intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-08 05:56:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Pass command stream to dispatch scheduler

Browse Source 
instead of taking it from CommandQueue

Change-Id: I8e43c3b7ed5cb46f79edf3290a84fc6ad41f3b57
Signed-off-by: Maciej Dziuban <maciej.dziuban@intel.com>
This commit is contained in:
Maciej Dziuban
2019-02-25 09:42:15 +01:00
committed by sys_ocldev
parent 10083d5a21
commit 0cf71414e2
12 changed files with 33 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
runtime/command_queue/enqueue_common.h
View File

@@ -454,6 +454,7 @@ void CommandQueueHw<GfxFamily>::processDeviceEnqueue(Kernel *parentKernel,
GpgpuWalkerHelper<GfxFamily>::dispatchScheduler( GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
*this, *this,
*this->commandStream,
*devQueueHw, *devQueueHw,
preemption, preemption,
scheduler, scheduler,

1
runtime/command_queue/gpgpu_walker.h
View File

@@ -196,6 +196,7 @@ class GpgpuWalkerHelper {
static void dispatchScheduler( static void dispatchScheduler(
CommandQueue &commandQueue, CommandQueue &commandQueue,
LinearStream &commandStream,
DeviceQueueHw<GfxFamily> &devQueueHw, DeviceQueueHw<GfxFamily> &devQueueHw,
PreemptionMode preemptionMode, PreemptionMode preemptionMode,
SchedulerKernel &scheduler, SchedulerKernel &scheduler,

24
runtime/command_queue/gpgpu_walker_base.inl
View File

@@ -53,6 +53,7 @@ inline size_t GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(
template <typename GfxFamily> template <typename GfxFamily>
void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler( void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
CommandQueue &commandQueue, CommandQueue &commandQueue,
LinearStream &commandStream,
DeviceQueueHw<GfxFamily> &devQueueHw, DeviceQueueHw<GfxFamily> &devQueueHw,
PreemptionMode preemptionMode, PreemptionMode preemptionMode,
SchedulerKernel &scheduler, SchedulerKernel &scheduler,
@@ -63,13 +64,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER; using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START; using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
OCLRT::LinearStream *commandStream = nullptr;
OCLRT::IndirectHeap *ioh = nullptr;
commandStream = &commandQueue.getCS(0);
bool dcFlush = false; bool dcFlush = false;
commandQueue.getCommandStreamReceiver().addPipeControl(*commandStream, dcFlush); commandQueue.getCommandStreamReceiver().addPipeControl(commandStream, dcFlush);
uint32_t interfaceDescriptorIndex = devQueueHw.schedulerIDIndex; uint32_t interfaceDescriptorIndex = devQueueHw.schedulerIDIndex;
const size_t offsetInterfaceDescriptorTable = devQueueHw.colorCalcStateSize; const size_t offsetInterfaceDescriptorTable = devQueueHw.colorCalcStateSize;
@@ -78,7 +74,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
// Program media interface descriptor load // Program media interface descriptor load
KernelCommandsHelper<GfxFamily>::sendMediaInterfaceDescriptorLoad( KernelCommandsHelper<GfxFamily>::sendMediaInterfaceDescriptorLoad(
*commandStream, commandStream,
offsetInterfaceDescriptor, offsetInterfaceDescriptor,
totalInterfaceDescriptorTableSize); totalInterfaceDescriptorTableSize);
@@ -123,16 +119,16 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
size_t curbeOffset = devQueueHw.setSchedulerCrossThreadData(scheduler); size_t curbeOffset = devQueueHw.setSchedulerCrossThreadData(scheduler);
IndirectHeap indirectObjectHeap(dsh->getCpuBase(), dsh->getMaxAvailableSpace()); IndirectHeap indirectObjectHeap(dsh->getCpuBase(), dsh->getMaxAvailableSpace());
indirectObjectHeap.getSpace(curbeOffset); indirectObjectHeap.getSpace(curbeOffset);
ioh = &indirectObjectHeap; IndirectHeap *ioh = &indirectObjectHeap;
// Program the walker. Invokes execution so all state should already be programmed // Program the walker. Invokes execution so all state should already be programmed
auto pGpGpuWalkerCmd = (GPGPU_WALKER *)commandStream->getSpace(sizeof(GPGPU_WALKER)); auto pGpGpuWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
*pGpGpuWalkerCmd = GfxFamily::cmdInitGpgpuWalker; *pGpGpuWalkerCmd = GfxFamily::cmdInitGpgpuWalker;
bool localIdsGenerationByRuntime = KernelCommandsHelper<GfxFamily>::isRuntimeLocalIdsGenerationRequired(1, globalWorkSizes, localWorkSizes); bool localIdsGenerationByRuntime = KernelCommandsHelper<GfxFamily>::isRuntimeLocalIdsGenerationRequired(1, globalWorkSizes, localWorkSizes);
bool inlineDataProgrammingRequired = KernelCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(scheduler); bool inlineDataProgrammingRequired = KernelCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(scheduler);
KernelCommandsHelper<GfxFamily>::sendIndirectState( KernelCommandsHelper<GfxFamily>::sendIndirectState(
*commandStream, commandStream,
*dsh, *dsh,
*ioh, *ioh,
*ssh, *ssh,
@@ -147,7 +143,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
localIdsGenerationByRuntime); localIdsGenerationByRuntime);
// Implement enabling special WA DisableLSQCROPERFforOCL if needed // Implement enabling special WA DisableLSQCROPERFforOCL if needed
GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(commandStream, scheduler, true); GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, true);
size_t globalOffsets[3] = {0, 0, 0}; size_t globalOffsets[3] = {0, 0, 0};
size_t workGroups[3] = {(scheduler.getGws() / scheduler.getLws()), 1, 1}; size_t workGroups[3] = {(scheduler.getGws() / scheduler.getLws()), 1, 1};
@@ -156,15 +152,15 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
*scheduler.getKernelInfo().patchInfo.threadPayload); *scheduler.getKernelInfo().patchInfo.threadPayload);
// Implement disabling special WA DisableLSQCROPERFforOCL if needed // Implement disabling special WA DisableLSQCROPERFforOCL if needed
GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(commandStream, scheduler, false); GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, false);
// Do not put BB_START only when returning in first Scheduler run // Do not put BB_START only when returning in first Scheduler run
if (devQueueHw.getSchedulerReturnInstance() != 1) { if (devQueueHw.getSchedulerReturnInstance() != 1) {
commandQueue.getCommandStreamReceiver().addPipeControl(*commandStream, true); commandQueue.getCommandStreamReceiver().addPipeControl(commandStream, true);
// Add BB Start Cmd to the SLB in the Primary Batch Buffer // Add BB Start Cmd to the SLB in the Primary Batch Buffer
auto *bbStart = (MI_BATCH_BUFFER_START *)commandStream->getSpace(sizeof(MI_BATCH_BUFFER_START)); auto *bbStart = static_cast<MI_BATCH_BUFFER_START *>(commandStream.getSpace(sizeof(MI_BATCH_BUFFER_START)));
*bbStart = GfxFamily::cmdInitBatchBufferStart; *bbStart = GfxFamily::cmdInitBatchBufferStart;
bbStart->setSecondLevelBatchBuffer(MI_BATCH_BUFFER_START::SECOND_LEVEL_BATCH_BUFFER_FIRST_LEVEL_BATCH); bbStart->setSecondLevelBatchBuffer(MI_BATCH_BUFFER_START::SECOND_LEVEL_BATCH_BUFFER_FIRST_LEVEL_BATCH);
uint64_t slbAddress = devQueueHw.getSlbBuffer()->getGpuAddress(); uint64_t slbAddress = devQueueHw.getSlbBuffer()->getGpuAddress();

4
runtime/device_queue/device_queue.cpp
View File

@@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2017-2018 Intel Corporation * Copyright (C) 2017-2019 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@@ -160,7 +160,7 @@ void DeviceQueue::resetDeviceQueue() {
return; return;
} }
void DeviceQueue::dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) { void DeviceQueue::dispatchScheduler(CommandQueue &cmdQ, LinearStream &commandStream, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) {
return; return;
} }

4
runtime/device_queue/device_queue.h
View File

@@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2017-2018 Intel Corporation * Copyright (C) 2017-2019 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@@ -80,7 +80,7 @@ class DeviceQueue : public BaseObject<_device_queue> {
} }
virtual void resetDeviceQueue(); virtual void resetDeviceQueue();
virtual void dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh); virtual void dispatchScheduler(CommandQueue &cmdQ, LinearStream &commandStream, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh);
virtual IndirectHeap *getIndirectHeap(IndirectHeap::Type type); virtual IndirectHeap *getIndirectHeap(IndirectHeap::Type type);
void acquireEMCriticalSection() { void acquireEMCriticalSection() {

5
runtime/device_queue/device_queue_hw.h
View File

@@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2017-2018 Intel Corporation * Copyright (C) 2017-2019 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@@ -16,6 +16,7 @@
#include "runtime/scheduler/scheduler_kernel.h" #include "runtime/scheduler/scheduler_kernel.h"
namespace OCLRT { namespace OCLRT {
template <typename GfxFamily> template <typename GfxFamily>
class DeviceQueueHw : public DeviceQueue { class DeviceQueueHw : public DeviceQueue {
using BaseClass = DeviceQueue; using BaseClass = DeviceQueue;
@@ -57,7 +58,7 @@ class DeviceQueueHw : public DeviceQueue {
void addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) override; void addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) override;
void resetDeviceQueue() override; void resetDeviceQueue() override;
void dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override; void dispatchScheduler(CommandQueue &cmdQ, LinearStream &commandStream, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override;
uint32_t getSchedulerReturnInstance() { uint32_t getSchedulerReturnInstance() {
return igilQueue->m_controls.m_SchedulerEarlyReturn; return igilQueue->m_controls.m_SchedulerEarlyReturn;

3
runtime/device_queue/device_queue_hw.inl
View File

@@ -368,8 +368,9 @@ size_t DeviceQueueHw<GfxFamily>::setSchedulerCrossThreadData(SchedulerKernel &sc
} }
template <typename GfxFamily> template <typename GfxFamily>
void DeviceQueueHw<GfxFamily>::dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) { void DeviceQueueHw<GfxFamily>::dispatchScheduler(CommandQueue &cmdQ, LinearStream &commandStream, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) {
GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(cmdQ, GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(cmdQ,
commandStream,
*this, *this,
preemptionMode, preemptionMode,
scheduler, scheduler,

1
runtime/helpers/task_information.cpp
View File

@@ -182,6 +182,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
devQueue->dispatchScheduler( devQueue->dispatchScheduler(
commandQueue, commandQueue,
*kernelOperation->commandStream,
scheduler, scheduler,
preemptionMode, preemptionMode,
ssh, ssh,

6
unit_tests/device_queue/device_queue_tests.cpp
View File

@@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2017-2018 Intel Corporation * Copyright (C) 2017-2019 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@@ -305,7 +305,9 @@ TEST_F(DeviceQueueTest, dispatchScheduler) {
CommandQueue cmdQ(nullptr, nullptr, 0); CommandQueue cmdQ(nullptr, nullptr, 0);
KernelInfo info; KernelInfo info;
MockSchedulerKernel *kernel = new MockSchedulerKernel(&program, info, *device); MockSchedulerKernel *kernel = new MockSchedulerKernel(&program, info, *device);
devQueue.dispatchScheduler(cmdQ, *kernel, device->getPreemptionMode(), nullptr, nullptr); LinearStream cmdStream;
devQueue.dispatchScheduler(cmdQ, cmdStream, *kernel, device->getPreemptionMode(), nullptr, nullptr);
delete kernel; delete kernel;
} }

3
unit_tests/execution_model/scheduler_dispatch_tests.cpp
View File

@@ -60,6 +60,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ExecutionModelSchedulerFixture, dispatchScheduler) {
GpgpuWalkerHelper<FamilyType>::dispatchScheduler( GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
*pCmdQ, *pCmdQ,
pCmdQ->getCS(0),
*pDevQueueHw, *pDevQueueHw,
pDevice->getPreemptionMode(), pDevice->getPreemptionMode(),
scheduler, scheduler,
@@ -178,6 +179,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ExecutionModelSchedulerFixture, dispatchSchedulerDoe
GpgpuWalkerHelper<FamilyType>::dispatchScheduler( GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
*pCmdQ, *pCmdQ,
pCmdQ->getCS(0),
*pDevQueueHw, *pDevQueueHw,
pDevice->getPreemptionMode(), pDevice->getPreemptionMode(),
scheduler, scheduler,
@@ -211,6 +213,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ParentKernelCommandQueueFixture, dispatchSchedulerWi
GpgpuWalkerHelper<FamilyType>::dispatchScheduler( GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
*pCmdQ, *pCmdQ,
pCmdQ->getCS(0),
mockDevQueue, mockDevQueue,
device->getPreemptionMode(), device->getPreemptionMode(),
scheduler, scheduler,

4
unit_tests/execution_model/submit_blocked_parent_kernel_tests.cpp
View File

@@ -62,9 +62,9 @@ class MockDeviceQueueHwWithCriticalSectionRelease : public DeviceQueueHw<GfxFami
timestampAddedInCleanupSection = hwTimeStamp ? hwTimeStamp->tagForCpuAccess : nullptr; timestampAddedInCleanupSection = hwTimeStamp ? hwTimeStamp->tagForCpuAccess : nullptr;
return BaseClass::addExecutionModelCleanUpSection(parentKernel, hwTimeStamp, taskCount); return BaseClass::addExecutionModelCleanUpSection(parentKernel, hwTimeStamp, taskCount);
} }
void dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override { void dispatchScheduler(CommandQueue &cmdQ, LinearStream &commandStream, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override {
schedulerDispatched = true; schedulerDispatched = true;
return BaseClass::dispatchScheduler(cmdQ, scheduler, preemptionMode, ssh, dsh); return BaseClass::dispatchScheduler(cmdQ, commandStream, scheduler, preemptionMode, ssh, dsh);
} }
uint32_t criticalSectioncheckCounter = 0; uint32_t criticalSectioncheckCounter = 0;

3
unit_tests/gen8/scheduler_dispatch_tests_gen8.cpp
View File

@@ -1,5 +1,5 @@
/* /*
* Copyright (C) 2017-2018 Intel Corporation * Copyright (C) 2017-2019 Intel Corporation
* *
* SPDX-License-Identifier: MIT * SPDX-License-Identifier: MIT
* *
@@ -38,6 +38,7 @@ BDWTEST_F(BdwSchedulerTest, givenCallToDispatchSchedulerWhenPipeControlWithCSSta
GpgpuWalkerHelper<FamilyType>::dispatchScheduler( GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
*pCmdQ, *pCmdQ,
commandStream,
*pDevQueueHw, *pDevQueueHw,
pDevice->getPreemptionMode(), pDevice->getPreemptionMode(),
scheduler, scheduler,