Reworked ExecuteCommandList(s) to take multiple command lists instead of one

4 files changed, 52 insertions, 22 deletions

diff --git a/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp b/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
index d9a2c2f2..a385f8aa 100644
--- a/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
+++ b/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
@@ -233,8 +233,9 @@ public:
     /// Implementation of IDeviceContext::FinishCommandList() in Vulkan backend.
     virtual void DILIGENT_CALL_TYPE FinishCommandList(class ICommandList** ppCommandList) override final;
 
-    /// Implementation of IDeviceContext::ExecuteCommandList() in Vulkan backend.
-    virtual void DILIGENT_CALL_TYPE ExecuteCommandList(class ICommandList* pCommandList) override final;
+    /// Implementation of IDeviceContext::ExecuteCommandLists() in Vulkan backend.
+    virtual void DILIGENT_CALL_TYPE ExecuteCommandLists(Uint32               NumCommandLists,
+                                                        ICommandList* const* ppCommandLists) override final;
 
     /// Implementation of IDeviceContext::SignalFence() in Vulkan backend.
     virtual void DILIGENT_CALL_TYPE SignalFence(IFence* pFence, Uint64 Value) override final;
diff --git a/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp b/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
index 37052725..60b8ebf1 100644
--- a/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
+++ b/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
@@ -230,12 +230,18 @@ private:
 
     virtual void TestTextureFormat(TEXTURE_FORMAT TexFormat) override final;
 
-    // Submits command buffer for execution to the command queue
-    // Returns the submitted command buffer number and the fence value
+    // Submits command buffer(s) for execution to the command queue and
+    // returns the submitted command buffer(s) number and the fence value.
+    // If SubmitInfo contains multiple command buffers, they all are treated
+    // like one and sumbitted atomically.
     // Parameters:
     //      * SubmittedCmdBuffNumber - submitted command buffer number
     //      * SubmittedFenceValue    - fence value associated with the submitted command buffer
-    void SubmitCommandBuffer(Uint32 QueueIndex, const VkSubmitInfo& SubmitInfo, Uint64& SubmittedCmdBuffNumber, Uint64& SubmittedFenceValue, std::vector<std::pair<Uint64, RefCntAutoPtr<IFence>>>* pFences);
+    void SubmitCommandBuffer(Uint32                                                 QueueIndex,
+                             const VkSubmitInfo&                                    SubmitInfo,
+                             Uint64&                                                SubmittedCmdBuffNumber,
+                             Uint64&                                                SubmittedFenceValue,
+                             std::vector<std::pair<Uint64, RefCntAutoPtr<IFence>>>* pFences);
 
     std::shared_ptr<VulkanUtilities::VulkanInstance>       m_VulkanInstance;
     std::unique_ptr<VulkanUtilities::VulkanPhysicalDevice> m_PhysicalDevice;
diff --git a/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
index 8516a174..3259e07e 100644
--- a/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
@@ -26,7 +26,10 @@
  */
 
 #include "pch.h"
+
 #include <sstream>
+#include <vector>
+
 #include "RenderDeviceVkImpl.hpp"
 #include "DeviceContextVkImpl.hpp"
 #include "PipelineStateVkImpl.hpp"
@@ -2098,7 +2101,8 @@ void DeviceContextVkImpl::FinishCommandList(class ICommandList** ppCommandList)
     InvalidateState();
 }
 
-void DeviceContextVkImpl::ExecuteCommandList(class ICommandList* pCommandList)
+void DeviceContextVkImpl::ExecuteCommandLists(Uint32               NumCommandLists,
+                                              ICommandList* const* ppCommandLists)
 {
     if (m_bIsDeferred)
     {
@@ -2106,31 +2110,50 @@ void DeviceContextVkImpl::ExecuteCommandList(class ICommandList* pCommandList)
         return;
     }
 
+    if (NumCommandLists == 0)
+        return;
+    DEV_CHECK_ERR(ppCommandLists != nullptr, "ppCommandLists must not be null when NumCommandLists is not zero");
+
     Flush();
 
     InvalidateState();
 
-    CommandListVkImpl* pCmdListVk = ValidatedCast<CommandListVkImpl>(pCommandList);
-    VkCommandBuffer    vkCmdBuff  = VK_NULL_HANDLE;
+    // TODO: replace with small_vector
+    std::vector<VkCommandBuffer>               vkCmdBuffs;
+    std::vector<RefCntAutoPtr<IDeviceContext>> DeferredCtxs;
+    vkCmdBuffs.reserve(NumCommandLists);
+    DeferredCtxs.reserve(NumCommandLists);
+    for (Uint32 i = 0; i < NumCommandLists; ++i)
+    {
+        auto* pCmdListVk = ValidatedCast<CommandListVkImpl>(ppCommandLists[i]);
+        DEV_CHECK_ERR(pCmdListVk != nullptr, "Command list must not be null");
+        VkCommandBuffer               vkCmdBuff = VK_NULL_HANDLE;
+        RefCntAutoPtr<IDeviceContext> pDeferredCtx;
+        pCmdListVk->Close(vkCmdBuff, pDeferredCtx);
+        VERIFY(vkCmdBuff != VK_NULL_HANDLE, "Trying to execute empty command buffer");
+        VERIFY_EXPR(pDeferredCtx);
+        vkCmdBuffs.emplace_back(vkCmdBuff);
+        DeferredCtxs.emplace_back(std::move(pDeferredCtx));
+    }
 
-    RefCntAutoPtr<IDeviceContext> pDeferredCtx;
-    pCmdListVk->Close(vkCmdBuff, pDeferredCtx);
-    VERIFY(vkCmdBuff != VK_NULL_HANDLE, "Trying to execute empty command buffer");
-    VERIFY_EXPR(pDeferredCtx);
     VkSubmitInfo SubmitInfo = {};
 
     SubmitInfo.sType              = VK_STRUCTURE_TYPE_SUBMIT_INFO;
     SubmitInfo.pNext              = nullptr;
-    SubmitInfo.commandBufferCount = 1;
-    SubmitInfo.pCommandBuffers    = &vkCmdBuff;
+    SubmitInfo.commandBufferCount = NumCommandLists;
+    SubmitInfo.pCommandBuffers    = vkCmdBuffs.data();
     VERIFY_EXPR(m_PendingFences.empty());
-    auto pDeferredCtxVkImpl  = pDeferredCtx.RawPtr<DeviceContextVkImpl>();
     auto SubmittedFenceValue = m_pDevice->ExecuteCommandBuffer(m_CommandQueueId, SubmitInfo, this, nullptr);
-    // Set the bit in the deferred context cmd queue mask corresponding to cmd queue of this context
-    pDeferredCtxVkImpl->m_SubmittedBuffersCmdQueueMask |= Uint64{1} << m_CommandQueueId;
-    // It is OK to dispose command buffer from another thread. We are not going to
-    // record any commands and only need to add the buffer to the queue
-    pDeferredCtxVkImpl->DisposeVkCmdBuffer(m_CommandQueueId, vkCmdBuff, SubmittedFenceValue);
+
+    for (Uint32 i = 0; i < NumCommandLists; ++i)
+    {
+        auto pDeferredCtxVkImpl = DeferredCtxs[i].RawPtr<DeviceContextVkImpl>();
+        // Set the bit in the deferred context cmd queue mask corresponding to cmd queue of this context
+        pDeferredCtxVkImpl->m_SubmittedBuffersCmdQueueMask |= Uint64{1} << m_CommandQueueId;
+        // It is OK to dispose command buffer from another thread. We are not going to
+        // record any commands and only need to add the buffer to the queue
+        pDeferredCtxVkImpl->DisposeVkCmdBuffer(m_CommandQueueId, std::move(vkCmdBuffs[i]), SubmittedFenceValue);
+    }
 }
 
 void DeviceContextVkImpl::SignalFence(IFence* pFence, Uint64 Value)
diff --git a/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
index 90c1a88e..b09d8068 100644
--- a/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
@@ -399,7 +399,7 @@ void RenderDeviceVkImpl::SubmitCommandBuffer(Uint32
 )
 {
     // Submit the command list to the queue
-    auto CmbBuffInfo       = TRenderDeviceBase::SubmitCommandBuffer(QueueIndex, SubmitInfo, true);
+    auto CmbBuffInfo       = TRenderDeviceBase::SubmitCommandBuffer(QueueIndex, true, SubmitInfo);
     SubmittedFenceValue    = CmbBuffInfo.FenceValue;
     SubmittedCmdBuffNumber = CmbBuffInfo.CmdBufferNumber;
     if (pFences != nullptr)
@@ -443,7 +443,7 @@ void RenderDeviceVkImpl::FlushStaleResources(Uint32 CmdQueueIndex)
     // Submit empty command buffer to the queue. This will effectively signal the fence and
     // discard all resources
     VkSubmitInfo DummySumbitInfo = {};
-    TRenderDeviceBase::SubmitCommandBuffer(0, DummySumbitInfo, true);
+    TRenderDeviceBase::SubmitCommandBuffer(0, true, DummySumbitInfo);
 }
 
 void RenderDeviceVkImpl::ReleaseStaleResources(bool ForceRelease)