Support for syncBuffer implicit argument

aratajew · sys_zuul · commit e31699d360fa · 2019-10-25T05:07:00.000-07:00
Change-Id: I3c6ad0ed1e2c9c8f3f929a4c188495c0f02553c6
diff --git a/IGC/AdaptorCommon/ImplicitArgs.cpp b/IGC/AdaptorCommon/ImplicitArgs.cpp
@@ -300,6 +300,7 @@ ImplicitArgs::ImplicitArgs(const llvm::Function& func , const MetaDataUtils* pMd
         IMPLICIT_ARGS.push_back(ImplicitArg(ImplicitArg::STAGE_IN_GRID_ORIGIN, "stageInGridOrigin", ImplicitArg::INT, WIAnalysis::UNIFORM, 3, ImplicitArg::ALIGN_GRF, true));
         IMPLICIT_ARGS.push_back(ImplicitArg(ImplicitArg::STAGE_IN_GRID_SIZE, "stageInGridSize", ImplicitArg::INT, WIAnalysis::UNIFORM, 3, ImplicitArg::ALIGN_GRF, true));
 
+        IMPLICIT_ARGS.push_back(ImplicitArg(ImplicitArg::SYNC_BUFFER, "syncBuffer", ImplicitArg::GLOBALPTR, WIAnalysis::UNIFORM, 1, ImplicitArg::ALIGN_PTR, false));
 
         assert(IMPLICIT_ARGS.size() == ImplicitArg::NUM_IMPLICIT_ARGS && "Mismatch in NUM_IMPLICIT_ARGS and IMPLICIT_ARGS vector");
 
diff --git a/IGC/AdaptorCommon/ImplicitArgs.hpp b/IGC/AdaptorCommon/ImplicitArgs.hpp
@@ -122,6 +122,7 @@ namespace IGC
             STAGE_IN_GRID_ORIGIN,
             STAGE_IN_GRID_SIZE,
 
+            SYNC_BUFFER,
 
             NUM_IMPLICIT_ARGS
         };
diff --git a/IGC/AdaptorOCL/OCL/KernelAnnotations.hpp b/IGC/AdaptorOCL/OCL/KernelAnnotations.hpp
@@ -222,6 +222,13 @@ struct PrintfBufferAnnotation : KernelArgumentAnnotation
     DWORD  PayloadPosition;
 };
 
+// Annotation for sync buffer.
+struct SyncBufferAnnotation : KernelArgumentAnnotation
+{
+    DWORD DataSize;
+    DWORD PayloadPosition;
+};
+
 // Generated by front end
 struct KernelConstantRegisterAnnotation
 {
diff --git a/IGC/AdaptorOCL/OCL/Patch/patch_parser.cpp b/IGC/AdaptorOCL/OCL/Patch/patch_parser.cpp
@@ -1005,6 +1005,27 @@ void DebugPatchList(
             }
             break;
 
+        case iOpenCL::PATCH_TOKEN_ALLOCATE_SYNC_BUFFER:
+            {
+                const iOpenCL::SPatchAllocateSyncBuffer* pPatchItem =
+                    (const iOpenCL::SPatchAllocateSyncBuffer*)pHeader;
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "PATCH_TOKEN_ALLOCATE_SYNC_BUFFER (%08X) (size = %d)\n",
+                    pPatchItem->Token,
+                    pPatchItem->Size);
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "\tSurfaceStateHeapOffset = %d\n",
+                    pPatchItem->SurfaceStateHeapOffset);
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "\tDataParamOffset = %d\n",
+                    pPatchItem->DataParamOffset);
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "\tDataParamSize = %d\n",
+                    pPatchItem->DataParamSize);
+            }
+            break;
+
         // Stateless Tokens
         case iOpenCL::PATCH_TOKEN_STATELESS_GLOBAL_MEMORY_OBJECT_KERNEL_ARGUMENT:
             {
diff --git a/IGC/AdaptorOCL/OCL/sp/sp_g8.cpp b/IGC/AdaptorOCL/OCL/sp/sp_g8.cpp
@@ -758,6 +758,21 @@ RETVAL CGen8OpenCLStateProcessor::CreateSurfaceStateHeap(
                     false)));
     }
 
+    if (annotations.m_syncBufferAnnotation != NULL)
+    {
+        unsigned int bti = annotations.m_argIndexMap.at(annotations.m_syncBufferAnnotation->ArgumentNumber);
+        context.Surface.SurfaceOffset[bti] = (DWORD)membuf.Size();
+
+        SurfaceStates.insert(
+            std::make_pair(
+                bti,
+                SurfaceState(
+                    SURFACE_BUFFER,
+                    SURFACE_FORMAT_RAW,
+                    0,
+                    false)));
+    }
+
     // If GT-Pin is enabled, assign btis to GT-Pin's surfaces
     if (gtpinEnabled)
     {
@@ -1659,6 +1674,32 @@ RETVAL CGen8OpenCLStateProcessor::CreatePatchList(
         }
     }
 
+    // Patch for Sync Buffer Offset
+    if (retValue.Success)
+    {
+        if (annotations.m_syncBufferAnnotation != nullptr)
+        {
+            iOpenCL::SyncBufferAnnotation* syncBufAnn = annotations.m_syncBufferAnnotation;
+
+            iOpenCL::SPatchAllocateSyncBuffer  patch;
+            memset(&patch, 0, sizeof(patch));
+
+            unsigned int bti = annotations.m_argIndexMap.at(syncBufAnn->ArgumentNumber);
+
+            patch.Token = iOpenCL::PATCH_TOKEN_ALLOCATE_SYNC_BUFFER;
+            patch.Size = sizeof(patch);
+            patch.SurfaceStateHeapOffset = context.Surface.SurfaceOffset[bti];
+            patch.DataParamOffset = syncBufAnn->PayloadPosition;
+            patch.DataParamSize = syncBufAnn->DataSize;
+
+            dataParameterStreamSize = std::max(
+                dataParameterStreamSize,
+                syncBufAnn->PayloadPosition + syncBufAnn->DataSize);
+
+            retValue = AddPatchItem(patch, membuf);
+        }
+    }
+
     // Pointer inputs with initializer
     if( retValue.Success )
     {
diff --git a/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h b/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h
@@ -33,7 +33,7 @@ Abstract:   Contains common patch structure definitions
 
 namespace iOpenCL
 {
-const uint32_t CURRENT_ICBE_VERSION = 1061;
+const uint32_t CURRENT_ICBE_VERSION = 1062;
 
 const uint32_t MAGIC_CL = 0x494E5443;      // 'I', 'N', 'T', 'C'
 const uint32_t INVALID_INDEX = 0xFFFFFFFF;
@@ -150,12 +150,13 @@ enum PATCH_TOKEN
     PATCH_TOKEN_PROGRAM_SYMBOL_TABLE,                             // 53  @SPatchFunctionTableInfo@
     PATCH_TOKEN_PROGRAM_RELOCATION_TABLE,                         // 54  @SPatchFunctionTableInfo@
     PATCH_TOKEN_MEDIA_VFE_STATE_SLOT1,                            // 55  @SPatchMediaVFEState of slot1@
+    PATCH_TOKEN_ALLOCATE_SYNC_BUFFER,                             // 56  @SPatchAllocateSyncBuffer@
 
     NUM_PATCH_TOKENS
 };
 
 // Update CURRENT_ICBE_VERSION when modifying the patch list
-static_assert( NUM_PATCH_TOKENS == 56, "NUM_PATCH_TOKENS has invalid value");
+static_assert( NUM_PATCH_TOKENS == 57, "NUM_PATCH_TOKENS has invalid value");
 
 /*****************************************************************************\
 ENUM: IMAGE_MEMORY_OBJECT_TYPE
diff --git a/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_shared.h b/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_shared.h
@@ -444,6 +444,17 @@ struct SPatchAllocateStatelessPrintfSurface :
 // Update CURRENT_ICBE_VERSION when modifying the patch list
 static_assert( sizeof( SPatchAllocateStatelessPrintfSurface ) == ( 16 + sizeof( SPatchItemHeader ) ) , "The size of SPatchAllocateStatelessPrintfSurface is not what is expected" );
 
+struct SPatchAllocateSyncBuffer :
+       SPatchItemHeader
+{
+    uint32_t   SurfaceStateHeapOffset;
+    uint32_t   DataParamOffset;
+    uint32_t   DataParamSize;
+};
+
+// Update CURRENT_ICBE_VERSION when modifying the patch list
+static_assert( sizeof( SPatchAllocateSyncBuffer ) == ( 12 + sizeof( SPatchItemHeader ) ), "The size of SPatchAllocateSyncBuffer is not what is expected" );
+
 /*****************************************************************************\
 STRUCT: SPatchAllocateStatelessPrivateSurface
 \*****************************************************************************/
diff --git a/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp b/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp
@@ -194,6 +194,13 @@ namespace IGC
             m_kernelInfo.m_printfBufferAnnotation = nullptr;
         }
 
+        // Sync Buffer Annotation
+        if (m_kernelInfo.m_syncBufferAnnotation != nullptr)
+        {
+            delete m_kernelInfo.m_syncBufferAnnotation;
+            m_kernelInfo.m_syncBufferAnnotation = nullptr;
+        }
+
         // StartGASAnnotationAnnotation
         if (m_kernelInfo.m_startGAS != nullptr)
         {
@@ -1147,6 +1154,23 @@ namespace IGC
             m_kernelInfo.m_threadPayload.UnusedPerThreadConstantPresent = true;
             break;
 
+        case KernelArg::ArgType::IMPLICIT_SYNC_BUFFER:
+        {
+            int argNo = kernelArg->getAssociatedArgNo();
+            SOpenCLKernelInfo::SResourceInfo resInfo = getResourceInfo(argNo);
+            m_kernelInfo.m_argIndexMap[argNo] = getBTI(resInfo);
+
+            iOpenCL::SyncBufferAnnotation* syncBuffer = new iOpenCL::SyncBufferAnnotation();
+
+            syncBuffer->AnnotationSize = sizeof(syncBuffer);
+            syncBuffer->ArgumentNumber = argNo;
+            syncBuffer->PayloadPosition = payloadPosition;
+            syncBuffer->DataSize = kernelArg->getAllocateSize();
+
+            m_kernelInfo.m_syncBufferAnnotation = syncBuffer;
+        }
+        break;
+
         case KernelArg::ArgType::IMPLICIT_PRINTF_BUFFER:
         {
             int argNo = kernelArg->getAssociatedArgNo();
@@ -1408,6 +1432,7 @@ namespace IGC
         m_kernelInfo.m_threadPayload.HasLocalID = false;
         m_kernelInfo.m_threadPayload.UnusedPerThreadConstantPresent = false;
         m_kernelInfo.m_printfBufferAnnotation = nullptr;
+        m_kernelInfo.m_syncBufferAnnotation = nullptr;
         m_kernelInfo.m_threadPayload.HasStageInGridOrigin = false;
         m_kernelInfo.m_threadPayload.HasStageInGridSize = false;
 
diff --git a/IGC/Compiler/CodeGenPublic.h b/IGC/Compiler/CodeGenPublic.h
@@ -513,6 +513,7 @@ namespace IGC
         SOpenCLKernelInfo()
         {
             m_printfBufferAnnotation = nullptr;
+            m_syncBufferAnnotation = nullptr;
             m_startGAS = nullptr;
             m_WindowSizeGAS = nullptr;
             m_PrivateMemSize = nullptr;
@@ -537,6 +538,7 @@ namespace IGC
         std::vector<iOpenCL::PrintfStringAnnotation*>       m_printfStringAnnotations;
 
         iOpenCL::PrintfBufferAnnotation* m_printfBufferAnnotation;
+        iOpenCL::SyncBufferAnnotation* m_syncBufferAnnotation;
         iOpenCL::StartGASAnnotation* m_startGAS = NULL;
         iOpenCL::WindowSizeGASAnnotation* m_WindowSizeGAS = NULL;
         iOpenCL::PrivateMemSizeAnnotation* m_PrivateMemSize = NULL;
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs.cpp
@@ -244,6 +244,8 @@ KernelArg::ArgType KernelArg::calcArgType(const ImplicitArg& arg) const
         return KernelArg::ArgType::IMPLICIT_CONSTANT_BASE;
     case ImplicitArg::PRINTF_BUFFER:
         return KernelArg::ArgType::IMPLICIT_PRINTF_BUFFER;
+    case ImplicitArg::SYNC_BUFFER:
+        return KernelArg::ArgType::IMPLICIT_SYNC_BUFFER;
     case ImplicitArg::BUFFER_OFFSET:
         return KernelArg::ArgType::IMPLICIT_BUFFER_OFFSET;
     case ImplicitArg::GLOBAL_BASE:
@@ -763,6 +765,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::IMPLICIT_GLOBAL_BASE,
             KernelArg::ArgType::IMPLICIT_PRIVATE_BASE,
             KernelArg::ArgType::IMPLICIT_PRINTF_BUFFER,
+            KernelArg::ArgType::IMPLICIT_SYNC_BUFFER,
             KernelArg::ArgType::IMPLICIT_BUFFER_OFFSET,
             KernelArg::ArgType::IMPLICIT_WORK_DIM,
             KernelArg::ArgType::IMPLICIT_NUM_GROUPS,
@@ -873,6 +876,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::IMPLICIT_GLOBAL_BASE,
             KernelArg::ArgType::IMPLICIT_PRIVATE_BASE,
             KernelArg::ArgType::IMPLICIT_PRINTF_BUFFER,
+            KernelArg::ArgType::IMPLICIT_SYNC_BUFFER,
             KernelArg::ArgType::IMPLICIT_BUFFER_OFFSET,
             KernelArg::ArgType::IMPLICIT_WORK_DIM,
             KernelArg::ArgType::IMPLICIT_NUM_GROUPS,
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs.hpp b/IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs.hpp
@@ -75,6 +75,7 @@ namespace IGC
             IMPLICIT_PRIVATE_BASE,
 
             IMPLICIT_PRINTF_BUFFER,
+            IMPLICIT_SYNC_BUFFER,
 
             IMPLICIT_BUFFER_OFFSET,
 
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/ResourceAllocator/ResourceAllocator.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/ResourceAllocator/ResourceAllocator.cpp
@@ -175,6 +175,7 @@ static AllocationType getAllocationType(KernelArg::ArgType argType, BindlessAllo
     case KernelArg::ArgType::IMPLICIT_GLOBAL_BASE:
     case KernelArg::ArgType::IMPLICIT_PRIVATE_BASE:
     case KernelArg::ArgType::IMPLICIT_PRINTF_BUFFER:
+    case KernelArg::ArgType::IMPLICIT_SYNC_BUFFER:
     case KernelArg::ArgType::IMPLICIT_DEVICE_ENQUEUE_EVENT_POOL:
     case KernelArg::ArgType::IMPLICIT_DEVICE_ENQUEUE_DEFAULT_DEVICE_QUEUE:
         return AllocationType::Other;
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp
@@ -158,6 +158,10 @@ void WIFuncResolution::visitCallInst(CallInst& CI)
     {
         wiRes = getStageInGridSize(CI);
     }
+    else if (funcName.equals(WIFuncsAnalysis::GET_SYNC_BUFFER))
+    {
+        wiRes = getSyncBufferPtr(CI);
+    }
     else
     {
         // Non WI function, do nothing
@@ -378,6 +382,19 @@ Value* WIFuncResolution::getStageInGridSize(CallInst& CI)
     return globalSize;
 }
 
+Value* WIFuncResolution::getSyncBufferPtr(CallInst& CI)
+{
+    // Receives:
+    // call i8 addrspace(1)* @__builtin_IB_get_sync_buffer()
+
+    // Creates:
+    // i8 addrspace(1)* %syncBuffer
+
+    Argument* syncBuffer = getImplicitArg(CI, ImplicitArg::SYNC_BUFFER);
+
+    return syncBuffer;
+}
+
 Argument* WIFuncResolution::getImplicitArg(CallInst& CI, ImplicitArg::ArgType argType)
 {
     unsigned int numImplicitArgs = m_implicitArgs.size();
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.hpp b/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.hpp
@@ -144,6 +144,12 @@ namespace IGC
         /// @return A value representing the stage in grid size
         llvm::Value* getStageInGridSize(llvm::CallInst& CI);
 
+        /// @brief  Resolves get_sync_buffer().
+        ///         Adds the appropriate sequence of code before the given call instruction
+        /// @param  CI The call instruction.
+        /// @return A value representing the sync buffer
+        llvm::Value* getSyncBufferPtr(llvm::CallInst& CI);
+
         /// @brief  get vector of work group size if reqd_work_group_size is set.
         /// @param  F the function to check
         /// @return A vector with work group size (e.g., <i32 16, i32 1, i32 1>)
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncsAnalysis.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncsAnalysis.cpp
@@ -60,6 +60,7 @@ const llvm::StringRef WIFuncsAnalysis::GET_NUM_GROUPS = "__builtin_IB_get_num_gr
 const llvm::StringRef WIFuncsAnalysis::GET_ENQUEUED_LOCAL_SIZE = "__builtin_IB_get_enqueued_local_size";
 const llvm::StringRef WIFuncsAnalysis::GET_STAGE_IN_GRID_ORIGIN = "__builtin_IB_get_stage_in_grid_origin";
 const llvm::StringRef WIFuncsAnalysis::GET_STAGE_IN_GRID_SIZE = "__builtin_IB_get_stage_in_grid_size";
+const llvm::StringRef WIFuncsAnalysis::GET_SYNC_BUFFER = "__builtin_IB_get_sync_buffer";
 
 WIFuncsAnalysis::WIFuncsAnalysis() : ModulePass(ID)
 {
@@ -90,6 +91,7 @@ bool WIFuncsAnalysis::runOnFunction(Function& F)
     m_hasEnqueuedLocalSize = false;
     m_hasStageInGridOrigin = false;
     m_hasStageInGridSize = false;
+    m_hasSyncBuffer = false;
 
     // Visit the function
     visit(F);
@@ -136,6 +138,10 @@ bool WIFuncsAnalysis::runOnFunction(Function& F)
     {
         implicitArgs.push_back(ImplicitArg::STAGE_IN_GRID_SIZE);
     }
+    if (m_hasSyncBuffer)
+    {
+        implicitArgs.push_back(ImplicitArg::SYNC_BUFFER);
+    }
 
     // Create the metadata representing the implicit args needed by this function
     ImplicitArgs::addImplicitArgs(F, implicitArgs, getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils());
@@ -183,4 +189,7 @@ void WIFuncsAnalysis::visitCallInst(CallInst& CI)
     else if (funcName.equals(GET_STAGE_IN_GRID_SIZE)) {
         m_hasStageInGridSize = true;
     }
+    else if (funcName.equals(GET_SYNC_BUFFER)) {
+        m_hasSyncBuffer = true;
+    }
 }
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncsAnalysis.hpp b/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncsAnalysis.hpp
@@ -86,6 +86,7 @@ namespace IGC
         static const llvm::StringRef GET_ENQUEUED_LOCAL_SIZE;
         static const llvm::StringRef GET_STAGE_IN_GRID_ORIGIN;
         static const llvm::StringRef GET_STAGE_IN_GRID_SIZE;
+        static const llvm::StringRef GET_SYNC_BUFFER;
 
     private:
         /// @brief  Function entry point.
@@ -112,6 +113,8 @@ namespace IGC
         bool m_hasStageInGridOrigin;
         /// @brief  Marks whether stage_in_grid_size is needed by the current function
         bool m_hasStageInGridSize;
+        /// @brief  Marks whether sync buffer is needed by the current function
+        bool m_hasSyncBuffer;
     };
 
 } // namespace IGC
