Add new parameter to dispatch payload data

Change-Id: I0034c5a40de65a050e19691b13793b7053354757

Author: zzdanowicz

Jenkinsfile

@@ -1,5 +1,5 @@
 #!groovy
 neoDependenciesRev='800243-1090'
 strategy='EQUAL'
-allowedCD=271
+allowedCD=272
 allowedF=4

runtime/command_queue/gpgpu_walker.inl

@@ -553,13 +553,15 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
 
     // Send our indirect object data
     size_t localWorkSizes[3] = {scheduler.getLws(), 1, 1};
+    size_t globalWorkSizes[3] = {scheduler.getGws(), 1, 1};
 
     // Create indirectHeap for IOH that is located at the end of device enqueue DSH
     size_t curbeOffset = devQueueHw.setSchedulerCrossThreadData(scheduler);
     IndirectHeap indirectObjectHeap(dsh->getCpuBase(), dsh->getMaxAvailableSpace());
     indirectObjectHeap.getSpace(curbeOffset);
     ioh = &indirectObjectHeap;
 
+    bool localIdsGeneration = KernelCommandsHelper<GfxFamily>::isDispatchForLocalIdsGeneration(1, globalWorkSizes, localWorkSizes);
     auto offsetCrossThreadData = KernelCommandsHelper<GfxFamily>::sendIndirectState(
         *commandStream,
         *dsh,
@@ -571,7 +573,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
         offsetInterfaceDescriptorTable,
         interfaceDescriptorIndex,
         preemptionMode,
-        nullptr);
+        nullptr,
+        localIdsGeneration);
 
     // Implement enabling special WA DisableLSQCROPERFforOCL if needed
     GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(commandStream, scheduler, true);

runtime/command_queue/hardware_interface/hardware_interface.inl

@@ -174,6 +174,7 @@ void HardwareInterface<GfxFamily>::dispatchWalker(
 
         // Send our indirect object data
         size_t localWorkSizes[3] = {lws.x, lws.y, lws.z};
+        size_t globalWorkSizes[3] = {gws.x, gws.y, gws.z};
 
         dispatchProfilingPerfStartCommands(dispatchInfo, multiDispatchInfo, hwTimeStamps,
                                            hwPerfCounter, commandStream, commandQueue);
@@ -197,6 +198,7 @@ void HardwareInterface<GfxFamily>::dispatchWalker(
 
         auto idd = obtainInterfaceDescriptorData(pWalkerCmd);
 
+        bool localIdsGeneration = KernelCommandsHelper<GfxFamily>::isDispatchForLocalIdsGeneration(dim, globalWorkSizes, localWorkSizes);
         auto offsetCrossThreadData = KernelCommandsHelper<GfxFamily>::sendIndirectState(
             *commandStream,
             *dsh,
@@ -208,7 +210,8 @@ void HardwareInterface<GfxFamily>::dispatchWalker(
             offsetInterfaceDescriptorTable,
             interfaceDescriptorIndex,
             preemptionMode,
-            idd);
+            idd,
+            localIdsGeneration);
 
         size_t globalOffsets[3] = {offset.x, offset.y, offset.z};
         size_t startWorkGroups[3] = {swgs.x, swgs.y, swgs.z};
@@ -218,21 +221,24 @@ void HardwareInterface<GfxFamily>::dispatchWalker(
 
         DEBUG_BREAK_IF(offsetCrossThreadData % 64 != 0);
         setOffsetCrossThreadData(pWalkerCmd, offsetCrossThreadData, interfaceDescriptorIndex);
+        auto sizeCrossThreadData = kernel.getCrossThreadDataSize();
 
-        auto threadPayload = kernel.getKernelInfo().patchInfo.threadPayload;
-        DEBUG_BREAK_IF(nullptr == threadPayload);
+        size_t sizePerThreadDataTotal = 0;
+        if (localIdsGeneration) {
+            auto threadPayload = kernel.getKernelInfo().patchInfo.threadPayload;
+            DEBUG_BREAK_IF(nullptr == threadPayload);
 
-        auto numChannels = PerThreadDataHelper::getNumLocalIdChannels(*threadPayload);
-        auto localIdSizePerThread = PerThreadDataHelper::getLocalIdSizePerThread(simd, numChannels);
-        localIdSizePerThread = std::max(localIdSizePerThread, sizeof(GRF));
+            auto numChannels = PerThreadDataHelper::getNumLocalIdChannels(*threadPayload);
+            auto localIdSizePerThread = PerThreadDataHelper::getLocalIdSizePerThread(simd, numChannels);
+            localIdSizePerThread = std::max(localIdSizePerThread, sizeof(GRF));
 
-        auto sizePerThreadDataTotal = getThreadsPerWG(simd, localWorkSize) * localIdSizePerThread;
-        DEBUG_BREAK_IF(sizePerThreadDataTotal == 0); // Hardware requires at least 1 GRF of perThreadData for each thread in thread group
+            sizePerThreadDataTotal = getThreadsPerWG(simd, localWorkSize) * localIdSizePerThread;
+            DEBUG_BREAK_IF(sizePerThreadDataTotal == 0); // Hardware requires at least 1 GRF of perThreadData for each thread in thread group
+        }
 
-        auto sizeCrossThreadData = kernel.getCrossThreadDataSize();
-        auto IndirectDataLength = alignUp(static_cast<uint32_t>(sizeCrossThreadData + sizePerThreadDataTotal),
+        auto indirectDataLength = alignUp(static_cast<uint32_t>(sizeCrossThreadData + sizePerThreadDataTotal),
                                           WALKER_TYPE<GfxFamily>::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
-        pWalkerCmd->setIndirectDataLength(IndirectDataLength);
+        pWalkerCmd->setIndirectDataLength(indirectDataLength);
 
         dispatchWorkarounds(commandStream, commandQueue, kernel, false);
         currentDispatchIndex++;

runtime/helpers/kernel_commands.h

@@ -89,7 +89,8 @@ struct KernelCommandsHelper : public PerThreadDataHelper {
         const uint64_t offsetInterfaceDescriptorTable,
         const uint32_t interfaceDescriptorIndex,
         PreemptionMode preemptionMode,
-        INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor);
+        INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
+        bool localIdsGeneration);
 
     static size_t getSizeRequiredCS();
     static bool isPipeControlWArequired();
@@ -151,5 +152,7 @@ struct KernelCommandsHelper : public PerThreadDataHelper {
     static const uint32_t alignIndirectStatePointer = 64 * sizeof(uint8_t);
 
     static bool doBindingTablePrefetch();
+
+    static bool isDispatchForLocalIdsGeneration(uint32_t workDim, size_t *gws, size_t *lws);
 };
 } // namespace OCLRT

runtime/helpers/kernel_commands.inl

@@ -294,7 +294,8 @@ size_t KernelCommandsHelper<GfxFamily>::sendIndirectState(
     const uint64_t offsetInterfaceDescriptorTable,
     const uint32_t interfaceDescriptorIndex,
     PreemptionMode preemptionMode,
-    INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor) {
+    INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
+    bool localIdsGeneration) {
     using SAMPLER_STATE = typename GfxFamily::SAMPLER_STATE;
 
     DEBUG_BREAK_IF(simd != 8 && simd != 16 && simd != 32);
@@ -422,4 +423,9 @@ template <typename GfxFamily>
 bool KernelCommandsHelper<GfxFamily>::doBindingTablePrefetch() {
     return true;
 }
+
+template <typename GfxFamily>
+bool KernelCommandsHelper<GfxFamily>::isDispatchForLocalIdsGeneration(uint32_t workDim, size_t *gws, size_t *lws) {
+    return true;
+}
 } // namespace OCLRT

unit_tests/helpers/kernel_commands_tests.cpp

@@ -317,7 +317,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, sendIndirectStateResourceUsage)
         IDToffset,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     // It's okay these are EXPECT_GE as they're only going to be used for
     // estimation purposes to avoid OOM.
@@ -364,7 +365,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, givenKernelWithFourBindingTableE
         0,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     if (KernelCommandsHelper<FamilyType>::doBindingTablePrefetch()) {
@@ -404,7 +406,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, givenKernelThatIsSchedulerWhenIn
         0,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     EXPECT_EQ(0u, interfaceDescriptor->getBindingTableEntryCount());
@@ -438,7 +441,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, givenKernelWith100BindingTableEn
         0,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     if (KernelCommandsHelper<FamilyType>::doBindingTablePrefetch()) {
@@ -503,7 +507,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, whenSendingIndirectStateThenKern
         IDToffset,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
     size_t numThreads = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
     numThreads = (numThreads + modifiedKernelInfo.getMaxSimdSize() - 1) / modifiedKernelInfo.getMaxSimdSize();
     size_t expectedIohSize = ((modifiedKernelInfo.getMaxSimdSize() == 32) ? 32 : 16) * 3 * numThreads * sizeof(uint16_t);
@@ -575,7 +580,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, usedBindingTableStatePointer) {
         0,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     EXPECT_EQ(0x00000000u, *(&bindingTableStatesPointers[0]));
     EXPECT_EQ(0x00000040u, *(&bindingTableStatesPointers[1]));
@@ -728,7 +734,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, usedBindingTableStatePointersFor
             0,
             0,
             pDevice->getPreemptionMode(),
-            nullptr);
+            nullptr,
+            true);
 
         bti = reinterpret_cast<typename FamilyType::BINDING_TABLE_STATE *>(reinterpret_cast<unsigned char *>(ssh.getCpuBase()) + localSshOffset + btiOffset);
         for (uint32_t i = 0; i < numSurfaces; ++i) {
@@ -962,7 +969,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, KernelCommandsTest, GivenKernelWithSamplersWhenIndir
         interfaceDescriptorTableOffset,
         0,
         pDevice->getPreemptionMode(),
-        nullptr);
+        nullptr,
+        true);
 
     bool isMemorySame = memcmp(borderColorPointer, mockDsh, borderColorSize) == 0;
     EXPECT_TRUE(isMemorySame);