Merged master

9 files changed, 133 insertions, 66 deletions

diff --git a/Graphics/GraphicsEngineVulkan/CMakeLists.txt b/Graphics/GraphicsEngineVulkan/CMakeLists.txt
index 79e61acd..ff540add 100644
--- a/Graphics/GraphicsEngineVulkan/CMakeLists.txt
+++ b/Graphics/GraphicsEngineVulkan/CMakeLists.txt
@@ -266,12 +266,15 @@ if(PLATFORM_WIN32)
 elseif(PLATFORM_MACOS)
     # Use Volk
 elseif(PLATFORM_IOS)
-    # For the apparent lack of a better way, link with both arm64 and x64 libraries.
-    # The linker will issue a warning, but this will work for both device and the simulator.
-    list(APPEND PRIVATE_DEPENDENCIES
-        "${MoltenVK_FRAMEWORK}/ios-arm64/libMoltenVK.a"
-        "${MoltenVK_FRAMEWORK}/ios-x86_64-simulator/libMoltenVK.a"
-    )
+    if(CMAKE_OSX_SYSROOT STREQUAL "iphonesimulator" OR PLATFORM_IOS_SIMULATOR)
+        list(APPEND PRIVATE_DEPENDENCIES
+            "${MoltenVK_FRAMEWORK}/ios-arm64_x86_64-simulator/libMoltenVK.a"
+        )
+    else()
+        list(APPEND PRIVATE_DEPENDENCIES
+            "${MoltenVK_FRAMEWORK}/ios-arm64/libMoltenVK.a"
+        )
+    endif()
 
     find_library(CORE_GRAPHICS CoreGraphics)
     find_library(METAL_LIBRARY Metal)
diff --git a/Graphics/GraphicsEngineVulkan/include/CommandListVkImpl.hpp b/Graphics/GraphicsEngineVulkan/include/CommandListVkImpl.hpp
index 3b363e30..63707812 100644
--- a/Graphics/GraphicsEngineVulkan/include/CommandListVkImpl.hpp
+++ b/Graphics/GraphicsEngineVulkan/include/CommandListVkImpl.hpp
@@ -60,12 +60,12 @@ public:
         VERIFY(m_vkCmdBuff == VK_NULL_HANDLE && !m_pDeferredCtx, "Destroying command list that was never executed");
     }
 
-    void Close(VkCommandBuffer&               CmdBuff,
-               RefCntAutoPtr<IDeviceContext>& pDeferredCtx)
+    VkCommandBuffer Close(RefCntAutoPtr<IDeviceContext>& pDeferredCtx)
     {
-        CmdBuff      = m_vkCmdBuff;
-        m_vkCmdBuff  = VK_NULL_HANDLE;
-        pDeferredCtx = std::move(m_pDeferredCtx);
+        auto vkCmdBuff = m_vkCmdBuff;
+        m_vkCmdBuff    = VK_NULL_HANDLE;
+        pDeferredCtx   = std::move(m_pDeferredCtx);
+        return vkCmdBuff;
     }
 
 private:
diff --git a/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp b/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
index 88ebd252..eb298fb1 100644
--- a/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
+++ b/Graphics/GraphicsEngineVulkan/include/DeviceContextVkImpl.hpp
@@ -236,8 +236,9 @@ 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;
@@ -400,6 +401,9 @@ private:
     void               CommitViewports();
     void               CommitScissorRects();
 
+    void Flush(Uint32               NumCommandLists,
+               ICommandList* const* ppCommandLists);
+
     __forceinline void TransitionOrVerifyBufferState(BufferVkImpl&                  Buffer,
                                                      RESOURCE_STATE_TRANSITION_MODE TransitionMode,
                                                      RESOURCE_STATE                 RequiredState,
diff --git a/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp b/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
index 126bff8c..3c5abb22 100644
--- a/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
+++ b/Graphics/GraphicsEngineVulkan/include/RenderDeviceVkImpl.hpp
@@ -239,12 +239,18 @@ 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 fe8786ff..2ae180e7 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"
@@ -1235,6 +1238,12 @@ void DeviceContextVkImpl::FinishFrame()
 
 void DeviceContextVkImpl::Flush()
 {
+    Flush(0, nullptr);
+}
+
+void DeviceContextVkImpl::Flush(Uint32               NumCommandLists,
+                                ICommandList* const* ppCommandLists)
+{
     if (m_bIsDeferred)
     {
         LOG_ERROR_MESSAGE("Flush() should only be called for immediate contexts.");
@@ -1252,10 +1261,11 @@ void DeviceContextVkImpl::Flush()
         LOG_ERROR_MESSAGE("Flushing device context inside an active render pass.");
     }
 
-    VkSubmitInfo SubmitInfo = {};
-
-    SubmitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
-    SubmitInfo.pNext = nullptr;
+    // TODO: replace with small_vector
+    std::vector<VkCommandBuffer>               vkCmdBuffs;
+    std::vector<RefCntAutoPtr<IDeviceContext>> DeferredCtxs;
+    vkCmdBuffs.reserve(NumCommandLists + 1);
+    DeferredCtxs.reserve(NumCommandLists + 1);
 
     auto vkCmdBuff = m_CommandBuffer.GetVkCmdBuffer();
     if (vkCmdBuff != VK_NULL_HANDLE)
@@ -1275,14 +1285,32 @@ void DeviceContextVkImpl::Flush()
             m_CommandBuffer.FlushBarriers();
             m_CommandBuffer.EndCommandBuffer();
 
-            SubmitInfo.commandBufferCount = 1;
-            SubmitInfo.pCommandBuffers    = &vkCmdBuff;
+            vkCmdBuffs.push_back(vkCmdBuff);
         }
     }
 
+    // Add command buffers from deferred contexts
+    for (Uint32 i = 0; i < NumCommandLists; ++i)
+    {
+        auto* pCmdListVk = ValidatedCast<CommandListVkImpl>(ppCommandLists[i]);
+        DEV_CHECK_ERR(pCmdListVk != nullptr, "Command list must not be null");
+        RefCntAutoPtr<IDeviceContext> pDeferredCtx;
+        vkCmdBuffs.emplace_back(pCmdListVk->Close(pDeferredCtx));
+        VERIFY(vkCmdBuffs.back() != VK_NULL_HANDLE, "Trying to execute empty command buffer");
+        VERIFY_EXPR(pDeferredCtx);
+        DeferredCtxs.emplace_back(std::move(pDeferredCtx));
+    }
+
     VERIFY_EXPR(m_VkWaitSemaphores.size() == m_WaitSemaphores.size());
     VERIFY_EXPR(m_VkSignalSemaphores.size() == m_SignalSemaphores.size());
 
+    VkSubmitInfo SubmitInfo = {};
+
+    SubmitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+    SubmitInfo.pNext = nullptr;
+
+    SubmitInfo.commandBufferCount = static_cast<uint32_t>(vkCmdBuffs.size());
+    SubmitInfo.pCommandBuffers    = vkCmdBuffs.data();
     SubmitInfo.waitSemaphoreCount = static_cast<uint32_t>(m_WaitSemaphores.size());
     VERIFY_EXPR(m_WaitSemaphores.size() == m_WaitDstStageMasks.size());
     SubmitInfo.pWaitSemaphores      = SubmitInfo.waitSemaphoreCount != 0 ? m_VkWaitSemaphores.data() : nullptr;
@@ -1301,11 +1329,25 @@ void DeviceContextVkImpl::Flush()
     m_VkSignalSemaphores.clear();
     m_PendingFences.clear();
 
+    size_t buff_idx = 0;
     if (vkCmdBuff != VK_NULL_HANDLE)
     {
+        VERIFY_EXPR(vkCmdBuffs[buff_idx] == vkCmdBuff);
         DisposeCurrentCmdBuffer(m_CommandQueueId, SubmittedFenceValue);
+        ++buff_idx;
     }
 
+    for (Uint32 i = 0; i < NumCommandLists; ++i, ++buff_idx)
+    {
+        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.fetch_or(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[buff_idx]), SubmittedFenceValue);
+    }
+    VERIFY_EXPR(buff_idx == vkCmdBuffs.size());
+
     for (auto& BindInfo : m_DescrSetBindInfo)
         BindInfo = DescriptorSetBindInfo{};
 
@@ -2378,7 +2420,8 @@ void DeviceContextVkImpl::FinishCommandList(class ICommandList** ppCommandList)
     InvalidateState();
 }
 
-void DeviceContextVkImpl::ExecuteCommandList(class ICommandList* pCommandList)
+void DeviceContextVkImpl::ExecuteCommandLists(Uint32               NumCommandLists,
+                                              ICommandList* const* ppCommandLists)
 {
     if (m_bIsDeferred)
     {
@@ -2386,31 +2429,13 @@ void DeviceContextVkImpl::ExecuteCommandList(class ICommandList* pCommandList)
         return;
     }
 
-    Flush();
-
-    InvalidateState();
-
-    CommandListVkImpl* pCmdListVk = ValidatedCast<CommandListVkImpl>(pCommandList);
-    VkCommandBuffer    vkCmdBuff  = VK_NULL_HANDLE;
+    if (NumCommandLists == 0)
+        return;
+    DEV_CHECK_ERR(ppCommandLists != nullptr, "ppCommandLists must not be null when NumCommandLists is not zero");
 
-    RefCntAutoPtr<IDeviceContext> pDeferredCtx;
-    pCmdListVk->Close(vkCmdBuff, pDeferredCtx);
-    VERIFY(vkCmdBuff != VK_NULL_HANDLE, "Trying to execute empty command buffer");
-    VERIFY_EXPR(pDeferredCtx);
-    VkSubmitInfo SubmitInfo = {};
+    Flush(NumCommandLists, ppCommandLists);
 
-    SubmitInfo.sType              = VK_STRUCTURE_TYPE_SUBMIT_INFO;
-    SubmitInfo.pNext              = nullptr;
-    SubmitInfo.commandBufferCount = 1;
-    SubmitInfo.pCommandBuffers    = &vkCmdBuff;
-    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);
+    InvalidateState();
 }
 
 void DeviceContextVkImpl::SignalFence(IFence* pFence, Uint64 Value)
diff --git a/Graphics/GraphicsEngineVulkan/src/PipelineStateVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/PipelineStateVkImpl.cpp
index 8d884c85..f54c8e27 100644
--- a/Graphics/GraphicsEngineVulkan/src/PipelineStateVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/PipelineStateVkImpl.cpp
@@ -45,6 +45,18 @@
 
 namespace Diligent
 {
+
+#if !DILIGENT_NO_HLSL
+namespace GLSLangUtils
+{
+void SpvOptimizerMessageConsumer(
+    spv_message_level_t level,
+    const char* /* source */,
+    const spv_position_t& /* position */,
+    const char* message);
+}
+#endif
+
 namespace
 {
 
@@ -63,6 +75,7 @@ bool StripReflection(const VulkanUtilities::VulkanLogicalDevice& LogicalDevice,
         Target = SPV_ENV_VULKAN_1_1_SPIRV_1_4;
 
     spvtools::Optimizer SpirvOptimizer(Target);
+    SpirvOptimizer.SetMessageConsumer(GLSLangUtils::SpvOptimizerMessageConsumer);
     // Decorations defined in SPV_GOOGLE_hlsl_functionality1 are the only instructions
     // removed by strip-reflect-info pass. SPIRV offsets become INVALID after this operation.
     SpirvOptimizer.RegisterPass(spvtools::CreateStripReflectInfoPass());
diff --git a/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
index bd956877..cc255fba 100644
--- a/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/RenderDeviceVkImpl.cpp
@@ -401,7 +401,7 @@ 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)
@@ -445,7 +445,7 @@ 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)
diff --git a/Graphics/GraphicsEngineVulkan/src/SwapChainVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/SwapChainVkImpl.cpp
index ee266edc..bd7c0d55 100644
--- a/Graphics/GraphicsEngineVulkan/src/SwapChainVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/SwapChainVkImpl.cpp
@@ -331,19 +331,39 @@ void SwapChainVkImpl::CreateVulkanSwapChain()
     m_SwapChainDesc.Width  = swapchainExtent.width;
     m_SwapChainDesc.Height = swapchainExtent.height;
 
-    // Mailbox is the lowest latency non-tearing presentation mode.
-    VkPresentModeKHR swapchainPresentMode = m_VSyncEnabled ? VK_PRESENT_MODE_FIFO_KHR : VK_PRESENT_MODE_MAILBOX_KHR;
-
-    bool PresentModeSupported = std::find(presentModes.begin(), presentModes.end(), swapchainPresentMode) != presentModes.end();
-    if (!PresentModeSupported)
+    // The FIFO present mode is guaranteed by the spec to always be supported.
+    VkPresentModeKHR PresentMode = VK_PRESENT_MODE_FIFO_KHR;
     {
-        VERIFY(swapchainPresentMode != VK_PRESENT_MODE_FIFO_KHR, "The FIFO present mode is guaranteed by the spec to be supported");
+        std::vector<VkPresentModeKHR> PreferredPresentModes;
+        if (m_VSyncEnabled)
+        {
+            // FIFO relaxed waits for the next VSync, but if the frame is late,
+            // it still shows it even if VSync has already passed, which may
+            // result in tearing.
+            PreferredPresentModes.push_back(VK_PRESENT_MODE_FIFO_RELAXED_KHR);
+            PreferredPresentModes.push_back(VK_PRESENT_MODE_FIFO_KHR);
+        }
+        else
+        {
+            // Mailbox is the lowest latency non-tearing presentation mode.
+            PreferredPresentModes.push_back(VK_PRESENT_MODE_MAILBOX_KHR);
+            PreferredPresentModes.push_back(VK_PRESENT_MODE_IMMEDIATE_KHR);
+            PreferredPresentModes.push_back(VK_PRESENT_MODE_FIFO_KHR);
+        }
 
-        const char* PresentModeName = nullptr;
+        for (auto PreferredMode : PreferredPresentModes)
+        {
+            if (std::find(presentModes.begin(), presentModes.end(), PreferredMode) != presentModes.end())
+            {
+                PresentMode = PreferredMode;
+                break;
+            }
+        }
 
+        const char* PresentModeName = nullptr;
 #define PRESENT_MODE_CASE(Mode) \
     case Mode: PresentModeName = #Mode; break;
-        switch (swapchainPresentMode)
+        switch (PresentMode)
         {
             PRESENT_MODE_CASE(VK_PRESENT_MODE_IMMEDIATE_KHR)
             PRESENT_MODE_CASE(VK_PRESENT_MODE_MAILBOX_KHR)
@@ -354,11 +374,7 @@ void SwapChainVkImpl::CreateVulkanSwapChain()
             default: PresentModeName = "<UNKNOWN>";
         }
 #undef PRESENT_MODE_CASE
-        LOG_WARNING_MESSAGE(PresentModeName, " is not supported. Defaulting to VK_PRESENT_MODE_FIFO_KHR");
-
-        swapchainPresentMode = VK_PRESENT_MODE_FIFO_KHR;
-        // The FIFO present mode is guaranteed by the spec to be supported
-        VERIFY(std::find(presentModes.begin(), presentModes.end(), swapchainPresentMode) != presentModes.end(), "FIFO present mode must be supported");
+        LOG_INFO_MESSAGE("Using ", PresentModeName, " swap chain present mode");
     }
 
     // Determine the number of VkImage's to use in the swap chain.
@@ -417,7 +433,7 @@ void SwapChainVkImpl::CreateVulkanSwapChain()
     swapchain_ci.preTransform       = vkPreTransform;
     swapchain_ci.compositeAlpha     = compositeAlpha;
     swapchain_ci.imageArrayLayers   = 1;
-    swapchain_ci.presentMode        = swapchainPresentMode;
+    swapchain_ci.presentMode        = PresentMode;
     swapchain_ci.oldSwapchain       = oldSwapchain;
     swapchain_ci.clipped            = VK_TRUE;
     swapchain_ci.imageColorSpace    = ColorSpace;
diff --git a/Graphics/GraphicsEngineVulkan/src/VulkanUtilities/VulkanCommandBuffer.cpp b/Graphics/GraphicsEngineVulkan/src/VulkanUtilities/VulkanCommandBuffer.cpp
index b0d72e70..b646a2ed 100644
--- a/Graphics/GraphicsEngineVulkan/src/VulkanUtilities/VulkanCommandBuffer.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/VulkanUtilities/VulkanCommandBuffer.cpp
@@ -152,12 +152,12 @@ static VkPipelineStageFlags PipelineStageFromAccessFlags(VkAccessFlags
 }
 
 
-static VkPipelineStageFlags AccessMaskFromImageLayout(VkImageLayout Layout,
-                                                      bool          IsDstMask // false - source mask
-                                                                              // true  - destination mask
+static VkAccessFlags AccessMaskFromImageLayout(VkImageLayout Layout,
+                                               bool          IsDstMask // false - source mask
+                                                                       // true  - destination mask
 )
 {
-    VkPipelineStageFlags AccessMask = 0;
+    VkAccessFlags AccessMask = 0;
     switch (Layout)
     {
         // does not support device access. This layout must only be used as the initialLayout member