Merge pull request #17055 from unknownbrackets/depth-zero-scale

GPU: Correct depth clip/cull for zero scale

Author: hrydgard

GPU/Common/ShaderUniforms.cpp

@@ -252,14 +252,15 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool flipView
 
 		// These are just the reverse of the formulas in GPUStateUtils.
 		float halfActualZRange = gstate_c.vpDepthScale != 0.0f ? vpZScale / gstate_c.vpDepthScale : 0.0f;
+		float inverseDepthScale = gstate_c.vpDepthScale != 0.0f ? 1.0f / gstate_c.vpDepthScale : 0.0f;
 		float minz = -((gstate_c.vpZOffset * halfActualZRange) - vpZCenter) - halfActualZRange;
 		float viewZScale = halfActualZRange * 2.0f;
 		float viewZCenter = minz;
 
 		ub->depthRange[0] = viewZScale;
 		ub->depthRange[1] = viewZCenter;
 		ub->depthRange[2] = gstate_c.vpZOffset * 0.5f + 0.5f;
-		ub->depthRange[3] = 2.0f * (1.0f / gstate_c.vpDepthScale);
+		ub->depthRange[3] = 2.0f * inverseDepthScale;
 	}
 
 	if (dirtyUniforms & DIRTY_CULLRANGE) {

GPU/Common/VertexShaderGenerator.cpp

@@ -1350,7 +1350,8 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 		}
 		if (gstate_c.Use(GPU_USE_CULL_DISTANCE)) {
 			// Cull any triangle fully outside in the same direction when depth clamp enabled.
-			WRITE(p, "  if (u_cullRangeMin.w > 0.0) {\n");
+			// We check u_depthRange in case depthScale was zero - in that case we can't work out the cull distance.
+			WRITE(p, "  if (u_cullRangeMin.w > 0.0 && u_depthRange.w != 0.0f) {\n");
 			WRITE(p, "    %sgl_CullDistance%s = projPos.z - u_cullRangeMin.z;\n", compat.vsOutPrefix, cull0);
 			WRITE(p, "    %sgl_CullDistance%s = u_cullRangeMax.z - projPos.z;\n", compat.vsOutPrefix, cull1);
 			WRITE(p, "  } else {\n");

GPU/Directx9/ShaderManagerDX9.cpp

@@ -466,7 +466,7 @@ void ShaderManagerDX9::VSUpdateUniforms(u64 dirtyUniforms) {
 		float minz = -((gstate_c.vpZOffset * halfActualZRange) - vpZCenter) - halfActualZRange;
 		float viewZScale = halfActualZRange * 2.0f;
 		float viewZCenter = minz;
-		float reverseScale = 2.0f * (1.0f / gstate_c.vpDepthScale);
+		float reverseScale = gstate_c.vpDepthScale != 0.0f ? 2.0f * (1.0f / gstate_c.vpDepthScale) : 0.0f;
 		float reverseTranslate = gstate_c.vpZOffset * 0.5f + 0.5f;
 
 		float data[4] = { viewZScale, viewZCenter, reverseTranslate, reverseScale };

GPU/GLES/ShaderManagerGLES.cpp

@@ -589,6 +589,7 @@ void LinkedShader::UpdateUniforms(const ShaderID &vsid, bool useBufferedRenderin
 
 		// These are just the reverse of the formulas in GPUStateUtils.
 		float halfActualZRange = gstate_c.vpDepthScale != 0.0f ? vpZScale / gstate_c.vpDepthScale : 0.0f;
+		float inverseDepthScale = gstate_c.vpDepthScale != 0.0f ? 1.0f / gstate_c.vpDepthScale : 0.0f;
 		float minz = -((gstate_c.vpZOffset * halfActualZRange) - vpZCenter) - halfActualZRange;
 		float viewZScale = halfActualZRange;
 		float viewZCenter = minz + halfActualZRange;
@@ -598,7 +599,7 @@ void LinkedShader::UpdateUniforms(const ShaderID &vsid, bool useBufferedRenderin
 			viewZCenter = vpZCenter;
 		}
 
-		float data[4] = { viewZScale, viewZCenter, gstate_c.vpZOffset, 1.0f / gstate_c.vpDepthScale };
+		float data[4] = { viewZScale, viewZCenter, gstate_c.vpZOffset, inverseDepthScale };
 		SetFloatUniform4(render_, &u_depthRange, data);
 	}
 	if (dirty & DIRTY_CULLRANGE) {