Grf configuration

Change-Id: I3741f53a38c6707b0c8ad82ae553ea65ae6917e4
Signed-off-by: Pawel Wilma <pawel.wilma@intel.com>

runtime/command_queue/enqueue_common.h

@@ -534,6 +534,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     }
 
     auto mediaSamplerRequired = false;
+    uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
     Kernel *kernel = nullptr;
     for (auto &dispatchInfo : multiDispatchInfo) {
         if (kernel != dispatchInfo.getKernel()) {
@@ -544,6 +545,8 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
         kernel->makeResident(commandStreamReceiver);
         requiresCoherency |= kernel->requiresCoherency();
         mediaSamplerRequired |= kernel->isVmeKernel();
+        auto numGrfRequiredByKernel = kernel->getKernelInfo().patchInfo.executionEnvironment->NumGRFRequired;
+        numGrfRequired = std::max(numGrfRequired, numGrfRequiredByKernel);
     }
 
     if (mediaSamplerRequired) {
@@ -593,7 +596,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     if (commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
         dispatchFlags.outOfDeviceDependencies = &eventsRequest;
     }
-
+    dispatchFlags.numGrfRequired = numGrfRequired;
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 
     if (gtpinIsGTPinInitialized()) {

runtime/command_stream/command_stream_receiver.h

@@ -15,6 +15,7 @@
 #include "runtime/helpers/completion_stamp.h"
 #include "runtime/helpers/flat_batch_buffer_helper.h"
 #include "runtime/helpers/options.h"
+#include "runtime/kernel/grf_config.h"
 #include "runtime/indirect_heap/indirect_heap.h"
 #include <cstddef>
 #include <cstdint>
@@ -179,6 +180,7 @@ class CommandStreamReceiver {
     int8_t lastMediaSamplerConfig = -1;
     PreemptionMode lastPreemptionMode = PreemptionMode::Initial;
     uint32_t latestSentStatelessMocsConfig = 0;
+    uint32_t lastSentNumGrfRequired = GrfConfig::DefaultGrfNumber;
 
     LinearStream commandStream;
 

runtime/command_stream/command_stream_receiver_hw.h

@@ -50,9 +50,9 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     size_t getCmdSizeForPreemption(const DispatchFlags &dispatchFlags) const;
     size_t getCmdSizeForL3Config() const;
     size_t getCmdSizeForPipelineSelect() const;
-    size_t getCmdSizeForCoherency();
+    size_t getCmdSizeForComputeMode();
     size_t getCmdSizeForMediaSampler(bool mediaSamplerRequired) const;
-    void programCoherency(LinearStream &csr, DispatchFlags &dispatchFlags);
+    void programComputeMode(LinearStream &csr, DispatchFlags &dispatchFlags);
 
     void waitForTaskCountWithKmdNotifyFallback(uint32_t taskCountToWait, FlushStamp flushStampToWait, bool useQuickKmdSleep, OsContext &osContext) override;
     const HardwareInfo &peekHwInfo() const { return hwInfo; }

runtime/command_stream/command_stream_receiver_hw.inl

@@ -228,6 +228,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     csrSizeRequestFlags.coherencyRequestChanged = this->lastSentCoherencyRequest != static_cast<int8_t>(dispatchFlags.requiresCoherency);
     csrSizeRequestFlags.preemptionRequestChanged = this->lastPreemptionMode != dispatchFlags.preemptionMode;
     csrSizeRequestFlags.mediaSamplerConfigChanged = this->lastMediaSamplerConfig != static_cast<int8_t>(dispatchFlags.mediaSamplerRequired);
+    csrSizeRequestFlags.numGrfRequiredChanged = this->lastSentNumGrfRequired != dispatchFlags.numGrfRequired;
 
     size_t requiredScratchSizeInBytes = requiredScratchSize * device.getDeviceInfo().computeUnitsUsedForScratch;
 
@@ -255,7 +256,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     }
     initPageTableManagerRegisters(commandStreamCSR);
     programPreemption(commandStreamCSR, device, dispatchFlags);
-    programCoherency(commandStreamCSR, dispatchFlags);
+    programComputeMode(commandStreamCSR, dispatchFlags);
     programL3(commandStreamCSR, dispatchFlags, newL3Config);
     programPipelineSelect(commandStreamCSR, dispatchFlags);
     programPreamble(commandStreamCSR, device, dispatchFlags, newL3Config);
@@ -628,7 +629,7 @@ size_t CommandStreamReceiverHw<GfxFamily>::getRequiredCmdStreamSize(const Dispat
     size += sizeof(typename GfxFamily::MI_BATCH_BUFFER_START);
 
     size += getCmdSizeForL3Config();
-    size += getCmdSizeForCoherency();
+    size += getCmdSizeForComputeMode();
     size += getCmdSizeForMediaSampler(dispatchFlags.mediaSamplerRequired);
     size += getCmdSizeForPipelineSelect();
     size += getCmdSizeForPreemption(dispatchFlags);

runtime/command_stream/csr_definitions.h

@@ -9,6 +9,7 @@
 #include "runtime/memory_manager/memory_constants.h"
 #include "runtime/helpers/hw_info.h"
 #include "runtime/helpers/properties_helper.h"
+#include "runtime/kernel/grf_config.h"
 #include <limits>
 
 namespace OCLRT {
@@ -43,6 +44,7 @@ struct DispatchFlags {
     FlushStampTrackingObj *flushStampReference = nullptr;
     PreemptionMode preemptionMode = PreemptionMode::Disabled;
     EventsRequest *outOfDeviceDependencies = nullptr;
+    uint32_t numGrfRequired = GrfConfig::DefaultGrfNumber;
 };
 
 struct CsrSizeRequestFlags {
@@ -51,5 +53,6 @@ struct CsrSizeRequestFlags {
     bool preemptionRequestChanged = false;
     bool mediaSamplerConfigChanged = false;
     bool hasSharedHandles = false;
+    bool numGrfRequiredChanged = false;
 };
 } // namespace OCLRT

runtime/gen10/command_stream_receiver_hw_gen10.cpp

@@ -16,7 +16,7 @@ typedef CNLFamily Family;
 static auto gfxCore = IGFX_GEN10_CORE;
 
 template <>
-size_t CommandStreamReceiverHw<Family>::getCmdSizeForCoherency() {
+size_t CommandStreamReceiverHw<Family>::getCmdSizeForComputeMode() {
     if (csrSizeRequestFlags.coherencyRequestChanged) {
         return sizeof(typename Family::MI_LOAD_REGISTER_IMM);
     }
@@ -24,7 +24,7 @@ size_t CommandStreamReceiverHw<Family>::getCmdSizeForCoherency() {
 }
 
 template <>
-void CommandStreamReceiverHw<Family>::programCoherency(LinearStream &stream, DispatchFlags &dispatchFlags) {
+void CommandStreamReceiverHw<Family>::programComputeMode(LinearStream &stream, DispatchFlags &dispatchFlags) {
     if (csrSizeRequestFlags.coherencyRequestChanged) {
         LriHelper<Family>::program(&stream, gen10HdcModeRegisterAddresss, DwordBuilder::build(4, true, !dispatchFlags.requiresCoherency));
         this->lastSentCoherencyRequest = static_cast<int8_t>(dispatchFlags.requiresCoherency);

runtime/gen8/command_stream_receiver_hw_gen8.cpp

@@ -15,12 +15,12 @@ typedef BDWFamily Family;
 static auto gfxCore = IGFX_GEN8_CORE;
 
 template <>
-size_t CommandStreamReceiverHw<Family>::getCmdSizeForCoherency() {
+size_t CommandStreamReceiverHw<Family>::getCmdSizeForComputeMode() {
     return 0;
 }
 
 template <>
-void CommandStreamReceiverHw<Family>::programCoherency(LinearStream &stream, DispatchFlags &dispatchFlags) {
+void CommandStreamReceiverHw<Family>::programComputeMode(LinearStream &stream, DispatchFlags &dispatchFlags) {
 }
 
 template <>

runtime/gen9/command_stream_receiver_hw_gen9.cpp

@@ -15,12 +15,12 @@ typedef SKLFamily Family;
 static auto gfxCore = IGFX_GEN9_CORE;
 
 template <>
-size_t CommandStreamReceiverHw<Family>::getCmdSizeForCoherency() {
+size_t CommandStreamReceiverHw<Family>::getCmdSizeForComputeMode() {
     return 0;
 }
 
 template <>
-void CommandStreamReceiverHw<Family>::programCoherency(LinearStream &stream, DispatchFlags &dispatchFlags) {
+void CommandStreamReceiverHw<Family>::programComputeMode(LinearStream &stream, DispatchFlags &dispatchFlags) {
 }
 
 template <>

runtime/kernel/CMakeLists.txt

@@ -7,11 +7,13 @@
 set(RUNTIME_SRCS_KERNEL
   ${CMAKE_CURRENT_SOURCE_DIR}/CMakeLists.txt
   ${CMAKE_CURRENT_SOURCE_DIR}/dynamic_kernel_info.h
+  ${CMAKE_CURRENT_SOURCE_DIR}/grf_config.h
   ${CMAKE_CURRENT_SOURCE_DIR}/image_transformer.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/image_transformer.h
   ${CMAKE_CURRENT_SOURCE_DIR}/kernel.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/kernel.h
   ${CMAKE_CURRENT_SOURCE_DIR}/kernel.inl
+  ${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}/kernel_reconfiguration.cpp
 )
 target_sources(${NEO_STATIC_LIB_NAME} PRIVATE ${RUNTIME_SRCS_KERNEL})
 set_property(GLOBAL PROPERTY RUNTIME_SRCS_KERNEL ${RUNTIME_SRCS_KERNEL})

runtime/kernel/grf_config.h

@@ -0,0 +1,12 @@
+/*
+ * Copyright (C) 2018 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#pragma once
+
+namespace GrfConfig {
+constexpr uint32_t DefaultGrfNumber = 128u;
+}

runtime/kernel/kernel.cpp

@@ -354,6 +354,8 @@ cl_int Kernel::initialize() {
             program->allocateBlockPrivateSurfaces();
         }
 
+        reconfigureKernel();
+
         retVal = CL_SUCCESS;
 
     } while (false);

runtime/kernel/kernel.h

@@ -458,6 +458,8 @@ class Kernel : public BaseObject<_cl_kernel> {
 
     void resolveArgs();
 
+    void reconfigureKernel();
+
     Program *program;
     Context *context;
     const Device &device;

runtime/kernel/kernel_reconfiguration.cpp

@@ -0,0 +1,13 @@
+/*
+ * Copyright (C) 2018 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "runtime/kernel/kernel.h"
+
+namespace OCLRT {
+void Kernel::reconfigureKernel() {
+}
+} // namespace OCLRT

runtime/program/kernel_info.h

@@ -239,5 +239,6 @@ struct KernelInfo {
     bool isKernelHeapSubstituted = false;
     GraphicsAllocation *kernelAllocation = nullptr;
     DebugData debugData;
+    bool computeMode = false;
 };
 } // namespace OCLRT

runtime/program/process_gen_binary.cpp

@@ -535,7 +535,8 @@ cl_int Program::parsePatchList(KernelInfo &kernelInfo) {
                     "\n  .SubgroupIndependentForwardProgressRequired", kernelInfo.patchInfo.executionEnvironment->SubgroupIndependentForwardProgressRequired,
                     "\n  .WorkgroupWalkOrderDim0", kernelInfo.workgroupWalkOrder[0],
                     "\n  .WorkgroupWalkOrderDim1", kernelInfo.workgroupWalkOrder[1],
-                    "\n  .WorkgroupWalkOrderDim2", kernelInfo.workgroupWalkOrder[2]);
+                    "\n  .WorkgroupWalkOrderDim2", kernelInfo.workgroupWalkOrder[2],
+                    "\n  .NumGRFRequired", kernelInfo.patchInfo.executionEnvironment->NumGRFRequired);
             break;
 
         case PATCH_TOKEN_DATA_PARAMETER_STREAM:

unit_tests/gen10/coherency_tests_gen10.cpp

@@ -42,19 +42,19 @@ struct Gen10CoherencyRequirements : public ::testing::Test {
 GEN10TEST_F(Gen10CoherencyRequirements, coherencyCmdSize) {
     auto lriSize = sizeof(MI_LOAD_REGISTER_IMM);
     overrideCoherencyRequest(false, false);
-    auto retSize = csr->getCmdSizeForCoherency();
+    auto retSize = csr->getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
 
     overrideCoherencyRequest(false, true);
-    retSize = csr->getCmdSizeForCoherency();
+    retSize = csr->getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
 
     overrideCoherencyRequest(true, true);
-    retSize = csr->getCmdSizeForCoherency();
+    retSize = csr->getCmdSizeForComputeMode();
     EXPECT_EQ(lriSize, retSize);
 
     overrideCoherencyRequest(true, false);
-    retSize = csr->getCmdSizeForCoherency();
+    retSize = csr->getCmdSizeForComputeMode();
     EXPECT_EQ(lriSize, retSize);
 }
 
@@ -68,14 +68,14 @@ GEN10TEST_F(Gen10CoherencyRequirements, coherencyCmdValues) {
     expectedCmd.setDataDword(DwordBuilder::build(4, true));
 
     overrideCoherencyRequest(true, false);
-    csr->programCoherency(stream, flags);
+    csr->programComputeMode(stream, flags);
     EXPECT_EQ(lriSize, stream.getUsed());
 
     auto cmd = reinterpret_cast<MI_LOAD_REGISTER_IMM *>(stream.getCpuBase());
     EXPECT_TRUE(memcmp(&expectedCmd, cmd, lriSize) == 0);
 
     overrideCoherencyRequest(true, true);
-    csr->programCoherency(stream, flags);
+    csr->programComputeMode(stream, flags);
     EXPECT_EQ(lriSize * 2, stream.getUsed());
 
     cmd = reinterpret_cast<MI_LOAD_REGISTER_IMM *>(ptrOffset(stream.getCpuBase(), lriSize));

unit_tests/gen8/coherency_tests_gen8.cpp

@@ -19,14 +19,14 @@ GEN8TEST_F(Gen8CoherencyRequirements, noCoherencyProgramming) {
     LinearStream stream;
     DispatchFlags flags = {};
 
-    auto retSize = csr.getCmdSizeForCoherency();
+    auto retSize = csr.getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
-    csr.programCoherency(stream, flags);
+    csr.programComputeMode(stream, flags);
     EXPECT_EQ(0u, stream.getUsed());
 
     flags.requiresCoherency = true;
-    retSize = csr.getCmdSizeForCoherency();
+    retSize = csr.getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
-    csr.programCoherency(stream, flags);
+    csr.programComputeMode(stream, flags);
     EXPECT_EQ(0u, stream.getUsed());
 }

unit_tests/gen9/coherency_tests_gen9.cpp

@@ -19,14 +19,14 @@ GEN9TEST_F(Gen9CoherencyRequirements, noCoherencyProgramming) {
     LinearStream stream;
     DispatchFlags flags = {};
 
-    auto retSize = csr.getCmdSizeForCoherency();
+    auto retSize = csr.getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
-    csr.programCoherency(stream, flags);
+    csr.programComputeMode(stream, flags);
     EXPECT_EQ(0u, stream.getUsed());
 
     flags.requiresCoherency = true;
-    retSize = csr.getCmdSizeForCoherency();
+    retSize = csr.getCmdSizeForComputeMode();
     EXPECT_EQ(0u, retSize);
-    csr.programCoherency(stream, flags);
+    csr.programComputeMode(stream, flags);
     EXPECT_EQ(0u, stream.getUsed());
 }

unit_tests/mocks/mock_kernel.h

@@ -8,6 +8,7 @@
 #pragma once
 
 #include "runtime/helpers/string.h"
+#include "runtime/kernel/grf_config.h"
 #include "runtime/kernel/kernel.h"
 #include "runtime/scheduler/scheduler_kernel.h"
 #include "runtime/device/device.h"
@@ -118,6 +119,7 @@ class MockKernel : public Kernel {
         SPatchExecutionEnvironment *executionEnvironment = new SPatchExecutionEnvironment;
         memset(executionEnvironment, 0, sizeof(SPatchExecutionEnvironment));
         executionEnvironment->HasDeviceEnqueue = 0;
+        executionEnvironment->NumGRFRequired = GrfConfig::DefaultGrfNumber;
         info->patchInfo.executionEnvironment = executionEnvironment;
 
         info->crossThreadData = new char[crossThreadSize];
@@ -243,6 +245,8 @@ class MockKernelWithInternals {
         memset(&executionEnvironment, 0, sizeof(SPatchExecutionEnvironment));
         memset(&executionEnvironmentBlock, 0, sizeof(SPatchExecutionEnvironment));
         memset(&dataParameterStream, 0, sizeof(SPatchDataParameterStream));
+        executionEnvironment.NumGRFRequired = GrfConfig::DefaultGrfNumber;
+        executionEnvironmentBlock.NumGRFRequired = GrfConfig::DefaultGrfNumber;
         kernelHeader.SurfaceStateHeapSize = sizeof(sshLocal);
         threadPayload.LocalIDXPresent = 1;
         threadPayload.LocalIDYPresent = 1;
@@ -324,6 +328,7 @@ class MockParentKernel : public Kernel {
         SPatchExecutionEnvironment *executionEnvironment = new SPatchExecutionEnvironment;
         *executionEnvironment = {};
         executionEnvironment->HasDeviceEnqueue = 1;
+        executionEnvironment->NumGRFRequired = GrfConfig::DefaultGrfNumber;
         info->patchInfo.executionEnvironment = executionEnvironment;
 
         SPatchAllocateStatelessDefaultDeviceQueueSurface *allocateDeviceQueue = new SPatchAllocateStatelessDefaultDeviceQueueSurface;