dlssnr: multi-pass cost slider and matched residual transfer

OptiScaler/Config.cpp

@@ -317,6 +317,8 @@ bool Config::Reload(std::filesystem::path iniPath)
 
             // --- DLSS 5 Neural Rendering (OptiScaler/dlssnr) ---
             DlssNrEnabled.set_from_config(readBool("DlssNr", "Enabled"));
+            DlssNrMode.set_from_config(readUInt("DlssNr", "Mode"));
+            DlssNrFeature1Pipeline.set_from_config(readUInt("DlssNr", "Feature1Pipeline"));
             DlssNrToggleKey.set_from_config(readInt("DlssNr", "ToggleKey"));
             DlssNrTransferStrength.set_from_config(readFloat("DlssNr", "TransferStrength"));
             DlssNrColourStrength.set_from_config(readFloat("DlssNr", "ColourStrength"));
@@ -327,6 +329,14 @@ bool Config::Reload(std::filesystem::path iniPath)
             DlssNrCompareZoom.set_from_config(readFloat("DlssNr", "CompareZoom"));
             DlssNrCompareSwap.set_from_config(readBool("DlssNr", "CompareSwap"));
             DlssNrWorkingScale.set_from_config(readFloat("DlssNr", "WorkingScale"));
+            DlssNrWorkAtNative.set_from_config(readBool("DlssNr", "WorkAtNative"));
+            DlssNrTransfer.set_from_config(readUInt("DlssNr", "Transfer"));
+            DlssNrHdrLift.set_from_config(readUInt("DlssNr", "HdrLift"));
+            if (DlssNrWorkingScale.has_value() && DlssNrWorkingScale.value() <= 0.0f)
+            {
+                DlssNrWorkAtNative = true;
+                DlssNrWorkingScale = 1.0f;
+            }
             DlssNrProxyProbe.set_from_config(readBool("DlssNr", "ProxyProbe"));
             DlssNrUseProxy.set_from_config(readBool("DlssNr", "UseProxy"));
             DlssNrAutoCapture.set_from_config(readBool("DlssNr", "AutoCapture"));
@@ -1166,38 +1176,43 @@ bool Config::SaveIni()
     {
         ini.SetValue("DLSS", "Enabled", GetBoolValue(Instance()->DLSSEnabled.value_for_config()).c_str());
 
-    // --- DLSS 5 Neural Rendering (OptiScaler/dlssnr) ---
-    ini.SetValue("DlssNr", "Enabled", GetBoolValue(Instance()->DlssNrEnabled.value_for_config()).c_str());
-    {
-        auto toggle = Instance()->DlssNrToggleKey.value_for_config();
-        ini.SetValue("DlssNr", "ToggleKey", GetIntValue(toggle, toggle > 0).c_str());
-    }
-    ini.SetValue("DlssNr", "TransferStrength",
-                 GetFloatValue(Instance()->DlssNrTransferStrength.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "ColourStrength",
-                 GetFloatValue(Instance()->DlssNrColourStrength.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "MaxRatio", GetFloatValue(Instance()->DlssNrMaxRatio.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "DebugView", GetIntValue(Instance()->DlssNrDebugView.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "Compare", GetIntValue(Instance()->DlssNrCompare.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "CompareSplit",
-                 GetFloatValue(Instance()->DlssNrCompareSplit.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "CompareZoom",
-                 GetFloatValue(Instance()->DlssNrCompareZoom.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "CompareSwap",
-                 GetBoolValue(Instance()->DlssNrCompareSwap.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "WorkingScale", GetFloatValue(Instance()->DlssNrWorkingScale.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "AutoCapture", GetBoolValue(Instance()->DlssNrAutoCapture.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "WhitePointScale",
-                 GetFloatValue(Instance()->DlssNrWhitePointScale.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "Preset", GetIntValue(Instance()->DlssNrPreset.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "Intensity", GetFloatValue(Instance()->DlssNrIntensity.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "Style", GetIntValue(Instance()->DlssNrStyle.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "LocalStructure",
-                 GetFloatValue(Instance()->DlssNrLocalStructure.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "LocalTone", GetFloatValue(Instance()->DlssNrLocalTone.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "SkinStructure",
-                 GetFloatValue(Instance()->DlssNrSkinStructure.value_for_config()).c_str());
-    ini.SetValue("DlssNr", "AutoMask", GetBoolValue(Instance()->DlssNrAutoMask.value_for_config()).c_str());
+        // --- DLSS 5 Neural Rendering (OptiScaler/dlssnr) ---
+        ini.SetValue("DlssNr", "Enabled", GetBoolValue(Instance()->DlssNrEnabled.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Mode", GetIntValue(Instance()->DlssNrMode.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Feature1Pipeline",
+                     GetIntValue(Instance()->DlssNrFeature1Pipeline.value_for_config()).c_str());
+        {
+            auto toggle = Instance()->DlssNrToggleKey.value_for_config();
+            ini.SetValue("DlssNr", "ToggleKey", GetIntValue(toggle, toggle > 0).c_str());
+        }
+        ini.SetValue("DlssNr", "TransferStrength",
+                     GetFloatValue(Instance()->DlssNrTransferStrength.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "ColourStrength",
+                     GetFloatValue(Instance()->DlssNrColourStrength.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "MaxRatio", GetFloatValue(Instance()->DlssNrMaxRatio.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "DebugView", GetIntValue(Instance()->DlssNrDebugView.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Compare", GetIntValue(Instance()->DlssNrCompare.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "CompareSplit",
+                     GetFloatValue(Instance()->DlssNrCompareSplit.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "CompareZoom", GetFloatValue(Instance()->DlssNrCompareZoom.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "CompareSwap", GetBoolValue(Instance()->DlssNrCompareSwap.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "WorkingScale",
+                     GetFloatValue(Instance()->DlssNrWorkingScale.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "WorkAtNative", GetBoolValue(Instance()->DlssNrWorkAtNative.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Transfer", GetIntValue(Instance()->DlssNrTransfer.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "HdrLift", GetIntValue(Instance()->DlssNrHdrLift.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "AutoCapture", GetBoolValue(Instance()->DlssNrAutoCapture.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "WhitePointScale",
+                     GetFloatValue(Instance()->DlssNrWhitePointScale.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Preset", GetIntValue(Instance()->DlssNrPreset.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Intensity", GetFloatValue(Instance()->DlssNrIntensity.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "Style", GetIntValue(Instance()->DlssNrStyle.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "LocalStructure",
+                     GetFloatValue(Instance()->DlssNrLocalStructure.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "LocalTone", GetFloatValue(Instance()->DlssNrLocalTone.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "SkinStructure",
+                     GetFloatValue(Instance()->DlssNrSkinStructure.value_for_config()).c_str());
+        ini.SetValue("DlssNr", "AutoMask", GetBoolValue(Instance()->DlssNrAutoMask.value_for_config()).c_str());
         ini.SetValue("DLSS", "RenderPresetOverride",
                      GetBoolValue(Instance()->RenderPresetOverride.value_for_config()).c_str());
         ini.SetValue("DLSS", "RenderPresetForAll",

OptiScaler/Config.h

@@ -257,6 +257,12 @@ class Config
     // DLSS Neural Rendering: a detail-synthesis pass over the upscaler's output. Off by default -- it is
     // an undocumented feature driven directly through its snippet, not something NVIDIA exposes.
     CustomOptional<bool> DlssNrEnabled { false };
+    // 0 post-process (after the upscaler), 1 multi-pass (1:1 first pass, then the model, then a
+    // spatial enlarge). OptiScaler's own Dx12 DLSS / Ray Reconstruction only; anything else falls
+    // back to post-process.
+    CustomOptional<uint32_t> DlssNrMode { 0 };
+    // Which upscaler the multi-pass first pass is: 0 Ray Reconstruction, 1 Super Resolution.
+    CustomOptional<uint32_t> DlssNrFeature1Pipeline { 0 };
     // Toggles the pass in game. Unbound by default -- a key that does something unexpected is worse
     // than one that does nothing.
     CustomOptional<int> DlssNrToggleKey { UnboundKey };
@@ -276,12 +282,11 @@ class Config
     CustomOptional<float> DlssNrTransferStrength { 1.0f };
     CustomOptional<float> DlssNrColourStrength { 1.0f };
 
-
     // The most the pass may multiply or divide a pixel by. A detail pass has no business restyling a
     // light source, whatever the model returns.
     CustomOptional<float> DlssNrMaxRatio { 2.0f };
 
-    // 0 off, 1 the picture the model was shown, 2 its raw answer, 3 what it changed, amplified.
+    // 0–5: off, proxy, model, amplified difference, full-res proxy, matched T.
     CustomOptional<uint32_t> DlssNrDebugView { 0 };
 
     // Showing the pass against itself, without having to toggle it and remember what the last frame
@@ -300,10 +305,18 @@ class Config
     // Which side the edited frame sits on, in both comparison modes.
     CustomOptional<bool> DlssNrCompareSwap { false };
 
-    // The fraction of the frame's resolution the model works at. The frame itself is never reduced --
-    // only the model's contribution is computed small and enlarged, so the picture underneath is
-    // untouched whatever this is set to. 1.0 is full resolution and behaves exactly as before.
+    // The model's working raster as a fraction of display resolution, not of the colour buffer.
+    // The frame itself is never reduced -- only the model's contribution is computed small and
+    // enlarged. 1.0 is display resolution, then clamped to the colour buffer so Multi-pass cannot
+    // invent pixels past Feature 1.
     CustomOptional<float> DlssNrWorkingScale { 1.0f };
+    // Lock the model to the game's render (native) raster. The slider then shows the live R/D
+    // ratio and is not adjustable. WorkingScale is left alone so unchecking restores it.
+    CustomOptional<bool> DlssNrWorkAtNative { false };
+    // 0 Classic, 1 Matched residual (default). How a below-frame model is brought back.
+    CustomOptional<uint32_t> DlssNrTransfer { 1 };
+    // 0 UpgradeToneMap (H0), 1 additive headroom (H1). Used only if Transfer == 1.
+    CustomOptional<uint32_t> DlssNrHdrLift { 0 };
 
     // Ask the driver's own nvngx.dll whether it will dispatch Neural Rendering, once per session.
     //
@@ -337,18 +350,10 @@ class Config
     // folder is cleared at the start of each run, so it holds one session's worth and never grows.
     CustomOptional<bool> DlssNrAutoCapture { true };
 
-
-
-
-
     // Multiplies the (auto or manual) white point before the encode: what the model considers "white".
     // Higher means highlights sit lower on the curve and the model treats them as less extreme.
     CustomOptional<float> DlssNrWhitePointScale { 1.0f };
 
-
-
-
-
     // --- end DLSS 5 Neural Rendering -------------------------------------------------------------
 
     // DLSS

OptiScaler/OptiScaler.vcxproj

@@ -824,6 +824,7 @@ copy NUL "$(SolutionDir)x64\Release\a\!! EXTRACT ALL FILES TO GAME FOLDER !!" /Y
     <ClCompile Include="shaders\dlssnr\DlssNr_Dx12.cpp" />
     <ClCompile Include="dlssnr\DlssNr_Proxy.cpp" />
     <ClCompile Include="dlssnr\DlssNr_Menu.cpp" />
+    <ClCompile Include="dlssnr\DlssNr_Modes.cpp" />
     <ClCompile Include="inputs\NVNGX_DLSS_Dx12.cpp" />
     <ClCompile Include="dllmain.cpp" />
     <ClCompile Include="inputs\NVNGX_DLSS_Vk.cpp" />

OptiScaler/dlssnr/DlssNr.h

@@ -14,7 +14,8 @@
 //   shaders/dlssnr/DlssNr_Dx12   the composition pass, as a Shader_Dx12
 //
 // Call sites, for the record:
-//   inputs/NVNGX_DLSS_Dx12.cpp        the pass after an upscale
+//   inputs/NVNGX_DLSS_Dx12.cpp        the pass after an upscale (PostProcess)
+//   upscalers/IFeature_Dx12.cpp       multi-pass stage (resources in, at R)
 //   upscalers/IFeature_Dx11wDx12.cpp  the pass inside the D3D11-on-D3D12 bridge
 //   menu/menu_common.cpp              the settings panel
 

OptiScaler/dlssnr/DlssNrFeature_Dx12.h

@@ -33,21 +33,22 @@ namespace DlssNr
 void EvaluateAfterUpscale(ID3D12GraphicsCommandList* cmdList, NVSDK_NGX_Parameter* params,
                           ID3D12CommandQueue* timingQueue = nullptr);
 
-
+// Resources in. Builds the composition pass on first use if it is not already
+// on this device, then Dispatches. Call sites that already hold the textures
+// -- the multi-pass pipeline stage -- use this rather than reading an NGX
+// parameter block.
+void Run(ID3D12GraphicsCommandList* cmdList, ID3D12Resource* colour, ID3D12Resource* depth, ID3D12Resource* motion,
+         ID3D12Resource* output, const DlssNrFrameInfo& frame, ID3D12CommandQueue* timingQueue = nullptr);
 
 // Frame generation titles tag their UI layer through Streamline; a copy of it makes the HUD mask
 // exact at the finished frame. Called at tag time.
 
-
-
-
 // The settings panel, drawn inside OptiScaler's menu.
 void RenderMenu(::Config* config, float menuResScale);
 
 // Clears the session failure latch, so a failure caused by transient thrash does not cost a restart.
 void RetryAfterFailure();
 
-
 // Whether the model is loaded and running, for the overlay.
 bool IsRunning();
 

OptiScaler/dlssnr/DlssNr_Area.h

@@ -0,0 +1,64 @@
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+
+namespace DlssNr
+{
+
+inline float Overlap1D(float box0, float box1, int i)
+{
+    const float t0 = (float) i;
+    const float t1 = (float) (i + 1);
+    const float a = box0 > t0 ? box0 : t0;
+    const float b = box1 < t1 ? box1 : t1;
+    return b > a ? (b - a) : 0.0f;
+}
+
+struct AreaSample
+{
+    float r, g, b, a;
+};
+
+// src is row-major srcW*srcH. Matches spec §7.1 / HLSL Mode 2 Matched.
+inline AreaSample AreaDownsamplePixel(const AreaSample* src, int srcW, int srcH, int dstW, int dstH, int x, int y)
+{
+    if (dstW == srcW && dstH == srcH)
+        return src[y * srcW + x];
+
+    const float x0 = ((float) x * (float) srcW) / (float) dstW;
+    const float x1 = ((float) (x + 1) * (float) srcW) / (float) dstW;
+    const float y0 = ((float) y * (float) srcH) / (float) dstH;
+    const float y1 = ((float) (y + 1) * (float) srcH) / (float) dstH;
+    const float area = (x1 - x0) * (y1 - y0);
+
+    const int i0 = (int) floorf(x0);
+    const int i1 = (int) ceilf(x1) - 1;
+    const int j0 = (int) floorf(y0);
+    const int j1 = (int) ceilf(y1) - 1;
+
+    float ar = 0.f, ag = 0.f, ab = 0.f;
+    for (int j = j0; j <= j1; ++j)
+    {
+        const int jj = std::clamp(j, 0, srcH - 1);
+        const float wy = Overlap1D(y0, y1, j);
+        for (int i = i0; i <= i1; ++i)
+        {
+            const int ii = std::clamp(i, 0, srcW - 1);
+            const float w = Overlap1D(x0, x1, i) * wy;
+            const AreaSample& s = src[jj * srcW + ii];
+            ar += s.r * w;
+            ag += s.g * w;
+            ab += s.b * w;
+        }
+    }
+
+    const float cx = ((float) x + 0.5f) * (float) srcW / (float) dstW;
+    const float cy = ((float) y + 0.5f) * (float) srcH / (float) dstH;
+    const int acx = std::clamp((int) floorf(cx), 0, srcW - 1);
+    const int acy = std::clamp((int) floorf(cy), 0, srcH - 1);
+
+    return { ar / area, ag / area, ab / area, src[acy * srcW + acx].a };
+}
+
+} // namespace DlssNr

OptiScaler/dlssnr/DlssNr_Capture.h

@@ -216,13 +216,11 @@ class FrameCapture
         buf.SampleDesc.Count = 1;
         buf.Layout = D3D12_TEXTURE_LAYOUT_ROW_MAJOR;
 
-        return SUCCEEDED(device->CreateCommittedResource(&heap, D3D12_HEAP_FLAG_NONE, &buf,
-                                                         D3D12_RESOURCE_STATE_COPY_DEST, nullptr,
-                                                         IID_PPV_ARGS(&shot.readback)));
+        return SUCCEEDED(device->CreateCommittedResource(
+            &heap, D3D12_HEAP_FLAG_NONE, &buf, D3D12_RESOURCE_STATE_COPY_DEST, nullptr, IID_PPV_ARGS(&shot.readback)));
     }
 
-    static void copy(ID3D12GraphicsCommandList* cmd, ID3D12Resource* src, D3D12_RESOURCE_STATES state,
-                     Shot& shot)
+    static void copy(ID3D12GraphicsCommandList* cmd, ID3D12Resource* src, D3D12_RESOURCE_STATES state, Shot& shot)
     {
         if (shot.readback == nullptr)
             return;
@@ -320,11 +318,10 @@ class FrameCapture
         {
             std::fprintf(f, "frames %u\n", captured_);
             std::fprintf(f, "before width %llu height %u format %d rowPitch %u\n",
-                         (unsigned long long) beforeDesc_.Width, beforeDesc_.Height,
-                         (int) beforeDesc_.Format,
+                         (unsigned long long) beforeDesc_.Width, beforeDesc_.Height, (int) beforeDesc_.Format,
                          beforeShots_.empty() ? 0 : beforeShots_[0].layout.Footprint.RowPitch);
-            std::fprintf(f, "after width %llu height %u format %d rowPitch %u\n",
-                         (unsigned long long) afterDesc_.Width, afterDesc_.Height, (int) afterDesc_.Format,
+            std::fprintf(f, "after width %llu height %u format %d rowPitch %u\n", (unsigned long long) afterDesc_.Width,
+                         afterDesc_.Height, (int) afterDesc_.Format,
                          afterShots_.empty() ? 0 : afterShots_[0].layout.Footprint.RowPitch);
             std::fprintf(f, "\nbefore_NN.raw is the frame as the upscaler produced it.\n");
             std::fprintf(f, "after_NN.raw is the same frame once the model's edit was applied.\n");