Pass more information to programPipelineSelect.

Change-Id: Iaabe60742269b721f8defe71306dd6e87d60d546

Author: PiotrFusik

runtime/command_queue/enqueue_common.h

@@ -526,6 +526,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
 
     auto mediaSamplerRequired = false;
     uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
+    auto specialPipelineSelectMode = false;
     Kernel *kernel = nullptr;
     for (auto &dispatchInfo : multiDispatchInfo) {
         if (kernel != dispatchInfo.getKernel()) {
@@ -538,6 +539,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
         mediaSamplerRequired |= kernel->isVmeKernel();
         auto numGrfRequiredByKernel = kernel->getKernelInfo().patchInfo.executionEnvironment->NumGRFRequired;
         numGrfRequired = std::max(numGrfRequired, numGrfRequiredByKernel);
+        specialPipelineSelectMode |= kernel->requiresSpecialPipelineSelectMode();
     }
 
     if (mediaSamplerRequired) {
@@ -595,6 +597,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
         dispatchFlags.outOfDeviceDependencies = &eventsRequest;
     }
     dispatchFlags.numGrfRequired = numGrfRequired;
+    dispatchFlags.specialPipelineSelectMode = specialPipelineSelectMode;
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 
     if (gtpinIsGTPinInitialized()) {

runtime/command_stream/command_stream_receiver.h

@@ -218,6 +218,7 @@ class CommandStreamReceiver {
     bool stallingPipeControlOnNextFlushRequired = false;
     bool timestampPacketWriteEnabled = false;
     bool nTo1SubmissionModelEnabled = false;
+    bool lastSpecialPipelineSelectMode = false;
 };
 
 typedef CommandStreamReceiver *(*CommandStreamReceiverCreateFunc)(const HardwareInfo &hwInfoIn, bool withAubDump, ExecutionEnvironment &executionEnvironment);

runtime/command_stream/command_stream_receiver_hw.inl

@@ -121,7 +121,7 @@ inline typename GfxFamily::PIPE_CONTROL *CommandStreamReceiverHw<GfxFamily>::add
 template <typename GfxFamily>
 void CommandStreamReceiverHw<GfxFamily>::programPipelineSelect(LinearStream &commandStream, DispatchFlags &dispatchFlags) {
     if (csrSizeRequestFlags.mediaSamplerConfigChanged || !isPreambleSent) {
-        PreambleHelper<GfxFamily>::programPipelineSelect(&commandStream, dispatchFlags.mediaSamplerRequired);
+        PreambleHelper<GfxFamily>::programPipelineSelect(&commandStream, dispatchFlags);
         this->lastMediaSamplerConfig = dispatchFlags.mediaSamplerRequired;
     }
 }
@@ -239,6 +239,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     csrSizeRequestFlags.preemptionRequestChanged = this->lastPreemptionMode != dispatchFlags.preemptionMode;
     csrSizeRequestFlags.mediaSamplerConfigChanged = this->lastMediaSamplerConfig != static_cast<int8_t>(dispatchFlags.mediaSamplerRequired);
     csrSizeRequestFlags.numGrfRequiredChanged = this->lastSentNumGrfRequired != dispatchFlags.numGrfRequired;
+    csrSizeRequestFlags.specialPipelineSelectModeChanged = this->lastSpecialPipelineSelectMode != dispatchFlags.specialPipelineSelectMode;
 
     size_t requiredScratchSizeInBytes = requiredScratchSize * device.getDeviceInfo().computeUnitsUsedForScratch;
 

runtime/command_stream/csr_definitions.h

@@ -45,6 +45,7 @@ struct DispatchFlags {
     PreemptionMode preemptionMode = PreemptionMode::Disabled;
     EventsRequest *outOfDeviceDependencies = nullptr;
     uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
+    bool specialPipelineSelectMode = false;
 };
 
 struct CsrSizeRequestFlags {
@@ -54,5 +55,6 @@ struct CsrSizeRequestFlags {
     bool mediaSamplerConfigChanged = false;
     bool hasSharedHandles = false;
     bool numGrfRequiredChanged = false;
+    bool specialPipelineSelectModeChanged = false;
 };
 } // namespace OCLRT

runtime/gen10/preamble_gen10.cpp

@@ -5,8 +5,9 @@
  *
  */
 
-#include "runtime/helpers/preamble.inl"
 #include "runtime/command_queue/gpgpu_walker.h"
+#include "runtime/command_stream/csr_definitions.h"
+#include "runtime/helpers/preamble.inl"
 
 namespace OCLRT {
 
@@ -67,7 +68,7 @@ void PreambleHelper<CNLFamily>::addPipeControlBeforeVfeCmd(LinearStream *pComman
 }
 
 template <>
-void PreambleHelper<CNLFamily>::programPipelineSelect(LinearStream *pCommandStream, bool mediaSamplerRequired) {
+void PreambleHelper<CNLFamily>::programPipelineSelect(LinearStream *pCommandStream, const DispatchFlags &dispatchFlags) {
     typedef typename CNLFamily::PIPELINE_SELECT PIPELINE_SELECT;
 
     auto pCmd = (PIPELINE_SELECT *)pCommandStream->getSpace(sizeof(PIPELINE_SELECT));
@@ -76,7 +77,7 @@ void PreambleHelper<CNLFamily>::programPipelineSelect(LinearStream *pCommandStre
     auto mask = pipelineSelectEnablePipelineSelectMaskBits | pipelineSelectMediaSamplerDopClockGateMaskBits;
     pCmd->setMaskBits(mask);
     pCmd->setPipelineSelection(PIPELINE_SELECT::PIPELINE_SELECTION_GPGPU);
-    pCmd->setMediaSamplerDopClockGateEnable(!mediaSamplerRequired);
+    pCmd->setMediaSamplerDopClockGateEnable(!dispatchFlags.mediaSamplerRequired);
 }
 
 template struct PreambleHelper<CNLFamily>;

runtime/gen8/preamble_gen8.cpp

@@ -32,7 +32,7 @@ uint32_t PreambleHelper<BDWFamily>::getL3Config(const HardwareInfo &hwInfo, bool
 }
 
 template <>
-void PreambleHelper<BDWFamily>::programPipelineSelect(LinearStream *pCommandStream, bool mediaSamplerRequired) {
+void PreambleHelper<BDWFamily>::programPipelineSelect(LinearStream *pCommandStream, const DispatchFlags &dispatchFlags) {
     typedef typename BDWFamily::PIPELINE_SELECT PIPELINE_SELECT;
     auto pCmd = (PIPELINE_SELECT *)pCommandStream->getSpace(sizeof(PIPELINE_SELECT));
     *pCmd = PIPELINE_SELECT::sInit();

runtime/gen9/preamble_gen9.cpp

@@ -5,6 +5,7 @@
  *
  */
 
+#include "runtime/command_stream/csr_definitions.h"
 #include "runtime/helpers/preamble.inl"
 
 namespace OCLRT {
@@ -27,7 +28,7 @@ uint32_t PreambleHelper<SKLFamily>::getL3Config(const HardwareInfo &hwInfo, bool
 }
 
 template <>
-void PreambleHelper<SKLFamily>::programPipelineSelect(LinearStream *pCommandStream, bool mediaSamplerRequired) {
+void PreambleHelper<SKLFamily>::programPipelineSelect(LinearStream *pCommandStream, const DispatchFlags &dispatchFlags) {
     typedef typename SKLFamily::PIPELINE_SELECT PIPELINE_SELECT;
 
     auto pCmd = (PIPELINE_SELECT *)pCommandStream->getSpace(sizeof(PIPELINE_SELECT));
@@ -36,7 +37,7 @@ void PreambleHelper<SKLFamily>::programPipelineSelect(LinearStream *pCommandStre
     auto mask = pipelineSelectEnablePipelineSelectMaskBits | pipelineSelectMediaSamplerDopClockGateMaskBits;
     pCmd->setMaskBits(mask);
     pCmd->setPipelineSelection(PIPELINE_SELECT::PIPELINE_SELECTION_GPGPU);
-    pCmd->setMediaSamplerDopClockGateEnable(!mediaSamplerRequired);
+    pCmd->setMediaSamplerDopClockGateEnable(!dispatchFlags.mediaSamplerRequired);
 }
 
 template <>

runtime/helpers/pipeline_select_helper.h

@@ -6,7 +6,7 @@
  */
 
 #pragma once
-#include "stdint.h"
+#include <cstdint>
 
 namespace OCLRT {
 const uint32_t pipelineSelectEnablePipelineSelectMaskBits = 0x3;

runtime/helpers/preamble.h

@@ -7,14 +7,15 @@
 
 #pragma once
 #include "igfxfmid.h"
-#include "stdint.h"
 #include "runtime/helpers/pipeline_select_helper.h"
 #include <cstddef>
+#include <cstdint>
 
 namespace OCLRT {
 
 struct HardwareInfo;
 class Device;
+struct DispatchFlags;
 class GraphicsAllocation;
 class LinearStream;
 
@@ -26,7 +27,7 @@ struct PreambleHelper {
     static constexpr size_t getScratchSpaceOffsetFor64bit() { return 4096; }
 
     static void programL3(LinearStream *pCommandStream, uint32_t l3Config);
-    static void programPipelineSelect(LinearStream *pCommandStream, bool mediaSamplerRequired);
+    static void programPipelineSelect(LinearStream *pCommandStream, const DispatchFlags &dispatchFlags);
     static uint32_t getDefaultThreadArbitrationPolicy();
     static void programThreadArbitration(LinearStream *pCommandStream, uint32_t requiredThreadArbitrationPolicy);
     static void programPreemption(LinearStream *pCommandStream, Device &device, GraphicsAllocation *preemptionCsr);

runtime/helpers/task_information.cpp

@@ -206,6 +206,7 @@ CompletionStamp &CommandComputeKernel::submit(uint32_t taskLevel, bool terminate
     if (commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
         dispatchFlags.outOfDeviceDependencies = &eventsRequest;
     }
+    dispatchFlags.specialPipelineSelectMode = kernel->requiresSpecialPipelineSelectMode();
 
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 

runtime/kernel/kernel.h

@@ -276,15 +276,16 @@ class Kernel : public BaseObject<_cl_kernel> {
     const void *getKernelArg(uint32_t argIndex) const;
     const SimpleKernelArgInfo &getKernelArgInfo(uint32_t argIndex) const;
 
-    bool getAllowNonUniform() { return program->getAllowNonUniform(); }
-    bool isVmeKernel() { return kernelInfo.isVmeWorkload; };
+    bool getAllowNonUniform() const { return program->getAllowNonUniform(); }
+    bool isVmeKernel() const { return kernelInfo.isVmeWorkload; }
+    bool requiresSpecialPipelineSelectMode() const { return specialPipelineSelectMode; }
 
     //residency for kernel surfaces
     MOCKABLE_VIRTUAL void makeResident(CommandStreamReceiver &commandStreamReceiver);
     MOCKABLE_VIRTUAL void getResidency(std::vector<Surface *> &dst);
     bool requiresCoherency();
     void resetSharedObjectsPatchAddresses();
-    bool isUsingSharedObjArgs() { return usingSharedObjArgs; }
+    bool isUsingSharedObjArgs() const { return usingSharedObjArgs; }
 
     bool hasPrintfOutput() const;
 
@@ -490,5 +491,7 @@ class Kernel : public BaseObject<_cl_kernel> {
 
     std::vector<PatchInfoData> patchInfoDataList;
     std::unique_ptr<ImageTransformer> imageTransformer;
+
+    bool specialPipelineSelectMode = false;
 };
 } // namespace OCLRT

unit_tests/libult/mock_gfx_family.cpp

@@ -158,7 +158,7 @@ uint32_t PreambleHelper<GENX>::getL3Config(const HardwareInfo &hwInfo, bool useS
 }
 
 template <>
-void PreambleHelper<GENX>::programPipelineSelect(LinearStream *pCommandStream, bool mediaSamplerRequired) {
+void PreambleHelper<GENX>::programPipelineSelect(LinearStream *pCommandStream, const DispatchFlags &dispatchFlags) {
 }
 
 template <>

unit_tests/mocks/mock_kernel.h

@@ -225,6 +225,8 @@ class MockKernel : public Kernel {
         Kernel::releaseOwnership();
     }
 
+    void setSpecialPipelineSelectMode(bool value) { specialPipelineSelectMode = value; }
+
     uint32_t makeResidentCalls = 0;
     uint32_t getResidencyCalls = 0;
     mutable uint32_t takeOwnershipCalls = 0;