[d3d11] Add compile-time debug flag for lazy binding

src/d3d11/d3d11_context.cpp

@@ -3548,7 +3548,7 @@ namespace dxvk {
       m_state.lazy.shadersUsed.set(ShaderStage);
       m_state.lazy.bindingsUsed[ShaderStage] = pShaderModule->GetBindingMask();
 
-      if (!m_state.lazy.shadersDirty.test(ShaderStage)) {
+      if (!m_state.lazy.shadersDirty.test(ShaderStage) && (DebugLazyBinding != Tristate::False)) {
         if (!(m_state.lazy.bindingsDirty[ShaderStage] & m_state.lazy.bindingsUsed[ShaderStage]).empty())
           m_state.lazy.shadersDirty.set(ShaderStage);
       }
@@ -4368,11 +4368,15 @@ namespace dxvk {
           T&                                DirtyMask,
           T                                 DirtyBit,
           bool                              IsNull) {
+    // Forward immediately if lazy binding is forced off
+    if (DebugLazyBinding == Tristate::False)
+      return false;
+
     if ((BoundMask & ~DirtyMask) & DirtyBit) {
       // If we're binding a non-null resource to an active slot that has not been
       // marked for lazy binding yet, forward the call immediately in order to
       // avoid tracking overhead. This is by far the most common case.
-      if (likely(!IsNull))
+      if (likely(!IsNull && DebugLazyBinding != Tristate::True))
         return false;
 
       // If we are binding a null resource to an active slot, the app will likely

src/d3d11/d3d11_context.h

@@ -75,6 +75,11 @@ namespace dxvk {
     // Use a local staging buffer to handle tiny uploads, most
     // of the time we're fine with hitting the global allocator
     constexpr static VkDeviceSize StagingBufferSize = 256ull << 10;
+
+  protected:
+    // Compile-time debug flag to force lazy binding on (True) or off (False)
+    constexpr static Tristate DebugLazyBinding = Tristate::Auto;
+
   public:
 
     D3D11CommonContext(

src/d3d11/d3d11_context_imm.cpp

@@ -1053,6 +1053,10 @@ namespace dxvk {
 
   void D3D11ImmediateContext::ConsiderFlush(
           GpuFlushType                FlushType) {
+    // In stress test mode, behave as if this would always flush
+    if (DebugLazyBinding == Tristate::True)
+      ApplyDirtyNullBindings();
+
     uint64_t chunkId = GetCurrentSequenceNumber();
     uint64_t submissionId = m_submissionFence->value();