Fallback to software transform if draw intersects -W<Z<W

GPU/Common/DrawEngineCommon.cpp

@@ -349,8 +349,8 @@ static bool TestBoundingBoxFast(const float *worldViewProj, const void *vdata, i
 	const int offset = dec->posoff;
 	const s8 *data = (const s8 *)vdata + offset;
 
-	float vpZOffset = gstate.getViewportZCenter();
-	float vpZScale = gstate.getViewportZScale();
+	const float vpZOffset = gstate.getViewportZCenter();
+	const float vpZScale = gstate.getViewportZScale();
 
 	for (int i = 0; i < vertexCount; i++, data += stride) {
 		Vec4F32 objPos;
@@ -366,14 +366,6 @@ static bool TestBoundingBoxFast(const float *worldViewProj, const void *vdata, i
 			break;
 		}
 		Vec4F32 clipPos = objPos.AsVec3ByMatrix44(worldViewProjMat) & vertexMask;
-		const float projZ = vpZScale * clipPos[2] / clipPos[3];
-		if (projZ < minProjZ) {
-			minProjZ = projZ;
-		}
-		if (projZ > maxProjZ) {  // else ruins the minss/maxss optimization.
-			maxProjZ = projZ;
-		}
-
 		Vec4F32 posW = clipPos.ShuffleWWWW();
 		if (cull) {
 			Vec4F32 posXY = clipPos.ShuffleXXYY();
@@ -386,9 +378,17 @@ static bool TestBoundingBoxFast(const float *worldViewProj, const void *vdata, i
 		if (cull) {
 			insideMaskZ |= planeDistZ.CompareGe(Vec4F32::Zero());
 		}
+		const float projZ = vpZScale * clipPos[2] / clipPos[3];
+		if (projZ < minProjZ) {
+			minProjZ = projZ;
+		}
+		if (projZ > maxProjZ) {  // else ruins the minss/maxss optimization.
+			maxProjZ = projZ;
+		}
 	}
 
 	if (!cull || (AllCompareBitsSet(insideMaskXY) && AllCompareBitsSet(insideMaskZ))) {
+		depths->valid = true;
 		depths->hitClipSpaceZW = AnyCompareBitsSet(anyOutsideMaskZ);
 		// Before checking later, we apply a viewport to the projZ, so we can later compare against minZ/maxZ or the outer bounds.
 		// But to save operations we don't do it here.
@@ -449,7 +449,7 @@ bool DrawEngineCommon::TestBoundingBoxFast(const float *worldViewProj, const voi
 		applyViewport.wx = -(maxViewport.x + minViewport.x) * viewportInvSize.x;
 		applyViewport.wy = -(maxViewport.y + minViewport.y) * viewportInvSize.y;
 
-		// TODO: Optimize.
+		// TODO: Optimize. It's possible to scale/offset a matrix in a quicker way.
 		Matrix4ByMatrix4(mtx, worldViewProj, applyViewport.m);
 		worldViewProj = mtx;
 	}

GPU/Common/DrawEngineCommon.h

@@ -73,10 +73,15 @@ struct alignas(16) Plane8 {
 };
 
 struct BoundingDepths {
+	bool valid = false;
 	bool hitClipSpaceZW = false;
 	float minProjZ = FLT_MAX;
 	float maxProjZ = -FLT_MAX;
 	void Merge(const BoundingDepths &other) {
+		if (!valid) {
+			*this = other;
+			return;
+		}
 		hitClipSpaceZW |= other.hitClipSpaceZW;
 		if (other.minProjZ < minProjZ) {
 			minProjZ = other.minProjZ;
@@ -380,4 +385,6 @@ protected:
 	std::vector<DepthDraw> depthDraws_;
 
 	double rasterTimeStart_ = 0.0;
+
+	bool lastUseHwTransform_ = true;
 };

GPU/GPU.h

@@ -98,6 +98,7 @@ struct GPUStatsPerFrame {
 	int numDrawSyncs;
 	int numListSyncs;
 	int numFlushes;
+	int numSoftTransformedDraws;
 	int numSoftClippedTriangles;
 	int numBBOXJumps;
 	int numVertsSubmitted;

GPU/GPUCommonHW.cpp

@@ -1805,7 +1805,7 @@ void GPUCommonHW::FormatGPUStatsCommon(StringWriter &w) {
 	w.F(
 		"DL processing time: %0.2f ms, %d drawsync, %d listsync\n"
 		"Draw: %d (%d dec, %d culled), flushes %d, clears %d, bbox jumps %d\n"
-		"Vertices: %d dec: %d drawn: %d clipped tris: %d\n"
+		"%d soft. Vertices: %d dec: %d drawn: %d clipped tris: %d\n"
 		"FBOs active: %d (evaluations: %d, created %d)\n"
 		"Textures: %d, dec: %d, invalidated: %d, hashed: %d kB, clut %d\n"
 		"readbacks %d (%d non-block), upload %d (cached %d), depal %d\n"
@@ -1824,6 +1824,7 @@ void GPUCommonHW::FormatGPUStatsCommon(StringWriter &w) {
 		gpuStats.perFrame.numFlushes,
 		gpuStats.perFrame.numClears,
 		gpuStats.perFrame.numBBOXJumps,
+		gpuStats.perFrame.numSoftTransformedDraws,
 		gpuStats.perFrame.numVertsSubmitted,
 		gpuStats.perFrame.numVertsDecoded,
 		gpuStats.perFrame.numUncachedVertsDrawn,

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -242,6 +242,18 @@ void DrawEngineVulkan::Flush() {
 	}
 	bool useHWTransform = CanUseHardwareTransform(prim) && provokingVertexOk;
 
+	if (useHWTransform && boundingDepths_.valid) {
+		if (boundingDepths_.hitClipSpaceZW) {
+			// Revert to software so we can clip more accurately.
+			useHWTransform = false;
+		}
+	}
+
+	if (useHWTransform != lastUseHwTransform_) {
+		gstate_c.Dirty(DIRTY_VERTEXSHADER_STATE);
+		lastUseHwTransform_ = useHWTransform;
+	}
+
 	// TODO: Here we can check depths_ to see if we need to fall back to software transform for clipping.
 
 	// The optimization to avoid indexing isn't really worth it on Vulkan since it means creating more pipelines.
@@ -377,6 +389,8 @@ void DrawEngineVulkan::Flush() {
 			DepthRasterSubmitRaw(prim, dec_, dec_->VertexType(), vertexCount);
 		}
 	} else {
+		gpuStats.perFrame.numSoftTransformedDraws++;
+
 		PROFILE_THIS_SCOPE("soft");
 		const VertexDecoder *swDec = dec_;
 		if (swDec->nweights != 0) {