, Vec3<size_t>(1, 1, 1), Vec3<size_t>(1, 1, 1), Vec3<size_t>(0, 0, 0));
di1.setPipeControlRequired(true);
multiDispatchInfo.push(di1);
MultiDispatchInfo multiDispatchInfo;
MemObjsForAuxTranslation memObjsForAuxTranslation;
memObjsForAuxTranslation.insert(&mockBuffer[0]);
memObjsForAuxTranslation.insert(&mockBuffer[1]);
BuiltinOpParams builtinOpsParams;
builtinOpsParams.memObjsForAuxTranslation = &memObjsForAuxTranslation;
builtinOpsParams.auxTranslationDirection = AuxTranslationDirection::AuxToNonAux;
builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams);
HardwareInterface<FamilyType>::dispatchWalker(
*pCmdQ,
@@ -1305,7 +1320,66 @@ HWTEST_F(DispatchWalkerTest, givenKernelWhenAuxTranslationRequiredThenPipeContro
EXPECT_TRUE(beginPipeControl->getCommandStreamerStallEnable());
auto endPipeControl = genCmdCast<typename FamilyType::PIPE_CONTROL *>(*(pipeControls[1]));
EXPECT_FALSE(endPipeControl->getDcFlushEnable());
bool dcFlushRequired = (executionEnvironment->getHardwareInfo()->platform.eRenderCoreFamily == IGFX_GEN8_CORE);
EXPECT_EQ(dcFlushRequired, endPipeControl->getDcFlushEnable());
EXPECT_TRUE(endPipeControl->getCommandStreamerStallEnable());
}
HWTEST_F(DispatchWalkerTest, givenKernelWhenNonAuxToAuxWhenTranslationRequiredThenPipeControlWithStallAdded) {
MockContext context;
auto executionEnvironment = pDevice->getExecutionEnvironment();
auto builtIns = executionEnvironment->getBuiltIns();
BuiltinDispatchInfoBuilder &baseBuilder = builtIns->getBuiltinDispatchInfoBuilder(EBuiltInOps::AuxTranslation, context, *pDevice);
auto &builder = static_cast<BuiltInOp<EBuiltInOps::AuxTranslation> &>(baseBuilder);
MockKernel kernel(program.get(), kernelInfo, *pDevice);
kernelInfo.workloadInfo.workDimOffset = 0;
ASSERT_EQ(CL_SUCCESS, kernel.initialize());
auto &cmdStream = pCmdQ->getCS(0);
void *buffer = cmdStream.getCpuBase();
kernel.auxTranslationRequired = true;
MockBuffer mockBuffer[2];
MultiDispatchInfo multiDispatchInfo;
MemObjsForAuxTranslation memObjsForAuxTranslation;
memObjsForAuxTranslation.insert(&mockBuffer[0]);
memObjsForAuxTranslation.insert(&mockBuffer[1]);
BuiltinOpParams builtinOpsParams;
builtinOpsParams.memObjsForAuxTranslation = &memObjsForAuxTranslation;
builtinOpsParams.auxTranslationDirection = AuxTranslationDirection::NonAuxToAux;
builder.buildDispatchInfosForAuxTranslation<FamilyType>(multiDispatchInfo, builtinOpsParams);
HardwareInterface<FamilyType>::dispatchWalker(
*pCmdQ,
multiDispatchInfo,
CsrDependencies(),
nullptr,
nullptr,
nullptr,
nullptr,
nullptr,
pDevice->getPreemptionMode(),
false);
auto sizeUsed = cmdStream.getUsed();
GenCmdList cmdList;
ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(cmdList, buffer, sizeUsed));
auto pipeControls = findAll<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
ASSERT_EQ(2u, pipeControls.size());
bool dcFlushRequired = (executionEnvironment->getHardwareInfo()->platform.eRenderCoreFamily == IGFX_GEN8_CORE);
auto beginPipeControl = genCmdCast<typename FamilyType::PIPE_CONTROL *>(*(pipeControls[0]));
EXPECT_EQ(dcFlushRequired, beginPipeControl->getDcFlushEnable());
EXPECT_TRUE(beginPipeControl->getCommandStreamerStallEnable());
auto endPipeControl = genCmdCast<typename FamilyType::PIPE_CONTROL *>(*(pipeControls[1]));
EXPECT_EQ(dcFlushRequired, endPipeControl->getDcFlushEnable());
EXPECT_TRUE(endPipeControl->getCommandStreamerStallEnable());
}

20
unit_tests/command_queue/enqueue_kernel_2_tests.cpp
View File

@@ -655,6 +655,7 @@ struct EnqueueAuxKernelTests : public EnqueueKernelTest {
template <typename FamilyType>
class MyCmdQ : public CommandQueueHw<FamilyType> {
public:
using CommandQueueHw<FamilyType>::commandStream;
MyCmdQ(Context *context, Device *device) : CommandQueueHw<FamilyType>(context, device, nullptr) {}
void dispatchAuxTranslation(MultiDispatchInfo &multiDispatchInfo, MemObjsForAuxTranslation &memObjsForAuxTranslation,
AuxTranslationDirection auxTranslationDirection) override {
@@ -733,14 +734,19 @@ HWTEST_F(EnqueueAuxKernelTests, givenMultipleArgsWhenAuxTranslationIsRequiredThe
EXPECT_EQ(&buffer2, *std::get<MemObjsForAuxTranslation>(cmdQ.dispatchAuxTranslationInputs.at(0)).begin());
EXPECT_EQ(&buffer2, *std::get<MemObjsForAuxTranslation>(cmdQ.dispatchAuxTranslationInputs.at(1)).begin());
uint32_t pipeControlCount = 0;
for (auto dispatchInfo : cmdQ.dispatchInfos) {
if (dispatchInfo.isPipeControlRequired()) {
++pipeControlCount;
}
}
EXPECT_EQ(4u, pipeControlCount);
auto cmdStream = cmdQ.commandStream;
auto sizeUsed = cmdStream->getUsed();
GenCmdList cmdList;
ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(cmdList, cmdStream->getCpuBase(), sizeUsed));
auto pipeControls = findAll<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
auto additionalPcCount = PipeControlHelper<FamilyType>::getSizeForPipeControlWithPostSyncOperation() / sizeof(typename FamilyType::PIPE_CONTROL);
// |AuxToNonAux|NDR|NonAuxToAux|
ASSERT_EQ(4u + additionalPcCount, pipeControls.size());
ASSERT_EQ(2u, cmdQ.auxTranslationDirections.size());
EXPECT_EQ(AuxTranslationDirection::AuxToNonAux, cmdQ.auxTranslationDirections[0]);
EXPECT_EQ(AuxTranslationDirection::NonAuxToAux, cmdQ.auxTranslationDirections[1]);