Software: Fix zfreeze with CullMode::All

Source/Core/VideoBackends/Software/Clipper.cpp

@@ -289,10 +289,42 @@ void ProcessTriangle(OutputVertexData* v0, OutputVertexData* v1, OutputVertexDat
 {
   INCSTAT(g_stats.this_frame.num_triangles_in)
 
-  bool backface;
-
-  if (!CullTest(v0, v1, v2, backface))
+  if (IsTriviallyRejected(v0, v1, v2))
+  {
+    INCSTAT(g_stats.this_frame.num_triangles_rejected)
+    // NOTE: The slope used by zfreeze shouldn't be updated if the triangle is
+    // trivially rejected during clipping
     return;
+  }
+
+  bool backface = IsBackface(v0, v1, v2);
+
+  if (!backface)
+  {
+    if (bpmem.genMode.cullmode == CullMode::Back || bpmem.genMode.cullmode == CullMode::All)
+    {
+      // cull frontfacing - we still need to update the slope for zfreeze
+      PerspectiveDivide(v0);
+      PerspectiveDivide(v1);
+      PerspectiveDivide(v2);
+      Rasterizer::UpdateZSlope(v0, v1, v2);
+      INCSTAT(g_stats.this_frame.num_triangles_culled)
+      return;
+    }
+  }
+  else
+  {
+    if (bpmem.genMode.cullmode == CullMode::Front || bpmem.genMode.cullmode == CullMode::All)
+    {
+      // cull backfacing - we still need to update the slope for zfreeze
+      PerspectiveDivide(v0);
+      PerspectiveDivide(v2);
+      PerspectiveDivide(v1);
+      Rasterizer::UpdateZSlope(v0, v2, v1);
+      INCSTAT(g_stats.this_frame.num_triangles_culled)
+      return;
+    }
+  }
 
   int indices[NUM_INDICES] = {0,         1,         2,         SKIP_FLAG, SKIP_FLAG, SKIP_FLAG,
                               SKIP_FLAG, SKIP_FLAG, SKIP_FLAG, SKIP_FLAG, SKIP_FLAG, SKIP_FLAG,
@@ -461,19 +493,18 @@ void ProcessPoint(OutputVertexData* center)
   Rasterizer::DrawTriangleFrontFace(&ur, &lr, &ul);
 }
 
-bool CullTest(const OutputVertexData* v0, const OutputVertexData* v1, const OutputVertexData* v2,
-              bool& backface)
+bool IsTriviallyRejected(const OutputVertexData* v0, const OutputVertexData* v1,
+                         const OutputVertexData* v2)
 {
   int mask = CalcClipMask(v0);
   mask &= CalcClipMask(v1);
   mask &= CalcClipMask(v2);
 
-  if (mask)
-  {
-    INCSTAT(g_stats.this_frame.num_triangles_rejected)
-    return false;
-  }
+  return mask != 0;
+}
 
+bool IsBackface(const OutputVertexData* v0, const OutputVertexData* v1, const OutputVertexData* v2)
+{
   float x0 = v0->projectedPosition.x;
   float x1 = v1->projectedPosition.x;
   float x2 = v2->projectedPosition.x;
@@ -486,29 +517,14 @@ bool CullTest(const OutputVertexData* v0, const OutputVertexData* v1, const Outp
 
   float normalZDir = (x0 * w2 - x2 * w0) * y1 + (x2 * y0 - x0 * y2) * w1 + (y2 * w0 - y0 * w2) * x1;
 
-  backface = normalZDir <= 0.0f;
+  bool backface = normalZDir <= 0.0f;
   // Jimmie Johnson's Anything with an Engine has a positive viewport, while other games have a
   // negative viewport.  The positive viewport does not require vertices to be vertically mirrored,
   // but the backface test does need to be inverted for things to be drawn.
   if (xfmem.viewport.ht > 0)
     backface = !backface;
 
-  // TODO: Are these tests / the definition of backface above backwards?
-  if ((bpmem.genMode.cullmode == CullMode::Back || bpmem.genMode.cullmode == CullMode::All) &&
-      !backface)  // cull frontfacing
-  {
-    INCSTAT(g_stats.this_frame.num_triangles_culled)
-    return false;
-  }
-
-  if ((bpmem.genMode.cullmode == CullMode::Front || bpmem.genMode.cullmode == CullMode::All) &&
-      backface)  // cull backfacing
-  {
-    INCSTAT(g_stats.this_frame.num_triangles_culled)
-    return false;
-  }
-
-  return true;
+  return backface;
 }
 
 void PerspectiveDivide(OutputVertexData* vertex)

Source/Core/VideoBackends/Software/Clipper.h

@@ -15,8 +15,10 @@ void ProcessLine(OutputVertexData* v0, OutputVertexData* v1);
 
 void ProcessPoint(OutputVertexData* v);
 
-bool CullTest(const OutputVertexData* v0, const OutputVertexData* v1, const OutputVertexData* v2,
-              bool& backface);
+bool IsTriviallyRejected(const OutputVertexData* v0, const OutputVertexData* v1,
+                         const OutputVertexData* v2);
+
+bool IsBackface(const OutputVertexData* v0, const OutputVertexData* v1, const OutputVertexData* v2);
 
 void PerspectiveDivide(OutputVertexData* vertex);
 }  // namespace Clipper

Source/Core/VideoBackends/Software/Rasterizer.cpp

@@ -301,11 +301,27 @@ static void BuildBlock(s32 blockX, s32 blockY)
   }
 }
 
+void UpdateZSlope(const OutputVertexData* v0, const OutputVertexData* v1,
+                  const OutputVertexData* v2)
+{
+  if (!bpmem.genMode.zfreeze)
+  {
+    const s32 X1 = iround(16.0f * v0->screenPosition[0]) - 9;
+    const s32 Y1 = iround(16.0f * v0->screenPosition[1]) - 9;
+    const SlopeContext ctx(v0, v1, v2, (X1 + 0xF) >> 4, (Y1 + 0xF) >> 4);
+    ZSlope = Slope(v0->screenPosition.z, v1->screenPosition.z, v2->screenPosition.z, ctx);
+  }
+}
+
 void DrawTriangleFrontFace(const OutputVertexData* v0, const OutputVertexData* v1,
                            const OutputVertexData* v2)
 {
   INCSTAT(g_stats.this_frame.num_triangles_drawn);
 
+  // The zslope should be updated now, even if the triangle is rejected by the scissor test, as
+  // zfreeze depends on it
+  UpdateZSlope(v0, v1, v2);
+
   // adapted from http://devmaster.net/posts/6145/advanced-rasterization
 
   // 28.4 fixed-pou32 coordinates. rounded to nearest and adjusted to match hardware output
@@ -370,22 +386,13 @@ void DrawTriangleFrontFace(const OutputVertexData* v0, const OutputVertexData* v
   if (minx >= maxx || miny >= maxy)
     return;
 
-  // Set up slopes
+  // Set up the remaining slopes
   const SlopeContext ctx(v0, v1, v2, (X1 + 0xF) >> 4, (Y1 + 0xF) >> 4);
 
   float w[3] = {1.0f / v0->projectedPosition.w, 1.0f / v1->projectedPosition.w,
                 1.0f / v2->projectedPosition.w};
   WSlope = Slope(w[0], w[1], w[2], ctx);
 
-  // TODO: The zfreeze emulation is not quite correct, yet!
-  // Many things might prevent us from reaching this line (culling, clipping, scissoring).
-  // However, the zslope is always guaranteed to be calculated unless all vertices are trivially
-  // rejected during clipping!
-  // We're currently sloppy at this since we abort early if any of the culling/clipping/scissoring
-  // tests fail.
-  if (!bpmem.genMode.zfreeze)
-    ZSlope = Slope(v0->screenPosition.z, v1->screenPosition.z, v2->screenPosition.z, ctx);
-
   for (unsigned int i = 0; i < bpmem.genMode.numcolchans; i++)
   {
     for (int comp = 0; comp < 4; comp++)

Source/Core/VideoBackends/Software/Rasterizer.h

@@ -11,6 +11,8 @@ namespace Rasterizer
 {
 void Init();
 
+void UpdateZSlope(const OutputVertexData* v0, const OutputVertexData* v1,
+                  const OutputVertexData* v2);
 void DrawTriangleFrontFace(const OutputVertexData* v0, const OutputVertexData* v1,
                            const OutputVertexData* v2);
 

Source/Core/VideoBackends/Software/SWVertexLoader.cpp

@@ -32,6 +32,13 @@ SWVertexLoader::SWVertexLoader() = default;
 
 SWVertexLoader::~SWVertexLoader() = default;
 
+DataReader SWVertexLoader::PrepareForAdditionalData(OpcodeDecoder::Primitive primitive, u32 count,
+                                                    u32 stride, bool cullall)
+{
+  // The software renderer needs cullall to be false for zfreeze to work
+  return VertexManagerBase::PrepareForAdditionalData(primitive, count, stride, false);
+}
+
 void SWVertexLoader::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
 {
   DebugUtil::OnObjectBegin();

Source/Core/VideoBackends/Software/SWVertexLoader.h

@@ -19,6 +19,9 @@ public:
   SWVertexLoader();
   ~SWVertexLoader();
 
+  DataReader PrepareForAdditionalData(OpcodeDecoder::Primitive primitive, u32 count, u32 stride,
+                                      bool cullall) override;
+
 protected:
   void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 

Source/Core/VideoCommon/VertexManagerBase.h

@@ -99,8 +99,8 @@ public:
 
   PrimitiveType GetCurrentPrimitiveType() const { return m_current_primitive_type; }
   void AddIndices(OpcodeDecoder::Primitive primitive, u32 num_vertices);
-  DataReader PrepareForAdditionalData(OpcodeDecoder::Primitive primitive, u32 count, u32 stride,
-                                      bool cullall);
+  virtual DataReader PrepareForAdditionalData(OpcodeDecoder::Primitive primitive, u32 count,
+                                              u32 stride, bool cullall);
   void FlushData(u32 count, u32 stride);
 
   void Flush();