From 1e6950fd77ac3c23a37313efb9980209e7b74927 Mon Sep 17 00:00:00 2001
From: Laszlo Agocs <laszlo.agocs@qt.io>
Date: Mon, 22 Jul 2024 14:55:36 +0200
Subject: [PATCH] rhi: d3d12: Also default to max frame latency 2

Follow what's been done in the D3D11 backend. Slightly different
logic here and there due to always having the newer interfaces
available, but does the same thing. Use QT_D3D_MAX_FRAME_LATENCY
to override the default 2.

[ChangeLog][RHI] The D3D12 backend creates swapchains from now on with
DXGI_SWAP_CHAIN_FLAG_FRAME_LATENCY_WAITABLE_OBJECT by default, with a
max frame latency of 2.

Task-number: QTBUG-127267
Change-Id: I4d0361ab546a1c0041592389f8954281f588b0ba
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
(cherry picked from commit 91cae3f9cc9cb97eab69e39d405000f2fd1e533a)
Reviewed-by: Qt Cherry-pick Bot <cherrypick_bot@qt-project.org>
---
 src/gui/rhi/qrhid3d12.cpp | 25 +++++++++++++++++++++++++
 src/gui/rhi/qrhid3d12_p.h |  2 ++
 2 files changed, 27 insertions(+)

diff --git a/src/gui/rhi/qrhid3d12.cpp b/src/gui/rhi/qrhid3d12.cpp
index 0f176c683dc..2a178fbb9a0 100644
--- a/src/gui/rhi/qrhid3d12.cpp
+++ b/src/gui/rhi/qrhid3d12.cpp
@@ -223,6 +223,11 @@ bool QRhiD3D12::create(QRhi::Flags flags)
         }
     }
 
+    if (qEnvironmentVariableIsSet("QT_D3D_MAX_FRAME_LATENCY"))
+        maxFrameLatency = UINT(qMax(0, qEnvironmentVariableIntValue("QT_D3D_MAX_FRAME_LATENCY")));
+    if (maxFrameLatency != 0)
+        qCDebug(QRHI_LOG_INFO, "Using frame latency waitable object with max frame latency %u", maxFrameLatency);
+
     supportsAllowTearing = false;
     IDXGIFactory5 *factory5 = nullptr;
     if (SUCCEEDED(dxgiFactory->QueryInterface(__uuidof(IDXGIFactory5), reinterpret_cast<void **>(&factory5)))) {
@@ -1527,6 +1532,9 @@ QRhi::FrameOpResult QRhiD3D12::beginFrame(QRhiSwapChain *swapChain, QRhi::BeginF
     for (QD3D12SwapChain *sc : std::as_const(swapchains))
         sc->waitCommandCompletionForFrameSlot(currentFrameSlot); // note: swapChainD->currentFrameSlot, not sc's
 
+    if (swapChainD->frameLatencyWaitableObject)
+        WaitForSingleObjectEx(swapChainD->frameLatencyWaitableObject, 1000, true);
+
     HRESULT hr = cmdAllocators[currentFrameSlot]->Reset();
     if (FAILED(hr)) {
         qWarning("Failed to reset command allocator: %s",
@@ -6122,6 +6130,11 @@ void QD3D12SwapChain::destroy()
         dcompTarget = nullptr;
     }
 
+    if (frameLatencyWaitableObject) {
+        CloseHandle(frameLatencyWaitableObject);
+        frameLatencyWaitableObject = nullptr;
+    }
+
     QRHI_RES_RHI(QRhiD3D12);
     if (rhiD) {
         rhiD->swapchains.remove(this);
@@ -6338,6 +6351,14 @@ bool QD3D12SwapChain::createOrResize()
     if (swapInterval == 0 && rhiD->supportsAllowTearing)
         swapChainFlags |= DXGI_SWAP_CHAIN_FLAG_ALLOW_TEARING;
 
+    // maxFrameLatency 0 means no waitable object usage.
+    // Ignore it also when NoVSync is on, and when using WARP.
+    const bool useFrameLatencyWaitableObject = rhiD->maxFrameLatency != 0
+                                               && swapInterval != 0
+                                               && rhiD->driverInfoStruct.deviceType != QRhiDriverInfo::CpuDevice;
+    if (useFrameLatencyWaitableObject)
+        swapChainFlags |= DXGI_SWAP_CHAIN_FLAG_FRAME_LATENCY_WAITABLE_OBJECT;
+
     if (!swapChain) {
         chooseFormats();
 
@@ -6394,6 +6415,10 @@ bool QD3D12SwapChain::createOrResize()
                              qPrintable(QSystemError::windowsComString(hr)));
                 }
             }
+            if (useFrameLatencyWaitableObject) {
+                swapChain->SetMaximumFrameLatency(rhiD->maxFrameLatency);
+                frameLatencyWaitableObject = swapChain->GetFrameLatencyWaitableObject();
+            }
             if (dcompVisual) {
                 hr = dcompVisual->SetContent(swapChain);
                 if (SUCCEEDED(hr)) {
diff --git a/src/gui/rhi/qrhid3d12_p.h b/src/gui/rhi/qrhid3d12_p.h
index 3f9abbb5acb..72c6c95faaf 100644
--- a/src/gui/rhi/qrhid3d12_p.h
+++ b/src/gui/rhi/qrhid3d12_p.h
@@ -1044,6 +1044,7 @@ struct QD3D12SwapChain : public QRhiSwapChain
     QD3D12SwapChainRenderTarget rtWrapper;
     QD3D12SwapChainRenderTarget rtWrapperRight;
     QD3D12CommandBuffer cbWrapper;
+    HANDLE frameLatencyWaitableObject = nullptr;
 
     struct FrameResources {
         ID3D12Fence *fence = nullptr;
@@ -1194,6 +1195,7 @@ public:
     void bindShaderVisibleHeaps(QD3D12CommandBuffer *cbD);
 
     bool debugLayer = false;
+    UINT maxFrameLatency = 2; // 1-3, use 2 to keep CPU-GPU parallelism while reducing lag compared to tripple buffering
     ID3D12Device2 *dev = nullptr;
     D3D_FEATURE_LEVEL minimumFeatureLevel = D3D_FEATURE_LEVEL(0);
     LUID adapterLuid = {};