GS: Use integer to determine Z all equal in vertex trace.

Author: TJnotJT

pcsx2/GS/GSState.cpp

@@ -540,6 +540,11 @@ void GSState::DumpVertices(const std::string& filename)
 	v2 = m_vt.m_max.t;
 	file << "\tmax t (x,y,z,w): " << v2.x << DEL << v2.y << DEL << v2.z << DEL << v2.w << std::endl;
 
+	// FIXME: MAKE THIS A NEW PR
+	// file << std::endl;
+	// file << "\teq (r,g,b,a): " << m_vt.m_eq.r << DEL << m_vt.m_eq.g << DEL << m_vt.m_eq.b << DEL << m_vt.m_eq.a << std::endl;
+	// file << "\teq (x,y,z,f): " << m_vt.m_eq.x << DEL << m_vt.m_eq.y << DEL << m_vt.m_eq.z << DEL << m_vt.m_eq.f << std::endl;
+	// file << "\teq (s,t,q)  : " << m_vt.m_eq.s << DEL << m_vt.m_eq.t << DEL << m_vt.m_eq.q << std::endl;
 	file.close();
 }
 

pcsx2/GS/Renderers/Common/GSVertexTrace.cpp

@@ -43,18 +43,8 @@ void GSVertexTrace::Update(const void* vertex, const u16* index, int v_count, in
 		m_max.c.a = 128;
 	}
 
-	// FIXME: This could be moved into the FMM functions
-	m_eq.value = (m_min.c == m_max.c).mask() | ((m_min.p == m_max.p).mask() << 16) | ((m_min.t == m_max.t).mask() << 20);
-
 	m_alpha.valid = false;
 
-	// FIXME: IF we just did the accurate STQ check, we could totally remove this
-	// I'm not sure of the cost. In doubt let's do it only when depth is enabled
-	if (m_state->m_context->TEST.ZTE == 1 && m_state->m_context->TEST.ZTST > ZTST_ALWAYS)
-	{
-		CorrectDepthTrace(vertex, v_count);
-	}
-
 	if (tme)
 	{
 		const GIFRegTEX1& TEX1 = m_state->m_context->TEX1;
@@ -125,49 +115,3 @@ void GSVertexTrace::Update(const void* vertex, const u16* index, int v_count, in
 		}
 	}
 }
-
-void GSVertexTrace::CorrectDepthTrace(const void* vertex, int count)
-{
-	if (m_eq.z == 0)
-		return;
-
-	// FindMinMax isn't accurate for the depth value. Lsb bit is always 0.
-	// The code below will check that depth value is really constant
-	// and will update m_min/m_max/m_eq accordingly
-	//
-	// Really impact Xenosaga3
-	//
-	// Hopefully function is barely called so AVX/SSE will be useless here
-
-	const GSVertex* RESTRICT v = (GSVertex*)vertex;
-
-	const int sprite_step = (m_primclass == GS_SPRITE_CLASS) ? 1 : 0;
-
-	u32 z = v[sprite_step].XYZ.Z;
-
-	if (z & 1)
-	{
-		// Check that first bit is always 1
-		for (int i = sprite_step; i < count; i += (sprite_step + 1))
-		{
-			z &= v[i].XYZ.Z;
-		}
-	}
-	else
-	{
-		// Check that first bit is always 0
-		for (int i = sprite_step; i < count; i += (sprite_step + 1))
-		{
-			z |= v[i].XYZ.Z;
-		}
-	}
-
-	if (z == v[sprite_step].XYZ.Z)
-	{
-		m_eq.z = 1;
-	}
-	else
-	{
-		m_eq.z = 0;
-	}
-}

pcsx2/GS/Renderers/Common/GSVertexTraceFMM.cpp

@@ -11,7 +11,7 @@
 // #define PRINTF(...) \
 // do \
 // { \
-// 	if (GSState::s_n == 4) { \
+// 	if (GSState::s_n == 10) { \
 // 		if (!debug_log) { \
 // 			debug_log = fopen("E:\\bad.log", "w"); \
 // 		} \
@@ -20,8 +20,6 @@
 // 	} \
 // } while (0)
 
-
-
 class CURRENT_ISA::GSVertexTraceFMM
 {
 	template <GS_PRIM_CLASS primclass, u32 iip, u32 tme, u32 fst, u32 color, bool flat_swapped>
@@ -127,6 +125,9 @@ void GSVertexTraceFMM::FindMinMax(GSVertexTrace& vt, const void* vertex, const u
 				GSVector4 stq0 = GSVector4::cast(GSVector4i(v0.m[0]));
 				GSVector4 stq1 = GSVector4::cast(GSVector4i(v1.m[0]));
 
+				// PRINTF("stq0: %f %f %f %f\n", stq0.x, stq0.y, stq0.z, stq0.w);
+				// PRINTF("stq1: %f %f %f %f\n", stq1.x, stq1.y, stq1.z, stq1.w);
+
 				pxAssert(primclass != GS_SPRITE_CLASS || stq0.w == stq1.w);
 
 				GSVector4 q = stq0.wwww(stq1);
@@ -151,7 +152,6 @@ void GSVertexTraceFMM::FindMinMax(GSVertexTrace& vt, const void* vertex, const u
 				tmax = tmax.max(st0.max(st1));
 
 			}
-			// PRINTF("tmin: %f %f, tmax: %f %f\n", tmin.x, tmin.y, tmax.x, tmax.y);
 		}
 
 		GSVector4i xyzf0(v0.m[1]);
@@ -218,8 +218,6 @@ void GSVertexTraceFMM::FindMinMax(GSVertexTrace& vt, const void* vertex, const u
 	vt.m_min.p = (GSVector4(pmin) - o) * s;
 	vt.m_max.p = (GSVector4(pmax) - o) * s;
 
-	// FIXME: We can fully determine m_eq here without having to correct depth trace, etc.
-
 	// Fix signed int conversion
 	vt.m_min.p.z = (float)pmin.U32[2];
 	vt.m_max.p.z = (float)pmax.U32[2];
@@ -254,4 +252,12 @@ void GSVertexTraceFMM::FindMinMax(GSVertexTrace& vt, const void* vertex, const u
 		vt.m_min.c = GSVector4i::zero();
 		vt.m_max.c = GSVector4i::zero();
 	}
+
+
+	// Use pmin==pmax integer comparison to rounding issues for Z
+	// The cast is needed so that the mask operation gives 4 bits (instead of 16 for GSVector4i).
+	const u32 p_eq = GSVector4::cast(pmin == pmax).mask();
+	const u32 t_eq = (vt.m_min.t == vt.m_max.t).mask();
+	const u32 c_eq = (vt.m_min.c == vt.m_max.c).mask();
+	vt.m_eq.value = c_eq | (p_eq << 16) | (t_eq << 20);
 }