Store crossThreadData per root device in Kernel

Related-To: NEO-5001
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>

Author: JablonskiMateusz

opencl/source/built_ins/vme_dispatch_builder.h

@@ -13,6 +13,7 @@
 #include "opencl/source/accelerators/intel_motion_estimation.h"
 #include "opencl/source/built_ins/built_in_ops_vme.h"
 #include "opencl/source/built_ins/builtins_dispatch_builder.h"
+#include "opencl/source/cl_device/cl_device.h"
 #include "opencl/source/helpers/dispatch_info_builder.h"
 #include "opencl/source/mem_obj/buffer.h"
 #include "opencl/source/mem_obj/image.h"
@@ -167,7 +168,7 @@ class VmeBuiltinDispatchInfoBuilder : public BuiltinDispatchInfoBuilder {
         DEBUG_BREAK_IF(kai.kernelArgPatchInfoVector.size() != 1);
         const KernelArgPatchInfo &patchInfo = kai.kernelArgPatchInfoVector[0];
         DEBUG_BREAK_IF(sizeof(RetType) > patchInfo.size);
-        return *(RetType *)(vmeKernel->getCrossThreadData() + patchInfo.crossthreadOffset);
+        return *(RetType *)(vmeKernel->getCrossThreadData(clDevice.getRootDeviceIndex()) + patchInfo.crossthreadOffset);
     }
 
     cl_int validateImages(Vec3<size_t> inputRegion, Vec3<size_t> offset) const {

opencl/source/command_queue/gpgpu_walker_bdw_plus.inl

@@ -150,7 +150,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
         &cmdWalker,
         nullptr,
         true,
-        devQueueHw.getDevice().getHardwareInfo());
+        devQueueHw.getDevice());
 
     // Implement enabling special WA DisableLSQCROPERFforOCL if needed
     GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, true);

opencl/source/command_queue/hardware_interface_bdw_plus.inl

@@ -102,7 +102,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         &walkerCmd,
         nullptr,
         true,
-        commandQueue.getDevice().getHardwareInfo());
+        commandQueue.getDevice());
 
     GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&walkerCmd, globalOffsets, startWorkGroups,
                                                            numWorkGroups, localWorkSizes, simd, dim,

opencl/source/helpers/hardware_commands_helper.h

@@ -76,7 +76,8 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
         Kernel &kernel,
         bool inlineDataProgrammingRequired,
         WALKER_TYPE<GfxFamily> *walkerCmd,
-        uint32_t &sizeCrossThreadData);
+        uint32_t &sizeCrossThreadData,
+        uint32_t rootDeviceIndex);
 
     static size_t sendIndirectState(
         LinearStream &commandStream,
@@ -93,7 +94,7 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
         WALKER_TYPE<GfxFamily> *walkerCmd,
         INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
         bool localIdsGenerationByRuntime,
-        const HardwareInfo &hardwareInfo);
+        const Device &device);
 
     static void programPerThreadData(
         size_t &sizePerThreadData,
@@ -121,6 +122,7 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
     static size_t getSizeRequiredDSH(
         const Kernel &kernel);
     static size_t getSizeRequiredIOH(
+        uint32_t rootDeviceIndex,
         const Kernel &kernel,
         size_t localWorkSize = 256);
     static size_t getSizeRequiredSSH(

opencl/source/helpers/hardware_commands_helper_base.inl

@@ -58,6 +58,7 @@ size_t HardwareCommandsHelper<GfxFamily>::getSizeRequiredDSH(
 
 template <typename GfxFamily>
 size_t HardwareCommandsHelper<GfxFamily>::getSizeRequiredIOH(
+    uint32_t rootDeviceIndex,
     const Kernel &kernel,
     size_t localWorkSize) {
     typedef typename GfxFamily::WALKER_TYPE WALKER_TYPE;
@@ -67,7 +68,7 @@ size_t HardwareCommandsHelper<GfxFamily>::getSizeRequiredIOH(
 
     auto numChannels = PerThreadDataHelper::getNumLocalIdChannels(*threadPayload);
     uint32_t grfSize = sizeof(typename GfxFamily::GRF);
-    return alignUp((kernel.getCrossThreadDataSize() +
+    return alignUp((kernel.getCrossThreadDataSize(rootDeviceIndex) +
                     getPerThreadDataSizeTotal(kernel.getKernelInfo().getMaxSimdSize(), grfSize, numChannels, localWorkSize)),
                    WALKER_TYPE::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
 }
@@ -102,7 +103,10 @@ size_t HardwareCommandsHelper<GfxFamily>::getTotalSizeRequiredDSH(
 template <typename GfxFamily>
 size_t HardwareCommandsHelper<GfxFamily>::getTotalSizeRequiredIOH(
     const MultiDispatchInfo &multiDispatchInfo) {
-    return getSizeRequired(multiDispatchInfo, [](const DispatchInfo &dispatchInfo) { return getSizeRequiredIOH(*dispatchInfo.getKernel(), Math::computeTotalElementsCount(dispatchInfo.getLocalWorkgroupSize())); });
+    return getSizeRequired(multiDispatchInfo, [](const DispatchInfo &dispatchInfo) { return getSizeRequiredIOH(
+                                                                                         dispatchInfo.getClDevice().getRootDeviceIndex(),
+                                                                                         *dispatchInfo.getKernel(),
+                                                                                         Math::computeTotalElementsCount(dispatchInfo.getLocalWorkgroupSize())); });
 }
 
 template <typename GfxFamily>
@@ -215,10 +219,13 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
     WALKER_TYPE<GfxFamily> *walkerCmd,
     INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
     bool localIdsGenerationByRuntime,
-    const HardwareInfo &hardwareInfo) {
+    const Device &device) {
 
     using SAMPLER_STATE = typename GfxFamily::SAMPLER_STATE;
 
+    auto &hardwareInfo = device.getHardwareInfo();
+    auto rootDeviceIndex = device.getRootDeviceIndex();
+
     DEBUG_BREAK_IF(simd != 1 && simd != 8 && simd != 16 && simd != 32);
     auto inlineDataProgrammingRequired = HardwareCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(kernel);
 
@@ -227,7 +234,7 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
     const auto &patchInfo = kernelInfo.patchInfo;
 
     ssh.align(BINDING_TABLE_STATE::SURFACESTATEPOINTER_ALIGN_SIZE);
-    kernel.patchBindlessSurfaceStateOffsets(ssh.getUsed());
+    kernel.patchBindlessSurfaceStateOffsets(device, ssh.getUsed());
 
     auto dstBindingTablePointer = EncodeSurfaceState<GfxFamily>::pushBindingTableAndSurfaceStates(ssh, (kernelInfo.patchInfo.bindingTableState != nullptr) ? kernelInfo.patchInfo.bindingTableState->Count : 0,
                                                                                                   kernel.getSurfaceStateHeap(), kernel.getSurfaceStateHeapSize(),
@@ -248,11 +255,11 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
     auto threadsPerThreadGroup = static_cast<uint32_t>(getThreadsPerWG(simd, localWorkItems));
     auto numChannels = PerThreadDataHelper::getNumLocalIdChannels(*threadPayload);
 
-    uint32_t sizeCrossThreadData = kernel.getCrossThreadDataSize();
+    uint32_t sizeCrossThreadData = kernel.getCrossThreadDataSize(rootDeviceIndex);
 
     size_t offsetCrossThreadData = HardwareCommandsHelper<GfxFamily>::sendCrossThreadData(
         ioh, kernel, inlineDataProgrammingRequired,
-        walkerCmd, sizeCrossThreadData);
+        walkerCmd, sizeCrossThreadData, rootDeviceIndex);
 
     size_t sizePerThreadDataTotal = 0;
     size_t sizePerThreadData = 0;

opencl/source/helpers/hardware_commands_helper_bdw_plus.inl

@@ -125,12 +125,13 @@ size_t HardwareCommandsHelper<GfxFamily>::sendCrossThreadData(
     Kernel &kernel,
     bool inlineDataProgrammingRequired,
     WALKER_TYPE<GfxFamily> *walkerCmd,
-    uint32_t &sizeCrossThreadData) {
+    uint32_t &sizeCrossThreadData,
+    uint32_t rootDeviceIndex) {
     indirectHeap.align(WALKER_TYPE<GfxFamily>::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
 
     auto offsetCrossThreadData = indirectHeap.getUsed();
     char *pDest = static_cast<char *>(indirectHeap.getSpace(sizeCrossThreadData));
-    memcpy_s(pDest, sizeCrossThreadData, kernel.getCrossThreadData(), sizeCrossThreadData);
+    memcpy_s(pDest, sizeCrossThreadData, kernel.getCrossThreadData(rootDeviceIndex), sizeCrossThreadData);
 
     if (DebugManager.flags.AddPatchInfoCommentsForAUBDump.get()) {
         FlatBatchBufferHelper::fixCrossThreadDataInfo(kernel.getPatchInfoDataList(), offsetCrossThreadData, indirectHeap.getGraphicsAllocation()->getGpuAddress());