From 3b030d956931084690c624dba166eedc9851777a Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Sat, 30 Nov 2019 17:01:44 +0100
Subject: [PATCH] [dxbc] Implement workaround to replace NaN render target
 outputs by zero

---
 dxvk.conf                   |  8 ++++++++
 src/d3d11/d3d11_options.cpp |  1 +
 src/d3d11/d3d11_options.h   |  4 ++++
 src/dxbc/dxbc_compiler.cpp  | 19 +++++++++++++++----
 src/dxbc/dxbc_options.cpp   |  1 +
 src/dxbc/dxbc_options.h     |  3 +++
 6 files changed, 32 insertions(+), 4 deletions(-)

diff --git a/dxvk.conf b/dxvk.conf
index 83f36a22d..f72a0c36d 100644
--- a/dxvk.conf
+++ b/dxvk.conf
@@ -126,6 +126,14 @@
 # d3d11.strictDivision = False
 
 
+# Replaces NaN outputs from fragment shaders with zeroes for floating
+# point render target. Used in some games to prevent artifacting.
+#
+# Supported values: True, False
+
+# d3d11.enableRtOutputNanFixup = False
+
+
 # Clears workgroup memory in compute shaders to zero. Some games don't do
 # this and rely on undefined behaviour. Enabling may reduce performance.
 #
diff --git a/src/d3d11/d3d11_options.cpp b/src/d3d11/d3d11_options.cpp
index 2731b3cd7..86c2b5238 100644
--- a/src/d3d11/d3d11_options.cpp
+++ b/src/d3d11/d3d11_options.cpp
@@ -9,6 +9,7 @@ namespace dxvk {
 
     this->dcSingleUseMode       = config.getOption<bool>("d3d11.dcSingleUseMode", true);
     this->strictDivision           = config.getOption<bool>("d3d11.strictDivision", false);
+    this->enableRtOutputNanFixup   = config.getOption<bool>("d3d11.enableRtOutputNanFixup", false);
     this->zeroInitWorkgroupMemory  = config.getOption<bool>("d3d11.zeroInitWorkgroupMemory", false);
     this->relaxedBarriers       = config.getOption<bool>("d3d11.relaxedBarriers", false);
     this->maxTessFactor         = config.getOption<int32_t>("d3d11.maxTessFactor", 0);
diff --git a/src/d3d11/d3d11_options.h b/src/d3d11/d3d11_options.h
index b467a2eb6..e340f25ee 100644
--- a/src/d3d11/d3d11_options.h
+++ b/src/d3d11/d3d11_options.h
@@ -25,6 +25,10 @@ namespace dxvk {
     /// games may expect correct behaviour.
     bool strictDivision;
 
+    /// Enables workaround to replace NaN render target
+    /// outputs with zero
+    bool enableRtOutputNanFixup;
+
     /// Enables out-of-bounds access check for constant
     /// buffers. Workaround for a few broken games that
     /// access random data inside their shaders.
diff --git a/src/dxbc/dxbc_compiler.cpp b/src/dxbc/dxbc_compiler.cpp
index 68de83209..05ec96c3e 100644
--- a/src/dxbc/dxbc_compiler.cpp
+++ b/src/dxbc/dxbc_compiler.cpp
@@ -5834,10 +5834,21 @@ namespace dxvk {
         scalars[c] = m_module.opVectorExtractDynamic(compTypeId, vector.id, specId);
       }
 
-      vector.id = m_module.opCompositeConstruct(
-        getVectorTypeId(vector.type),
-        vector.type.ccount,
-        scalars.data());
+      uint32_t typeId = getVectorTypeId(vector.type);
+      vector.id = m_module.opCompositeConstruct(typeId, vector.type.ccount, scalars.data());
+
+      // Replace NaN by zero if requested
+      if (m_moduleInfo.options.enableRtOutputNanFixup && vector.type.ctype == DxbcScalarType::Float32) {
+        uint32_t boolType = m_module.defBoolType();
+
+        if (vector.type.ccount > 1)
+          boolType = m_module.defVectorType(boolType, vector.type.ccount);
+
+        uint32_t zero = emitBuildConstVecf32(0.0f, 0.0f, 0.0f, 0.0f,
+          DxbcRegMask((1u << vector.type.ccount) - 1)).id;
+        uint32_t isNan = m_module.opIsNan(boolType, vector.id);
+        vector.id = m_module.opSelect(typeId, isNan, zero, vector.id);
+      }
       
       emitValueStore(m_oRegs[i], vector,
         DxbcRegMask::firstN(vector.type.ccount));
diff --git a/src/dxbc/dxbc_options.cpp b/src/dxbc/dxbc_options.cpp
index 827e507ac..0e187d710 100644
--- a/src/dxbc/dxbc_options.cpp
+++ b/src/dxbc/dxbc_options.cpp
@@ -37,6 +37,7 @@ namespace dxvk {
       case Tristate::False: minSsboAlignment = ~0u; break;
     }
     
+    enableRtOutputNanFixup   = options.enableRtOutputNanFixup;
     strictDivision           = options.strictDivision;
     zeroInitWorkgroupMemory  = options.zeroInitWorkgroupMemory;
     dynamicIndexedConstantBufferAsSsbo = options.constantBufferRangeCheck;
diff --git a/src/dxbc/dxbc_options.h b/src/dxbc/dxbc_options.h
index afdb76c49..dac8d562c 100644
--- a/src/dxbc/dxbc_options.h
+++ b/src/dxbc/dxbc_options.h
@@ -32,6 +32,9 @@ namespace dxvk {
     /// dword offsets. Fixes RE2 and DMC5 on Nvidia drivers.
     bool useSdivForBufferIndex = false;
 
+    /// Enables NaN fixup for render target outputs
+    bool enableRtOutputNanFixup = false;
+
     /// Enables sm4-compliant division-by-zero behaviour
     bool strictDivision = false;