intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 09:09:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add missing cache flush for blocked path

Browse Source 
Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2021-12-07 08:40:35 +00:00
committed by Compute-Runtime-Automation
parent fca0d03c90
commit 0c49fa8203
5 changed files with 58 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
opencl/source/command_queue/command_queue.h
View File

@@ -350,6 +350,8 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
uint64_t dispatchHints = 0;
bool isTextureCacheFlushNeeded(uint32_t commandType) const;
protected:
void *enqueueReadMemObjForMap(TransferProperties &transferProperties, EventsRequest &eventsRequest, cl_int &errcodeRet);
cl_int enqueueWriteMemObjForUnmap(MemObj *memObj, void *mappedPtr, EventsRequest &eventsRequest);
@@ -370,7 +372,6 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
bool queueDependenciesClearRequired() const;
bool blitEnqueueAllowed(const CsrSelectionArgs &args) const;
bool isTextureCacheFlushNeeded(uint32_t commandType) const;
inline bool shouldFlushDC(uint32_t commandType, PrintfHandler *printfHandler) const {
return (commandType == CL_COMMAND_READ_BUFFER ||
commandType == CL_COMMAND_READ_BUFFER_RECT ||

2
opencl/source/command_queue/enqueue_common.h
View File

@@ -1032,7 +1032,7 @@ void CommandQueueHw<GfxFamily>::enqueueBlocked(
allSurfaces,
shouldFlushDC(commandType, printfHandler.get()),
slmUsed,
commandType == CL_COMMAND_NDRANGE_KERNEL,
commandType,
std::move(printfHandler),
preemptionMode,
multiDispatchInfo.peekMainKernel(),

8
opencl/source/helpers/task_information.cpp
View File

@@ -113,10 +113,10 @@ CompletionStamp &CommandMapUnmap::submit(uint32_t taskLevel, bool terminated) {
}
CommandComputeKernel::CommandComputeKernel(CommandQueue &commandQueue, std::unique_ptr<KernelOperation> &kernelOperation, std::vector<Surface *> &surfaces,
bool flushDC, bool usesSLM, bool ndRangeKernel, std::unique_ptr<PrintfHandler> &&printfHandler,
bool flushDC, bool usesSLM, uint32_t commandType, std::unique_ptr<PrintfHandler> &&printfHandler,
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount)
: Command(commandQueue, kernelOperation), flushDC(flushDC), slmUsed(usesSLM),
NDRangeKernel(ndRangeKernel), printfHandler(std::move(printfHandler)), kernel(kernel),
commandType(commandType), printfHandler(std::move(printfHandler)), kernel(kernel),
kernelCount(kernelCount), preemptionMode(preemptionMode) {
for (auto surface : surfaces) {
this->surfaces.push_back(surface);
@@ -241,7 +241,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
flushDC, //dcFlush
slmUsed, //useSLM
true, //guardCommandBufferWithPipeControl
NDRangeKernel, //GSBA32BitRequired
commandType == CL_COMMAND_NDRANGE_KERNEL, //GSBA32BitRequired
requiresCoherency, //requiresCoherency
commandQueue.getPriority() == QueuePriority::LOW, //lowPriority
false, //implicitFlush
@@ -252,7 +252,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
kernel->getKernelInfo().kernelDescriptor.kernelAttributes.flags.useGlobalAtomics, //useGlobalAtomics
kernel->areMultipleSubDevicesInContext(), //areMultipleSubDevicesInContext
kernel->requiresMemoryMigration(), //memoryMigrationRequired
false); //textureCacheFlush
commandQueue.isTextureCacheFlushNeeded(this->commandType)); //textureCacheFlush
if (commandQueue.getContext().getRootDeviceIndices().size() > 1) {
eventsRequest.fillCsrDependenciesForTaskCountContainer(dispatchFlags.csrDependencies, commandStreamReceiver);

4
opencl/source/helpers/task_information.h
View File

@@ -130,7 +130,7 @@ class CommandMapUnmap : public Command {
class CommandComputeKernel : public Command {
public:
CommandComputeKernel(CommandQueue &commandQueue, std::unique_ptr<KernelOperation> &kernelOperation, std::vector<Surface *> &surfaces,
bool flushDC, bool usesSLM, bool ndRangeKernel, std::unique_ptr<PrintfHandler> &&printfHandler,
bool flushDC, bool usesSLM, uint32_t commandType, std::unique_ptr<PrintfHandler> &&printfHandler,
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount);
~CommandComputeKernel() override;
@@ -145,7 +145,7 @@ class CommandComputeKernel : public Command {
std::vector<Surface *> surfaces;
bool flushDC;
bool slmUsed;
bool NDRangeKernel;
uint32_t commandType;
std::unique_ptr<PrintfHandler> printfHandler;
Kernel *kernel;
uint32_t kernelCount;