[d3d11] Add option to force compute shader UAV synchronization

src/d3d11/d3d11_options.cpp

@@ -15,6 +15,7 @@ namespace dxvk {
   D3D11Options::D3D11Options(const Config& config) {
     this->zeroInitWorkgroupMemory  = config.getOption<bool>("d3d11.zeroInitWorkgroupMemory", false);
     this->forceVolatileTgsmAccess = config.getOption<bool>("d3d11.forceVolatileTgsmAccess", false);
+    this->forceComputeUavBarriers = config.getOption<bool>("d3d11.forceComputeUavBarriers", false);
     this->relaxedBarriers       = config.getOption<bool>("d3d11.relaxedBarriers", false);
     this->relaxedGraphicsBarriers = config.getOption<bool>("d3d11.relaxedGraphicsBarriers", false);
     this->maxTessFactor         = config.getOption<int32_t>("d3d11.maxTessFactor", 0);

src/d3d11/d3d11_options.h

@@ -26,6 +26,13 @@ namespace dxvk {
     /// without explicit synchronization.
     bool forceVolatileTgsmAccess = false;
 
+    /// Force UAV synchronization insided compute shaders
+    ///
+    /// Workaround for compute shaders that access overlapping
+    /// memory regions within a UAV without proper workgroup
+    /// synchroniation. Will have a negative performance impact.
+    bool forceComputeUavBarriers = false;
+
     /// Use relaxed memory barriers
     ///
     /// May improve performance in some games,

src/dxbc/dxbc_options.cpp

@@ -35,6 +35,7 @@ namespace dxvk {
     invariantPosition        = options.invariantPosition;
     zeroInitWorkgroupMemory  = options.zeroInitWorkgroupMemory;
     forceVolatileTgsmAccess  = options.forceVolatileTgsmAccess;
+    forceComputeUavBarriers  = options.forceComputeUavBarriers;
     disableMsaa              = options.disableMsaa;
     forceSampleRateShading   = options.forceSampleRateShading;
     enableSampleShadingInterlock = device->features().extFragmentShaderInterlock.fragmentShaderSampleInterlock;