[dxbc] Add option to force memory barriers after shared memory stores

src/d3d11/d3d11_options.cpp

@@ -10,6 +10,7 @@ namespace dxvk {
     this->dcSingleUseMode       = config.getOption<bool>("d3d11.dcSingleUseMode", true);
     this->enableRtOutputNanFixup   = config.getOption<bool>("d3d11.enableRtOutputNanFixup", false);
     this->zeroInitWorkgroupMemory  = config.getOption<bool>("d3d11.zeroInitWorkgroupMemory", false);
+    this->forceTgsmBarriers     = config.getOption<bool>("d3d11.forceTgsmBarriers", false);
     this->relaxedBarriers       = config.getOption<bool>("d3d11.relaxedBarriers", false);
     this->maxTessFactor         = config.getOption<int32_t>("d3d11.maxTessFactor", 0);
     this->samplerAnisotropy     = config.getOption<int32_t>("d3d11.samplerAnisotropy", -1);

src/d3d11/d3d11_options.h

@@ -35,6 +35,13 @@ namespace dxvk {
     /// TGSM in compute shaders before reading it.
     bool zeroInitWorkgroupMemory;
 
+    /// Force thread-group shared memory barriers
+    ///
+    /// Workaround for compute shaders that read and
+    /// write from the same shared memory location
+    /// without explicit synchronization.
+    bool forceTgsmBarriers;
+
     /// Use relaxed memory barriers
     ///
     /// May improve performance in some games,

src/dxbc/dxbc_compiler.cpp

@@ -5221,6 +5221,15 @@ namespace dxvk {
         srcComponentIndex += 1;
       }
     }
+
+    // Make sure that shared memory stores are made visible in
+    // case the game does not synchronize invocations properly
+    if (isTgsm && m_moduleInfo.options.forceTgsmBarriers) {
+      m_module.opMemoryBarrier(
+        m_module.constu32(spv::ScopeWorkgroup),
+        m_module.constu32(spv::MemorySemanticsWorkgroupMemoryMask
+                        | spv::MemorySemanticsAcquireReleaseMask));
+    }
     
     // End conditional block
     if (!isTgsm) {

src/dxbc/dxbc_options.cpp

@@ -40,6 +40,7 @@ namespace dxvk {
     invariantPosition        = options.invariantPosition;
     enableRtOutputNanFixup   = options.enableRtOutputNanFixup;
     zeroInitWorkgroupMemory  = options.zeroInitWorkgroupMemory;
+    forceTgsmBarriers        = options.forceTgsmBarriers;
     dynamicIndexedConstantBufferAsSsbo = options.constantBufferRangeCheck;
 
     // Disable early discard on RADV (with LLVM) due to GPU hangs

src/dxbc/dxbc_options.h

@@ -45,6 +45,9 @@ namespace dxvk {
     /// Declare vertex positions as invariant
     bool invariantPosition = false;
 
+    /// Insert memory barriers after TGSM stoes
+    bool forceTgsmBarriers = false;
+
     /// Minimum storage buffer alignment
     VkDeviceSize minSsboAlignment = 0;
   };