Convert the BBOX/BJUMP culling to use the worldviewproj matrix and a screen space check.

GPU/Common/DrawEngineCommon.cpp

@@ -163,47 +163,6 @@ void DrawEngineCommon::DispatchSubmitImm(GEPrimitiveType prim, TransformedVertex
 	}
 }
 
-// Gated by DIRTY_CULL_PLANES
-void DrawEngineCommon::UpdatePlanes(const float viewproj[16]) {
-	// Next, we need to apply viewport, scissor, region, and even offset - but only for X/Y.
-	// Note that the PSP does not clip against the viewport.
-	const Vec2f baseOffset = Vec2f(gstate.getOffsetX(), gstate.getOffsetY());
-	// Region1 (rate) is used as an X1/Y1 here, matching PSP behavior.
-	minOffset_ = baseOffset + Vec2f(std::max(gstate.getRegionRateX() - 0x100, gstate.getScissorX1()), std::max(gstate.getRegionRateY() - 0x100, gstate.getScissorY1())) - Vec2f(1.0f, 1.0f);
-	maxOffset_ = baseOffset + Vec2f(std::min(gstate.getRegionX2(), gstate.getScissorX2()), std::min(gstate.getRegionY2(), gstate.getScissorY2())) + Vec2f(1.0f, 1.0f);
-
-	// Let's not handle these special cases in the fast culler.
-	offsetOutsideEdge_ = maxOffset_.x >= 4096.0f || minOffset_.x < 1.0f || minOffset_.y < 1.0f || maxOffset_.y >= 4096.0f;
-
-	// Now let's apply the viewport to our scissor/region + offset range.
-	Vec2f inverseViewportScale = Vec2f(1.0f / gstate.getViewportXScale(), 1.0f / gstate.getViewportYScale());
-	Vec2f minViewport = (minOffset_ - Vec2f(gstate.getViewportXCenter(), gstate.getViewportYCenter())) * inverseViewportScale;
-	Vec2f maxViewport = (maxOffset_ - Vec2f(gstate.getViewportXCenter(), gstate.getViewportYCenter())) * inverseViewportScale;
-
-	Vec2f viewportInvSize = Vec2f(1.0f / (maxViewport.x - minViewport.x), 1.0f / (maxViewport.y - minViewport.y));
-
-	Lin::Matrix4x4 applyViewport{};
-	// Scale to the viewport's size.
-	applyViewport.xx = 2.0f * viewportInvSize.x;
-	applyViewport.yy = 2.0f * viewportInvSize.y;
-	applyViewport.zz = 1.0f;
-	applyViewport.ww = 1.0f;
-	// And offset to the viewport's centers.
-	applyViewport.wx = -(maxViewport.x + minViewport.x) * viewportInvSize.x;
-	applyViewport.wy = -(maxViewport.y + minViewport.y) * viewportInvSize.y;
-
-	float mtx[16];
-	Matrix4ByMatrix4(mtx, viewproj, applyViewport.m);
-	// I'm sure there's some fairly optimized way to set these. If we make a version of Matrix4ByMatrix4
-	// that returns a transpose, it looks like these will be more straightforward.
-	planes_.Set(0, mtx[3] - mtx[0], mtx[7] - mtx[4], mtx[11] - mtx[8],  mtx[15] - mtx[12]);  // Right
-	planes_.Set(1, mtx[3] + mtx[0], mtx[7] + mtx[4], mtx[11] + mtx[8],  mtx[15] + mtx[12]);  // Left
-	planes_.Set(2, mtx[3] + mtx[1], mtx[7] + mtx[5], mtx[11] + mtx[9],  mtx[15] + mtx[13]);  // Bottom
-	planes_.Set(3, mtx[3] - mtx[1], mtx[7] - mtx[5], mtx[11] - mtx[9],  mtx[15] - mtx[13]);  // Top
-	planes_.Set(4, mtx[3] + mtx[2], mtx[7] + mtx[6], mtx[11] + mtx[10], mtx[15] + mtx[14]);  // Near
-	planes_.Set(5, mtx[3] - mtx[2], mtx[7] - mtx[6], mtx[11] - mtx[10], mtx[15] - mtx[14]);  // Far
-}
-
 // This code has plenty of potential for optimization.
 //
 // It does the simplest and safest test possible: If all points of a bbox is outside a single of
@@ -229,14 +188,6 @@ bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int
 	if (gstate_c.Use(GPU_USE_VIRTUAL_REALITY))
 		return true;
 
-	// Due to world matrix updates per "thing", this isn't quite as effective as it could be if we did world transform
-	// in here as well. Though, it still does cut down on a lot of updates in Tekken 6.
-	if (gstate_c.IsDirty(DIRTY_CULL_PLANES)) {
-		UpdatePlanes(gstate_c.viewproj);
-		gpuStats.numPlaneUpdates++;
-		gstate_c.Clean(DIRTY_CULL_PLANES);
-	}
-
 	// Try to skip NormalizeVertices if it's pure positions. No need to bother with a vertex decoder
 	// and a large vertex format.
 	if ((vertType & 0xFFFFFF) == GE_VTYPE_POS_FLOAT && !inds) {
@@ -301,51 +252,60 @@ bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int
 		}
 	}
 
-	// Pretransform the verts in-place so we don't have to do it inside the loop.
-	// We do this differently in the fast version below since we skip the max/minOffset checks there
-	// making it easier to get the whole thing ready for SIMD.
+	// TODO: How accurate should we be?
+	int insideCount[5] = {0};
 	for (int i = 0; i < vertexCount; i++) {
-		float worldpos[3];
-		Vec3ByMatrix43(worldpos, &verts[i * 3], gstate.worldMatrix);
-		memcpy(&verts[i * 3], worldpos, 12);
-	}
+		// Complete the transform to see if the vertex should be ignored. Not sure if we need to go to these lengths...
+		const float *objpos = verts + i * 3;
 
-	// Note: near/far are not checked without clamp/clip enabled, so we skip those planes.
-	int totalPlanes = gstate.isDepthClampEnabled() ? 6 : 4;
-	for (int plane = 0; plane < totalPlanes; plane++) {
-		int inside = 0;
-		int out = 0;
-		for (int i = 0; i < vertexCount; i++) {
-			// Test against the frustum planes, and count.
-			// TODO: We should test 4 vertices at a time using SIMD.
-			// I guess could also test one vertex against 4 planes at a time, though a lot of waste at the common case of 6.
-			const float *worldpos = verts + i * 3;
-			float value = planes_.Test(plane, worldpos);
-			if (value <= -FLT_EPSILON)  // Not sure why we use exactly this value. Probably '< 0' would do.
-				out++;
-			else
-				inside++;
+		float projpos[4]{};
+		Vec3ByMatrix44(projpos, objpos, gstate_c.worldviewproj);
+
+		if (projpos[2] > -projpos[3] && projpos[2] < projpos[3]) {
+			insideCount[4]++;
 		}
 
-		// No vertices inside this one plane? Don't need to draw.
-		if (inside == 0) {
-			// All out - but check for X and Y if the offset was near the cullbox edge.
-			bool outsideEdge = false;
-			switch (plane) {
-			case 0: outsideEdge = maxOffset_.x >= 4096.0f; break;
-			case 1: outsideEdge = minOffset_.x < 1.0f; break;
-			case 2: outsideEdge = minOffset_.y < 1.0f; break;
-			case 3: outsideEdge = maxOffset_.y >= 4096.0f; break;
-			}
+		float invW = 1.0f / projpos[3];
+		const float screenpos[3] = {
+			projpos[0] * gstate.getViewportXScale() * invW + gstate.getViewportXCenter(),
+			projpos[1] * gstate.getViewportYScale() * invW + gstate.getViewportYCenter(),
+			projpos[2] * gstate.getViewportZScale() * invW + gstate.getViewportZCenter(),
+		};
 
-			// Only consider this outside if offset + scissor/region is fully inside the cullbox.
-			if (!outsideEdge)
-				return false;
+		const float drawX = screenpos[0] - gstate.getOffsetX();
+		const float drawY = screenpos[1] - gstate.getOffsetY();
+
+		const int regionX2 = gstate.getRegionX2();
+		const int regionY2 = gstate.getRegionY2();
+		if (drawX > 0.0f) {
+			insideCount[0]++;
+		}
+		if (drawX < regionX2) {
+			insideCount[1]++;
+		}
+		if (drawY > 0.0f) {
+			insideCount[2]++;
+		}
+		if (drawY < regionY2) {
+			insideCount[3]++;
 		}
-
-		// Any out. For testing that the planes are in the right locations.
-		// if (out != 0) return false;
 	}
+
+#if 0
+	// For debugging, the exclusive check. This should make it obvious where the culling borders are in screen space.
+	for (int i = 0; i < ARRAY_SIZE(insideCount); i++) {
+		if (insideCount[i] != vertexCount) {
+			return false;
+		}
+	}
+#endif
+
+	for (int i = 0; i < ARRAY_SIZE(insideCount); i++) {
+		if (insideCount[i] == 0) {
+			return false;
+		}
+	}
+
 	return true;
 }
 

GPU/Common/DrawEngineCommon.h

@@ -169,7 +169,6 @@ public:
 
 protected:
 	virtual bool UpdateUseHWTessellation(bool enabled) const { return enabled; }
-	void UpdatePlanes(const float viewproj[16]);
 
 	void DecodeVerts(const VertexDecoder *dec, u8 *dest);
 	int DecodeInds();
@@ -352,7 +351,6 @@ protected:
 	TessellationDataTransfer *tessDataTransfer = nullptr;
 
 	// Culling
-	Plane8 planes_{};
 	Vec2f minOffset_{};
 	Vec2f maxOffset_{};
 	bool offsetOutsideEdge_ = false;

GPU/Common/ShaderCommon.h

@@ -102,7 +102,7 @@ enum : uint64_t {
 	DIRTY_VIEW_PROJ_MATRIX = 1ULL << 40,
 	DIRTY_WORLD_VIEW_PROJ_MATRIX = 1ULL << 41,
 	// Free non-uniform bit 42.
-	DIRTY_CULL_PLANES = 1ULL << 43,
+	// Free non-uniform bit 43.
 	DIRTY_FRAMEBUF = 1ULL << 44,
 	DIRTY_TEXTURE_IMAGE = 1ULL << 45,  // Means that the definition of the texture image has changed (address, stride etc), and we need to look up again.
 	DIRTY_TEXTURE_PARAMS = 1ULL << 46,

GPU/GPU.h

@@ -90,7 +90,6 @@ struct GPUStatistics {
 		numTextureDataBytesHashed = 0;
 		numFlushes = 0;
 		numBBOXJumps = 0;
-		numPlaneUpdates = 0;
 		numTexturesDecoded = 0;
 		numFramebufferEvaluations = 0;
 		numFBOsCreated = 0;
@@ -132,7 +131,6 @@ struct GPUStatistics {
 	int numListSyncs;
 	int numFlushes;
 	int numBBOXJumps;
-	int numPlaneUpdates;
 	int numVertsSubmitted;
 	int numVertsDecoded;
 	int numUncachedVertsDrawn;

GPU/GPUCommonHW.cpp

@@ -191,12 +191,12 @@ const CommonCommandTableEntry commonCommandTable[] = {
 	{ GE_CMD_ANTIALIASENABLE, FLAG_FLUSHBEFOREONCHANGE },
 
 	// Viewport.
-	{ GE_CMD_OFFSETX, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
-	{ GE_CMD_OFFSETY, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
-	{ GE_CMD_VIEWPORTXSCALE, FLAG_FLUSHBEFOREONCHANGE,  DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULL_PLANES },
-	{ GE_CMD_VIEWPORTYSCALE, FLAG_FLUSHBEFOREONCHANGE,  DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULL_PLANES },
-	{ GE_CMD_VIEWPORTXCENTER, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULL_PLANES },
-	{ GE_CMD_VIEWPORTYCENTER, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULL_PLANES },
+	{ GE_CMD_OFFSETX, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
+	{ GE_CMD_OFFSETY, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
+	{ GE_CMD_VIEWPORTXSCALE, FLAG_FLUSHBEFOREONCHANGE,  DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
+	{ GE_CMD_VIEWPORTYSCALE, FLAG_FLUSHBEFOREONCHANGE,  DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
+	{ GE_CMD_VIEWPORTXCENTER, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
+	{ GE_CMD_VIEWPORTYCENTER, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
 	{ GE_CMD_VIEWPORTZSCALE, FLAG_FLUSHBEFOREONCHANGE,  DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_DEPTHRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
 	{ GE_CMD_VIEWPORTZCENTER, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_CULLRANGE | DIRTY_DEPTHRANGE | DIRTY_PROJMATRIX | DIRTY_VIEWPORTSCISSOR_STATE },
 	{ GE_CMD_DEPTHCLAMPENABLE, FLAG_FLUSHBEFOREONCHANGE, DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_RASTER_STATE },
@@ -206,12 +206,12 @@ const CommonCommandTableEntry commonCommandTable[] = {
 	{ GE_CMD_MAXZ, FLAG_FLUSHBEFOREONCHANGE, DIRTY_DEPTHRANGE | DIRTY_RASTER_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
 
 	// Region
-	{ GE_CMD_REGION1, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
-	{ GE_CMD_REGION2, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
+	{ GE_CMD_REGION1, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
+	{ GE_CMD_REGION2, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
 
 	// Scissor
-	{ GE_CMD_SCISSOR1, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
-	{ GE_CMD_SCISSOR2, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE | DIRTY_CULL_PLANES },
+	{ GE_CMD_SCISSOR1, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
+	{ GE_CMD_SCISSOR2, FLAG_FLUSHBEFOREONCHANGE, DIRTY_FRAMEBUF | DIRTY_TEXTURE_PARAMS | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE },
 
 	// Lighting base colors
 	{ GE_CMD_AMBIENTCOLOR, FLAG_FLUSHBEFOREONCHANGE, DIRTY_AMBIENT },
@@ -1586,7 +1586,7 @@ void GPUCommonHW::Execute_ViewMtxNum(u32 op, u32 diff) {
 			if (dst[i] != newVal) {
 				Flush();
 				dst[i] = newVal;
-				gstate_c.Dirty(DIRTY_VIEWMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX | DIRTY_CULL_PLANES);
+				gstate_c.Dirty(DIRTY_VIEWMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 			}
 			if (++i >= end) {
 				break;
@@ -1609,7 +1609,7 @@ void GPUCommonHW::Execute_ViewMtxData(u32 op, u32 diff) {
 	if (num < 12 && newVal != ((const u32 *)gstate.viewMatrix)[num]) {
 		Flush();
 		((u32 *)gstate.viewMatrix)[num] = newVal;
-		gstate_c.Dirty(DIRTY_VIEWMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX | DIRTY_CULL_PLANES);
+		gstate_c.Dirty(DIRTY_VIEWMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 	}
 	num++;
 	gstate.viewmtxnum = (GE_CMD_VIEWMATRIXNUMBER << 24) | (num & 0x00FFFFFF);
@@ -1639,7 +1639,7 @@ void GPUCommonHW::Execute_ProjMtxNum(u32 op, u32 diff) {
 			if (dst[i] != newVal) {
 				Flush();
 				dst[i] = newVal;
-				gstate_c.Dirty(DIRTY_PROJMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX | DIRTY_CULL_PLANES);
+				gstate_c.Dirty(DIRTY_PROJMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 			}
 			if (++i >= end) {
 				break;
@@ -1662,7 +1662,7 @@ void GPUCommonHW::Execute_ProjMtxData(u32 op, u32 diff) {
 	if (num < 16 && newVal != ((const u32 *)gstate.projMatrix)[num]) {
 		Flush();
 		((u32 *)gstate.projMatrix)[num] = newVal;
-		gstate_c.Dirty(DIRTY_PROJMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX | DIRTY_CULL_PLANES);
+		gstate_c.Dirty(DIRTY_PROJMATRIX | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 	}
 	num++;
 	if (num <= 16)
@@ -1822,7 +1822,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 	float vertexAverageCycles = gpuStats.numVertsSubmitted > 0 ? (float)gpuStats.vertexGPUCycles / (float)gpuStats.numVertsSubmitted : 0.0f;
 	return snprintf(buffer, size,
 		"DL processing time: %0.2f ms, %d drawsync, %d listsync\n"
-		"Draw: %d (%d dec, %d culled), flushes %d, clears %d, bbox jumps %d (%d updates)\n"
+		"Draw: %d (%d dec, %d culled), flushes %d, clears %d, bbox jumps %d\n"
 		"Vertices: %d dec: %d drawn: %d\n"
 		"FBOs active: %d (evaluations: %d, created %d)\n"
 		"Textures: %d, dec: %d, invalidated: %d, hashed: %d kB, clut %d\n"
@@ -1842,7 +1842,6 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		gpuStats.numFlushes,
 		gpuStats.numClears,
 		gpuStats.numBBOXJumps,
-		gpuStats.numPlaneUpdates,
 		gpuStats.numVertsSubmitted,
 		gpuStats.numVertsDecoded,
 		gpuStats.numUncachedVertsDrawn,

GPU/GPUState.cpp

@@ -116,7 +116,7 @@ void GPUgstate::Reset() {
 
 	savedContextVersion = 1;
 
-	gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+	gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 }
 
 void GPUgstate::Save(u32_le *ptr) {
@@ -248,7 +248,7 @@ void GPUgstate::Restore(const u32_le *ptr) {
 	if (gpu)
 		gpu->ResetMatrices();
 
-	gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+	gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 }
 
 bool vertTypeIsSkinningEnabled(u32 vertType) {
@@ -359,7 +359,7 @@ void GPUStateCache::DoState(PointerWrap &p) {
 	}
 
 	if (p.GetMode() == PointerWrap::MODE_READ)
-		gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+		gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 }
 
 static const char *const g_gpuUseFlagNames[32] = {

GPU/Software/SoftGpu.cpp

@@ -1038,7 +1038,6 @@ void SoftGPU::Execute_FramebufFormat(u32 op, u32 diff) {
 }
 
 void SoftGPU::Execute_BoundingBox(u32 op, u32 diff) {
-	gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 	GPUCommon::Execute_BoundingBox(op, diff);
 }
 
@@ -1094,7 +1093,7 @@ void SoftGPU::Execute_WorldMtxData(u32 op, u32 diff) {
 		if (newVal != *target) {
 			*target = newVal;
 			dirtyFlags_ |= SoftDirty::TRANSFORM_MATRIX;
-			gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+			gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 		}
 	}
 
@@ -1115,7 +1114,7 @@ void SoftGPU::Execute_ViewMtxData(u32 op, u32 diff) {
 		if (newVal != *target) {
 			*target = newVal;
 			dirtyFlags_ |= SoftDirty::TRANSFORM_MATRIX;
-			gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+			gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 		}
 	}
 
@@ -1136,7 +1135,7 @@ void SoftGPU::Execute_ProjMtxData(u32 op, u32 diff) {
 		if (newVal != *target) {
 			*target = newVal;
 			dirtyFlags_ |= SoftDirty::TRANSFORM_MATRIX;
-			gstate_c.Dirty(DIRTY_CULL_PLANES | DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
+			gstate_c.Dirty(DIRTY_WORLD_VIEW_PROJ_MATRIX | DIRTY_VIEW_PROJ_MATRIX);
 		}
 	}