[d3d9] Clip plane compaction

Compact clip planes to the smallest amount that are enabled.

Signed-off-by: Autumn Ashton <[email protected]>

Author: misyltoad

src/d3d9/d3d9_device.cpp

@@ -5933,17 +5933,21 @@ namespace dxvk {
     auto mapPtr = m_vsClipPlanes.AllocSlice();
     auto dst = reinterpret_cast<D3D9ClipPlane*>(mapPtr);
 
-    uint32_t clipPlaneMask = 0u;
+    uint32_t clipPlaneCount = 0u;
     for (uint32_t i = 0; i < caps::MaxClipPlanes; i++) {
-      dst[i] = (m_state.renderStates[D3DRS_CLIPPLANEENABLE] & (1 << i))
+      D3D9ClipPlane clipPlane = (m_state.renderStates[D3DRS_CLIPPLANEENABLE] & (1 << i))
         ? m_state.clipPlanes[i]
         : D3D9ClipPlane();
 
-      if (dst[i] != D3D9ClipPlane())
-        clipPlaneMask |= 1u << i;
+      if (clipPlane != D3D9ClipPlane())
+        dst[clipPlaneCount++] = clipPlane;
     }
 
-    if (m_specInfo.set<SpecClipPlaneMask>(clipPlaneMask))
+    // Write the rest to 0 for GPL.
+    for (uint32_t i = clipPlaneCount; i < caps::MaxClipPlanes; i++)
+      dst[i] = D3D9ClipPlane();
+
+    if (m_specInfo.set<SpecClipPlaneCount>(clipPlaneCount))
       m_flags.set(D3D9DeviceFlag::DirtySpecializationEntries);
   }
 

src/d3d9/d3d9_fixed_function.cpp

@@ -2407,6 +2407,9 @@ namespace dxvk {
 
     m_module.decorateBuiltIn(clipDistArray, spv::BuiltInClipDistance);
 
+    // Always consider clip planes enabled when doing GPL by forcing 6 the quick value.
+    uint32_t clipPlaneCount = m_spec.get(m_module, m_specUbo, SpecClipPlaneCount, 0, 32, m_module.constu32(6));
+
     // Compute clip distances
     for (uint32_t i = 0; i < caps::MaxClipPlanes; i++) {
       std::array<uint32_t, 2> blockMembers = {{
@@ -2421,9 +2424,7 @@ namespace dxvk {
 
       uint32_t distId = m_module.opDot(floatType, worldPos, planeId);
 
-      // Always consider clip planes enabled when doing GPL by forcing a mask of 0xffffffff for the quick value.
-      uint32_t clipPlaneEnabledBit = m_spec.get(m_module, m_specUbo, SpecClipPlaneMask, i, 1, m_module.constu32(0xffffffff));
-      uint32_t clipPlaneEnabled = m_module.opINotEqual(boolType, clipPlaneEnabledBit, m_module.constu32(0));
+      uint32_t clipPlaneEnabled = m_module.opULessThan(boolType, m_module.constu32(i), clipPlaneCount);
 
       uint32_t value = m_module.opSelect(floatType, clipPlaneEnabled, distId, m_module.constf32(0.0f));
 

src/d3d9/d3d9_spec_constants.h

@@ -30,7 +30,7 @@ namespace dxvk {
     SpecDrefClamp,          // 1 bit for 16 PS samplers       | Bits: 16
     SpecFetch4,             // 1 bit for 16 PS samplers       | Bits: 16
 
-    SpecClipPlaneMask,      // 6 bits for 6 clip planes       | Bits : 6
+    SpecClipPlaneCount,     // 3 bits for 6 clip planes       | Bits : 3
 
     SpecConstantCount,
   };
@@ -71,7 +71,7 @@ namespace dxvk {
       { 4, 0,  16 }, // DrefClamp
       { 4, 16, 16 }, // Fetch4
 
-      { 5, 0, 6 },   // ClipPlaneEnabled
+      { 5, 0, 3 },   // ClipPlaneCount
     }};
 
     template <D3D9SpecConstantId Id, typename T>

src/dxso/dxso_compiler.cpp

@@ -3537,6 +3537,9 @@ void DxsoCompiler::emitControlFlowGenericLoop(
     DxsoRegisterValue position;
     position.type = { DxsoScalarType::Float32, 4 };
     position.id = m_module.opLoad(vec4Type, positionPtr);
+
+    // Always consider clip planes enabled when doing GPL by forcing 6 the quick value.
+    uint32_t clipPlaneCount = m_spec.get(m_module, m_specUbo, SpecClipPlaneCount, 0, 32, m_module.constu32(6));
 
     for (uint32_t i = 0; i < caps::MaxClipPlanes; i++) {
       std::array<uint32_t, 2> blockMembers = {{
@@ -3552,9 +3555,7 @@ void DxsoCompiler::emitControlFlowGenericLoop(
 
       DxsoRegisterValue dist = emitDot(position, plane);
 
-      // Always consider clip planes enabled when doing GPL by forcing a mask of 0xffffffff for the quick value.
-      uint32_t clipPlaneEnabledBit = m_spec.get(m_module, m_specUbo, SpecClipPlaneMask, i, 1, m_module.constu32(0xffffffff));
-      uint32_t clipPlaneEnabled = m_module.opINotEqual(boolType, clipPlaneEnabledBit, m_module.constu32(0));
+      uint32_t clipPlaneEnabled = m_module.opULessThan(boolType, m_module.constu32(i), clipPlaneCount);
 
       uint32_t value = m_module.opSelect(floatType, clipPlaneEnabled, dist.id, m_module.constf32(0.0f));