intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-28 16:48:45 +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(
*this,
*this->commandStream,
*devQueueHw,
preemption,
scheduler,

1
runtime/command_queue/gpgpu_walker.h
View File

@@ -196,6 +196,7 @@ class GpgpuWalkerHelper {
static void dispatchScheduler(
CommandQueue &commandQueue,
LinearStream &commandStream,
DeviceQueueHw<GfxFamily> &devQueueHw,
PreemptionMode preemptionMode,
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>
void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
CommandQueue &commandQueue,
LinearStream &commandStream,
DeviceQueueHw<GfxFamily> &devQueueHw,
PreemptionMode preemptionMode,
SchedulerKernel &scheduler,
@@ -63,13 +64,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
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;
commandQueue.getCommandStreamReceiver().addPipeControl(*commandStream, dcFlush);
commandQueue.getCommandStreamReceiver().addPipeControl(commandStream, dcFlush);
uint32_t interfaceDescriptorIndex = devQueueHw.schedulerIDIndex;
const size_t offsetInterfaceDescriptorTable = devQueueHw.colorCalcStateSize;
@@ -78,7 +74,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
// Program media interface descriptor load
KernelCommandsHelper<GfxFamily>::sendMediaInterfaceDescriptorLoad(
*commandStream,
commandStream,
offsetInterfaceDescriptor,
totalInterfaceDescriptorTableSize);
@@ -123,16 +119,16 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
size_t curbeOffset = devQueueHw.setSchedulerCrossThreadData(scheduler);
IndirectHeap indirectObjectHeap(dsh->getCpuBase(), dsh->getMaxAvailableSpace());
indirectObjectHeap.getSpace(curbeOffset);
ioh = &indirectObjectHeap;
IndirectHeap *ioh = &indirectObjectHeap;
// 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;
bool localIdsGenerationByRuntime = KernelCommandsHelper<GfxFamily>::isRuntimeLocalIdsGenerationRequired(1, globalWorkSizes, localWorkSizes);
bool inlineDataProgrammingRequired = KernelCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(scheduler);
KernelCommandsHelper<GfxFamily>::sendIndirectState(
*commandStream,
commandStream,
*dsh,
*ioh,
*ssh,
@@ -147,7 +143,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
localIdsGenerationByRuntime);
// 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 workGroups[3] = {(scheduler.getGws() / scheduler.getLws()), 1, 1};
@@ -156,15 +152,15 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
*scheduler.getKernelInfo().patchInfo.threadPayload);
// 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
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
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->setSecondLevelBatchBuffer(MI_BATCH_BUFFER_START::SECOND_LEVEL_BATCH_BUFFER_FIRST_LEVEL_BATCH);
uint64_t slbAddress = devQueueHw.getSlbBuffer()->getGpuAddress();