optimize root view updates

4 files changed, 70 insertions, 67 deletions

diff --git a/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp b/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
index 5ed3a1dc..333dc912 100644
--- a/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
+++ b/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
@@ -391,27 +391,30 @@ private:
         using Bitfield = Uint8;
         static_assert(sizeof(Bitfield) * 8 >= MAX_RESOURCE_SIGNATURES, "not enought space to store MAX_RESOURCE_SIGNATURES bits");
 
-        //Bitfield             ActiveSRBMask      ;     // Indicates which SRBs are active in current PSO
-        bool                 bRootViewsCommitted; // Indicates if root views have been committed since the time SRB  has been committed.
+        Bitfield             ActiveSRBMask      = 0; // Indicates which SRBs are active in current PSO
+        Bitfield             DynamicBuffersMask = 0; // Indicates which SRBs have dynamic buffers
+        bool                 bRootViewsCommitted;    // Indicates if root views have been committed since the time SRB  has been committed.
         bool                 bRootTablesCommited;
-        bool                 IsCompute : 1;
         ID3D12RootSignature* pRootSig;
 
         std::array<class ShaderResourceBindingD3D12Impl*, MAX_RESOURCE_SIGNATURES> SRBs;
 
-        RootTableInfo(bool _IsCompute)
+        RootTableInfo()
         {
             memset(this, 0, sizeof(*this));
-            IsCompute = _IsCompute;
         }
 
         __forceinline bool RequireUpdate(bool DynamicBuffersIntact = false) const
         {
-            return true; //(StaleSRBMask & ActiveSRBMask) != 0 || ((DynamicBuffersMask & ActiveSRBMask) != 0 && !DynamicBuffersIntact);
+            return !bRootViewsCommitted || !bRootTablesCommited || ((DynamicBuffersMask & ActiveSRBMask) != 0 && !DynamicBuffersIntact);
         }
+
+        void SetDynamicBufferBit(Uint32 Index) { DynamicBuffersMask |= static_cast<Bitfield>(1u << Index); }
+        void ClearDynamicBufferBit(Uint32 Index) { DynamicBuffersMask &= static_cast<Bitfield>(~(1u << Index)); }
     };
     __forceinline RootTableInfo& GetRootTableInfo(PIPELINE_TYPE PipelineType);
 
+    template <bool IsCompute>
     __forceinline void CommitRootTables(RootTableInfo& RootInfo);
 #ifdef DILIGENT_DEVELOPMENT
     void DvpValidateCommittedShaderResources();
diff --git a/Graphics/GraphicsEngineD3D12/include/PipelineResourceSignatureD3D12Impl.hpp b/Graphics/GraphicsEngineD3D12/include/PipelineResourceSignatureD3D12Impl.hpp
index 8fd7ee7d..c747dda2 100644
--- a/Graphics/GraphicsEngineD3D12/include/PipelineResourceSignatureD3D12Impl.hpp
+++ b/Graphics/GraphicsEngineD3D12/include/PipelineResourceSignatureD3D12Impl.hpp
@@ -273,11 +273,11 @@ public:
                           bool                      IsCompute,
                           Uint32                    FirstRootIndex);
 
+    template <bool IsCompute>
     void CommitRootViews(ShaderResourceCacheD3D12& ResourceCache,
                          CommandContext&           Ctx,
                          DeviceContextD3D12Impl*   pDeviceCtx,
                          Uint32                    DeviceCtxId,
-                         bool                      IsCompute,
                          Uint32                    FirstRootIndex);
 
 private:
@@ -307,4 +307,32 @@ private:
     SRBMemoryAllocator m_SRBMemAllocator;
 };
 
+template <bool IsCompute>
+__forceinline void PipelineResourceSignatureD3D12Impl::CommitRootViews(ShaderResourceCacheD3D12& ResourceCache,
+                                                                       CommandContext&           CmdCtx,
+                                                                       DeviceContextD3D12Impl*   pDeviceCtx,
+                                                                       Uint32                    DeviceCtxId,
+                                                                       Uint32                    FirstRootIndex)
+{
+    for (Uint32 rv = 0; rv < m_RootParams.GetNumRootViews(); ++rv)
+    {
+        auto& RootView = m_RootParams.GetRootView(rv);
+        auto  RootInd  = RootView.RootIndex;
+
+        auto& Res = ResourceCache.GetRootTable(RootInd).GetResource(0, D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV);
+        if (auto* pBuffToTransition = Res.pObject.RawPtr<BufferD3D12Impl>())
+        {
+            bool IsDynamic = pBuffToTransition->GetDesc().Usage == USAGE_DYNAMIC;
+            if (IsDynamic)
+            {
+                D3D12_GPU_VIRTUAL_ADDRESS CBVAddress = pBuffToTransition->GetGPUAddress(DeviceCtxId, pDeviceCtx);
+                if (IsCompute)
+                    CmdCtx.GetCommandList()->SetComputeRootConstantBufferView(FirstRootIndex + RootInd, CBVAddress);
+                else
+                    CmdCtx.GetCommandList()->SetGraphicsRootConstantBufferView(FirstRootIndex + RootInd, CBVAddress);
+            }
+        }
+    }
+}
+
 } // namespace Diligent
diff --git a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
index 58e3262f..db878dd9 100644
--- a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
@@ -75,8 +75,6 @@ DeviceContextD3D12Impl::DeviceContextD3D12Impl(IReferenceCounters*          pRef
         bIsDeferred ? std::numeric_limits<decltype(m_NumCommandsToFlush)>::max() : EngineCI.NumCommandsToFlushCmdList,
         bIsDeferred
     },
-    m_GraphicsResources{false},
-    m_ComputeResources {true },
     m_DynamicHeap
     {
         pDeviceD3D12Impl->GetDynamicMemoryManager(),
@@ -236,6 +234,14 @@ void DeviceContextD3D12Impl::SetPipelineState(IPipelineState* pPipelineState)
         RootInfo.pRootSig            = pd3d12RootSig;
         RootInfo.bRootTablesCommited = false;
         RootInfo.bRootViewsCommitted = false;
+        RootInfo.ActiveSRBMask       = 0;
+
+        for (Uint32 i = 0, SignCount = pPipelineStateD3D12->GetSignatureCount(); i < SignCount; ++i)
+        {
+            auto* pSignature = pPipelineStateD3D12->GetSignature(i);
+            if (pSignature != nullptr)
+                RootInfo.ActiveSRBMask |= 1u << i;
+        }
     }
 
     switch (PSODesc.PipelineType)
@@ -293,6 +299,7 @@ void DeviceContextD3D12Impl::SetPipelineState(IPipelineState* pPipelineState)
     }
 }
 
+template <bool IsCompute>
 void DeviceContextD3D12Impl::CommitRootTables(RootTableInfo& RootInfo)
 {
     const auto& RootSig = *m_pPipelineState->GetRootSignature();
@@ -301,7 +308,6 @@ void DeviceContextD3D12Impl::CommitRootTables(RootTableInfo& RootInfo)
     if (!RootInfo.bRootTablesCommited)
     {
         RootInfo.bRootTablesCommited = true;
-
         for (Uint32 s = 0; s < RootSig.GetSignatureCount(); ++s)
         {
             auto* pSignature = RootSig.GetSignature(s);
@@ -311,25 +317,21 @@ void DeviceContextD3D12Impl::CommitRootTables(RootTableInfo& RootInfo)
                 continue;
 
             VERIFY_EXPR(pSRB != nullptr);
-            pSignature->CommitRootTables(pSRB->GetResourceCache(), CmdCtx, this, GetContextId(), RootInfo.IsCompute, RootSig.GetFirstRootIndex(s));
+            pSignature->CommitRootTables(pSRB->GetResourceCache(), CmdCtx, this, GetContextId(), IsCompute, RootSig.GetFirstRootIndex(s));
         }
     }
 
-    //if (!RootInfo.bRootViewsCommitted)
+    RootInfo.bRootViewsCommitted = true;
+    for (Uint32 s = 0; s < RootSig.GetSignatureCount(); ++s)
     {
-        RootInfo.bRootViewsCommitted = true;
-
-        for (Uint32 s = 0; s < RootSig.GetSignatureCount(); ++s)
-        {
-            auto* pSignature = RootSig.GetSignature(s);
-            auto* pSRB       = RootInfo.SRBs[s];
+        auto* pSignature = RootSig.GetSignature(s);
+        auto* pSRB       = RootInfo.SRBs[s];
 
-            if (pSignature == nullptr || pSignature->GetNumRootViews() == 0)
-                continue;
+        if (pSignature == nullptr || pSignature->GetNumRootViews() == 0)
+            continue;
 
-            VERIFY_EXPR(pSRB != nullptr);
-            pSignature->CommitRootViews(pSRB->GetResourceCache(), CmdCtx, this, GetContextId(), RootInfo.IsCompute, RootSig.GetFirstRootIndex(s));
-        }
+        VERIFY_EXPR(pSRB != nullptr);
+        pSignature->CommitRootViews<IsCompute>(pSRB->GetResourceCache(), CmdCtx, this, GetContextId(), RootSig.GetFirstRootIndex(s));
     }
 }
 
@@ -359,12 +361,6 @@ void DeviceContextD3D12Impl::CommitShaderResources(IShaderResourceBinding* pShad
     auto& CmdCtx               = GetCmdContext();
     auto* pSignature           = pResBindingD3D12Impl->GetSignature();
 
-    //if (pSignature->GetTotalRootCount() == 0)
-    //{
-    // Ignore SRBs that contain no resources
-    //    return;
-    //}
-
 #ifdef DILIGENT_DEBUG
     //ResourceCache.DbgVerifyDynamicBuffersCounter();
 #endif
@@ -380,9 +376,15 @@ void DeviceContextD3D12Impl::CommitShaderResources(IShaderResourceBinding* pShad
     }
 #endif
 
-    auto& RootInfo = GetRootTableInfo(pSignature->GetPipelineType());
+    const auto SRBIndex = pResBindingD3D12Impl->GetBindingIndex();
+    auto&      RootInfo = GetRootTableInfo(pSignature->GetPipelineType());
 
-    RootInfo.SRBs[pSignature->GetDesc().BindingIndex] = pResBindingD3D12Impl;
+    RootInfo.SRBs[SRBIndex] = pResBindingD3D12Impl;
+
+    if (ResourceCache.GetNumDynamicCBsBound() > 0)
+        RootInfo.SetDynamicBufferBit(SRBIndex);
+    else
+        RootInfo.ClearDynamicBufferBit(SRBIndex);
 
     RootInfo.bRootTablesCommited = false;
     RootInfo.bRootViewsCommitted = false;
@@ -581,7 +583,7 @@ void DeviceContextD3D12Impl::PrepareForDraw(GraphicsContext& GraphCtx, DRAW_FLAG
     auto& RootInfo = GetRootTableInfo(PIPELINE_TYPE_GRAPHICS);
     if (RootInfo.RequireUpdate(Flags & DRAW_FLAG_DYNAMIC_RESOURCE_BUFFERS_INTACT))
     {
-        CommitRootTables(RootInfo);
+        CommitRootTables<false>(RootInfo);
     }
 
 #ifdef DILIGENT_DEVELOPMENT
@@ -714,7 +716,7 @@ void DeviceContextD3D12Impl::PrepareForDispatchCompute(ComputeContext& ComputeCt
     auto& RootInfo = GetRootTableInfo(PIPELINE_TYPE_COMPUTE);
     if (RootInfo.RequireUpdate())
     {
-        CommitRootTables(RootInfo);
+        CommitRootTables<true>(RootInfo);
     }
 
 #ifdef DILIGENT_DEVELOPMENT
@@ -727,7 +729,7 @@ void DeviceContextD3D12Impl::PrepareForDispatchRays(GraphicsContext& GraphCtx)
     auto& RootInfo = GetRootTableInfo(PIPELINE_TYPE_RAY_TRACING);
     if (RootInfo.RequireUpdate())
     {
-        CommitRootTables(RootInfo);
+        CommitRootTables<true>(RootInfo);
     }
 
 #ifdef DILIGENT_DEVELOPMENT
@@ -894,10 +896,9 @@ void DeviceContextD3D12Impl::Flush(bool                 RequestNewCmdCtx,
     if (RequestNewCmdCtx)
         RequestCommandContext(m_pDevice);
 
-    m_State = State{};
-
-    m_GraphicsResources = RootTableInfo{false};
-    m_ComputeResources  = RootTableInfo{true};
+    m_State             = State{};
+    m_GraphicsResources = RootTableInfo{};
+    m_ComputeResources  = RootTableInfo{};
 
     // Setting pipeline state to null makes sure that render targets and other
     // states will be restored in the command list next time a PSO is bound.
diff --git a/Graphics/GraphicsEngineD3D12/src/PipelineResourceSignatureD3D12Impl.cpp b/Graphics/GraphicsEngineD3D12/src/PipelineResourceSignatureD3D12Impl.cpp
index 7cab8a5d..243cf9c0 100644
--- a/Graphics/GraphicsEngineD3D12/src/PipelineResourceSignatureD3D12Impl.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/PipelineResourceSignatureD3D12Impl.cpp
@@ -1060,7 +1060,6 @@ void PipelineResourceSignatureD3D12Impl::CommitRootTables(ShaderResourceCacheD3D
 
     Uint32 NumDynamicCbvSrvUavDescriptors = m_RootParams.GetTotalSrvCbvUavSlots(ROOT_PARAMETER_GROUP_DYNAMIC);
     Uint32 NumDynamicSamplerDescriptors   = m_RootParams.GetTotalSamplerSlots(ROOT_PARAMETER_GROUP_DYNAMIC);
-    //VERIFY_EXPR(NumDynamicCbvSrvUavDescriptors > 0 || NumDynamicSamplerDescriptors > 0);
 
     DescriptorHeapAllocation DynamicCbvSrvUavDescriptors, DynamicSamplerDescriptors;
     if (NumDynamicCbvSrvUavDescriptors > 0)
@@ -1204,34 +1203,6 @@ void PipelineResourceSignatureD3D12Impl::CommitRootTables(ShaderResourceCacheD3D
     }
 }
 
-void PipelineResourceSignatureD3D12Impl::CommitRootViews(ShaderResourceCacheD3D12& ResourceCache,
-                                                         CommandContext&           CmdCtx,
-                                                         DeviceContextD3D12Impl*   pDeviceCtx,
-                                                         Uint32                    DeviceCtxId,
-                                                         bool                      IsCompute,
-                                                         Uint32                    FirstRootIndex)
-{
-    for (Uint32 rv = 0; rv < m_RootParams.GetNumRootViews(); ++rv)
-    {
-        auto& RootView = m_RootParams.GetRootView(rv);
-        auto  RootInd  = RootView.RootIndex;
-
-        auto& Res = ResourceCache.GetRootTable(RootInd).GetResource(0, D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV);
-        if (auto* pBuffToTransition = Res.pObject.RawPtr<BufferD3D12Impl>())
-        {
-            bool IsDynamic = pBuffToTransition->GetDesc().Usage == USAGE_DYNAMIC;
-            if (IsDynamic)
-            {
-                D3D12_GPU_VIRTUAL_ADDRESS CBVAddress = pBuffToTransition->GetGPUAddress(DeviceCtxId, pDeviceCtx);
-                if (IsCompute)
-                    CmdCtx.GetCommandList()->SetComputeRootConstantBufferView(FirstRootIndex + RootInd, CBVAddress);
-                else
-                    CmdCtx.GetCommandList()->SetGraphicsRootConstantBufferView(FirstRootIndex + RootInd, CBVAddress);
-            }
-        }
-    }
-}
-
 
 namespace
 {