DeviceContextD3D12Impl: reworked flush to also execute deferred command lists to avoid redundant submit

2 files changed, 44 insertions, 39 deletions

diff --git a/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp b/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
index 15c85492..39ac2431 100644
--- a/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
+++ b/Graphics/GraphicsEngineD3D12/include/DeviceContextD3D12Impl.hpp
@@ -346,7 +346,9 @@ private:
     void CommitScissorRects(class GraphicsContext& GraphCtx, bool ScissorEnable);
     void TransitionSubpassAttachments(Uint32 NextSubpass);
     void CommitSubpassRenderTargets();
-    void Flush(bool RequestNewCmdCtx);
+    void Flush(bool                 RequestNewCmdCtx,
+               Uint32               NumCommandLists = 0,
+               ICommandList* const* ppCommandLists  = nullptr);
 
     __forceinline void RequestCommandContext(RenderDeviceD3D12Impl* pDeviceD3D12Impl);
 
diff --git a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
index b7415753..1cd2f766 100644
--- a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
@@ -773,26 +773,52 @@ void DeviceContextD3D12Impl::RequestCommandContext(RenderDeviceD3D12Impl* pDevic
     m_CurrCmdCtx->SetDynamicGPUDescriptorAllocators(m_DynamicGPUDescriptorAllocator);
 }
 
-void DeviceContextD3D12Impl::Flush(bool RequestNewCmdCtx)
+void DeviceContextD3D12Impl::Flush(bool                 RequestNewCmdCtx,
+                                   Uint32               NumCommandLists,
+                                   ICommandList* const* ppCommandLists)
 {
+    VERIFY(!m_bIsDeferred || NumCommandLists == 0 && ppCommandLists == nullptr, "Only immediate context can execute command lists");
+
+    if (m_ActiveQueriesCounter > 0)
+    {
+        LOG_ERROR_MESSAGE("Flushing device context that has ", m_ActiveQueriesCounter,
+                          " active queries. Direct3D12 requires that queries are begun and ended in the same command list");
+    }
+
+    // TODO: use small_vector
+    std::vector<RenderDeviceD3D12Impl::PooledCommandContext> Contexts;
+    Contexts.reserve(NumCommandLists + 1);
+
+    // First, execute current context
     if (m_CurrCmdCtx)
     {
         VERIFY(!m_bIsDeferred, "Deferred contexts cannot execute command lists directly");
         if (m_State.NumCommands != 0)
-        {
-            m_CurrCmdCtx->FlushResourceBarriers();
-            m_pDevice->CloseAndExecuteCommandContexts(m_CommandQueueId, 1, &m_CurrCmdCtx, true, &m_PendingFences);
-            VERIFY(!m_CurrCmdCtx, "Context must be disposed by CloseAndExecuteCommandContexts");
-            m_PendingFences.clear();
-        }
+            Contexts.emplace_back(std::move(m_CurrCmdCtx));
         else
             m_pDevice->DisposeCommandContext(std::move(m_CurrCmdCtx));
     }
 
-    if (m_ActiveQueriesCounter > 0)
+    // Next, add extra command lists from deferred contexts
+    for (Uint32 i = 0; i < NumCommandLists; ++i)
     {
-        LOG_ERROR_MESSAGE("Flushing device context that has ", m_ActiveQueriesCounter,
-                          " active queries. Direct3D12 requires that queries are begun and ended in the same command list");
+        auto* const pCmdListD3D12 = ValidatedCast<CommandListD3D12Impl>(ppCommandLists[i]);
+
+        RefCntAutoPtr<DeviceContextD3D12Impl> pDeferredCtx;
+        Contexts.emplace_back(pCmdListD3D12->Close(pDeferredCtx));
+        // Set the bit in the deferred context cmd queue mask corresponding to the cmd queue of this context
+        pDeferredCtx->m_SubmittedBuffersCmdQueueMask.fetch_or(Uint64{1} << m_CommandQueueId);
+    }
+
+    if (!Contexts.empty())
+    {
+        m_pDevice->CloseAndExecuteCommandContexts(m_CommandQueueId, static_cast<Uint32>(Contexts.size()), Contexts.data(), true, &m_PendingFences);
+        m_PendingFences.clear();
+
+#ifdef DILIGENT_DEBUG
+        for (Uint32 i = 0; i < NumCommandLists; ++i)
+            VERIFY(!Contexts[i], "All contexts must be disposed by CloseAndExecuteCommandContexts");
+#endif
     }
 
     // If there is no command list to submit, but there are pending fences, we need to signal them now
@@ -2036,7 +2062,7 @@ void DeviceContextD3D12Impl::GenerateMips(ITextureView* pTexView)
 
 void DeviceContextD3D12Impl::FinishCommandList(ICommandList** ppCommandList)
 {
-    VERIFY(m_pActiveRenderPass == nullptr, "Finishing command list inside an active render pass.");
+    DEV_CHECK_ERR(m_pActiveRenderPass == nullptr, "Finishing command list inside an active render pass.");
 
     CommandListD3D12Impl* pCmdListD3D12(NEW_RC_OBJ(m_CmdListAllocator, "CommandListD3D12Impl instance", CommandListD3D12Impl)(m_pDevice, this, std::move(m_CurrCmdCtx)));
     pCmdListD3D12->QueryInterface(IID_CommandList, reinterpret_cast<IObject**>(ppCommandList));
@@ -2058,43 +2084,20 @@ void DeviceContextD3D12Impl::ExecuteCommandLists(Uint32               NumCommand
         return;
     DEV_CHECK_ERR(ppCommandLists != nullptr, "ppCommandLists must not be null when NumCommandLists is not zero");
 
-    // First execute commands in this context
-    Flush(true);
+    Flush(true, NumCommandLists, ppCommandLists);
 
     InvalidateState();
-
-    VERIFY_EXPR(m_PendingFences.empty());
-
-    // TODO: use small_vector
-    std::vector<RenderDeviceD3D12Impl::PooledCommandContext> Contexts;
-    Contexts.reserve(NumCommandLists);
-    for (Uint32 i = 0; i < NumCommandLists; ++i)
-    {
-        auto* const pCmdListD3D12 = ValidatedCast<CommandListD3D12Impl>(ppCommandLists[i]);
-
-        RefCntAutoPtr<DeviceContextD3D12Impl> pDeferredCtx;
-        Contexts.emplace_back(pCmdListD3D12->Close(pDeferredCtx));
-        // Set the bit in the deferred context cmd queue mask corresponding to cmd queue of this context
-        pDeferredCtx->m_SubmittedBuffersCmdQueueMask |= Uint64{1} << m_CommandQueueId;
-    }
-
-    m_pDevice->CloseAndExecuteCommandContexts(m_CommandQueueId, NumCommandLists, Contexts.data(), true, nullptr);
-
-#ifdef DILIGENT_DEBUG
-    for (Uint32 i = 0; i < NumCommandLists; ++i)
-        VERIFY(!Contexts[i], "All contexts must be disposed by CloseAndExecuteCommandContexts");
-#endif
 }
 
 void DeviceContextD3D12Impl::SignalFence(IFence* pFence, Uint64 Value)
 {
-    VERIFY(!m_bIsDeferred, "Fence can only be signaled from immediate context");
+    DEV_CHECK_ERR(!m_bIsDeferred, "Fence can only be signaled from immediate context");
     m_PendingFences.emplace_back(Value, pFence);
 }
 
 void DeviceContextD3D12Impl::WaitForFence(IFence* pFence, Uint64 Value, bool FlushContext)
 {
-    VERIFY(!m_bIsDeferred, "Fence can only be waited from immediate context");
+    DEV_CHECK_ERR(!m_bIsDeferred, "Fence can only be waited from immediate context");
     if (FlushContext)
         Flush();
     auto* pFenceD3D12 = ValidatedCast<FenceD3D12Impl>(pFence);
@@ -2103,7 +2106,7 @@ void DeviceContextD3D12Impl::WaitForFence(IFence* pFence, Uint64 Value, bool Flu
 
 void DeviceContextD3D12Impl::WaitForIdle()
 {
-    VERIFY(!m_bIsDeferred, "Only immediate contexts can be idled");
+    DEV_CHECK_ERR(!m_bIsDeferred, "Only immediate contexts can be idled");
     Flush();
     m_pDevice->IdleCommandQueue(m_CommandQueueId, true);
 }