[dxbc] Rework TGSM workaround

src/d3d11/d3d11_options.cpp

@@ -16,7 +16,7 @@ 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->forceVolatileTgsmAccess = config.getOption<bool>("d3d11.forceVolatileTgsmAccess", false);
     this->relaxedBarriers       = config.getOption<bool>("d3d11.relaxedBarriers", false);
     this->ignoreGraphicsBarriers = config.getOption<bool>("d3d11.ignoreGraphicsBarriers", false);
     this->maxTessFactor         = config.getOption<int32_t>("d3d11.maxTessFactor", 0);

src/d3d11/d3d11_options.h

@@ -30,12 +30,12 @@ namespace dxvk {
     /// TGSM in compute shaders before reading it.
     bool zeroInitWorkgroupMemory;
 
-    /// Force thread-group shared memory barriers
+    /// Force thread-group shared memory accesses to be volatile
     ///
     /// Workaround for compute shaders that read and
     /// write from the same shared memory location
     /// without explicit synchronization.
-    bool forceTgsmBarriers;
+    bool forceVolatileTgsmAccess;
 
     /// Use relaxed memory barriers
     ///

src/dxbc/dxbc_compiler.cpp

@@ -5190,17 +5190,23 @@ namespace dxvk {
     SpirvImageOperands imageOperands;
     imageOperands.sparse = sparseFeedbackId != 0;
 
-    if (isTgsm)
+    uint32_t coherence = bufferInfo.coherence;
+
+    if (isTgsm || coherence)
       memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask;
 
-    if (bufferInfo.coherence) {
-      memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask
-                           | spv::MemoryAccessMakePointerVisibleMask;
-      memoryOperands.makeVisible = m_module.constu32(bufferInfo.coherence);
+    if (isTgsm && m_moduleInfo.options.forceVolatileTgsmAccess) {
+      memoryOperands.flags |= spv::MemoryAccessVolatileMask;
+      coherence = spv::ScopeWorkgroup;
+    }
+
+    if (coherence) {
+      memoryOperands.flags |= spv::MemoryAccessMakePointerVisibleMask;
+      memoryOperands.makeVisible = m_module.constu32(coherence);
 
       imageOperands.flags = spv::ImageOperandsNonPrivateTexelMask
                           | spv::ImageOperandsMakeTexelVisibleMask;
-      imageOperands.makeVisible = m_module.constu32(bufferInfo.coherence);
+      imageOperands.makeVisible = m_module.constu32(coherence);
     }
 
     sparseFeedbackId = 0;
@@ -5308,17 +5314,23 @@ namespace dxvk {
     SpirvMemoryOperands memoryOperands;
     SpirvImageOperands imageOperands;
 
-    if (isTgsm)
+    uint32_t coherence = bufferInfo.coherence;
+
+    if (isTgsm || coherence)
       memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask;
 
-    if (bufferInfo.coherence) {
-      memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask
-                           | spv::MemoryAccessMakePointerAvailableMask;
-      memoryOperands.makeAvailable = m_module.constu32(bufferInfo.coherence);
+    if (isTgsm && m_moduleInfo.options.forceVolatileTgsmAccess) {
+      memoryOperands.flags |= spv::MemoryAccessVolatileMask;
+      coherence = spv::ScopeWorkgroup;
+    }
+
+    if (coherence) {
+      memoryOperands.flags |= spv::MemoryAccessMakePointerAvailableMask;
+      memoryOperands.makeAvailable = m_module.constu32(coherence);
 
       imageOperands.flags = spv::ImageOperandsNonPrivateTexelMask
                           | spv::ImageOperandsMakeTexelAvailableMask;
-      imageOperands.makeAvailable = m_module.constu32(bufferInfo.coherence);
+      imageOperands.makeAvailable = m_module.constu32(coherence);
     }
 
     for (uint32_t i = 0; i < 4; i++) {

src/dxbc/dxbc_options.cpp

@@ -37,7 +37,7 @@ namespace dxvk {
     invariantPosition        = options.invariantPosition;
     enableRtOutputNanFixup   = options.enableRtOutputNanFixup;
     zeroInitWorkgroupMemory  = options.zeroInitWorkgroupMemory;
-    forceTgsmBarriers        = options.forceTgsmBarriers;
+    forceVolatileTgsmAccess  = options.forceVolatileTgsmAccess;
     disableMsaa              = options.disableMsaa;
 
     // Figure out float control flags to match D3D11 rules

src/dxbc/dxbc_options.h

@@ -41,7 +41,7 @@ namespace dxvk {
     bool invariantPosition = false;
 
     /// Insert memory barriers after TGSM stoes
-    bool forceTgsmBarriers = false;
+    bool forceVolatileTgsmAccess = false;
 
     /// Replace ld_ms with ld
     bool disableMsaa = false;

src/util/config/config.cpp

@@ -181,7 +181,7 @@ namespace dxvk {
     /* F1 games - do not synchronize TGSM access  *
      * in a compute shader, causing artifacts     */
     { R"(\\F1_20(1[89]|[2-9][0-9])\.exe$)", {{
-      { "d3d11.forceTgsmBarriers",          "True" },
+      { "d3d11.forceVolatileTgsmAccess",    "True" },
     }} },
     /* Darksiders Warmastered - apparently reads  *
      * from write-only mapped buffers             */