[dxbc,d3d11] Add option to explicitly emit long dot products

dxvk.conf

@@ -529,6 +529,17 @@
 # d3d11.longMad = False
 # d3d9.longMad = False
 
+
+# Long Dot
+#
+# Whether to emit dot products as an FMA chain or as a plain SPIR-V dot product.
+#
+# Supported values:
+# - True/False
+
+# d3d11.longDot = False
+
+
 # Device Local Constant Buffers
 #
 # Enables using device local, host accessible memory for constant buffers in D3D9.

src/d3d11/d3d11_options.cpp

@@ -32,6 +32,7 @@ namespace dxvk {
     this->maxFrameLatency       = config.getOption<int32_t>("dxgi.maxFrameLatency", 0);
     this->exposeDriverCommandLists = config.getOption<bool>("d3d11.exposeDriverCommandLists", true);
     this->longMad               = config.getOption<bool>("d3d11.longMad", false);
+    this->longDot               = config.getOption<bool>("d3d11.longDot", false);
     this->reproducibleCommandStream = config.getOption<bool>("d3d11.reproducibleCommandStream", false);
 
     // Clamp LOD bias so that people don't abuse this in unintended ways

src/d3d11/d3d11_options.h

@@ -118,9 +118,12 @@ namespace dxvk {
     /// Shader dump path
     std::string shaderDumpPath;
 
-    /// Should we make our Mads a FFma or do it the long way with an FMul and an FAdd?
+    /// Translate Mad/Dfma to separate FMul+FAdd
     bool longMad;
 
+    /// Translate DpX to a precise FMul+FFma chain
+    bool longDot;
+
     /// Ensure that for the same D3D commands the output VK commands
     /// don't change between runs. Useful for comparative benchmarking,
     /// can negatively affect performance.

src/dxbc/dxbc_compiler.cpp

@@ -2044,15 +2044,41 @@ namespace dxvk {
     DxbcRegisterValue dst;
     dst.type.ctype  = ins.dst[0].dataType;
     dst.type.ccount = 1;
-    
+    dst.id = 0;
-    dst.id = m_module.opDot(
+
-      getVectorTypeId(dst.type),
+    if (!m_moduleInfo.options.longDot) {
-      src.at(0).id,
+      dst.id = m_module.opDot(
-      src.at(1).id);
+        getVectorTypeId(dst.type),
-    
+        src.at(0).id,
-    if (ins.controls.precise() || m_precise)
+        src.at(1).id);
-      m_module.decorate(dst.id, spv::DecorationNoContraction);
+
-    
+      if (ins.controls.precise() || m_precise)
+        m_module.decorate(dst.id, spv::DecorationNoContraction);
+    } else {
+      uint32_t componentType = getVectorTypeId(dst.type);
+      uint32_t componentCount = srcMask.popCount();
+
+      for (uint32_t i = 1; i <= componentCount; i++) {
+        uint32_t idx = componentCount - i;
+
+        if (dst.id) {
+          dst.id = m_module.opFFma(componentType,
+            m_module.opCompositeExtract(componentType, src.at(0).id, 1, &idx),
+            m_module.opCompositeExtract(componentType, src.at(1).id, 1, &idx),
+            dst.id);
+        } else {
+          dst.id = m_module.opFMul(componentType,
+            m_module.opCompositeExtract(componentType, src.at(0).id, 1, &idx),
+            m_module.opCompositeExtract(componentType, src.at(1).id, 1, &idx));
+        }
+
+        // Unconditionally mark as precise since the exact order of operation
+        // matters for some games, even if the instruction itself is not marked
+        // as precise.
+        m_module.decorate(dst.id, spv::DecorationNoContraction);
+      }
+    }
+
     dst = emitDstOperandModifiers(dst, ins.modifiers);
     emitRegisterStore(ins.dst[0], dst);
   }

src/dxbc/dxbc_options.cpp

@@ -39,6 +39,7 @@ namespace dxvk {
     forceSampleRateShading   = options.forceSampleRateShading;
     enableSampleShadingInterlock = device->features().extFragmentShaderInterlock.fragmentShaderSampleInterlock;
     longMad                  = options.longMad;
+    longDot                  = options.longDot;
 
     // Figure out float control flags to match D3D11 rules
     if (options.floatControls) {

src/dxbc/dxbc_options.h

@@ -55,8 +55,11 @@ namespace dxvk {
     /// Minimum storage buffer alignment
     VkDeviceSize minSsboAlignment = 0;
 
-    /// Should we make our Mads a FFma or do it the long way with an FMul and an FAdd?
+    /// Translate Mad/Dfma to separate FMul+FAdd
     bool longMad;
+
+    /// Translate DpX to a precise FMul+FFma chain
+    bool longDot;
   };
   
 }