From 92f2f330dffb0d0f125d216a0be7bd76879a3de2 Mon Sep 17 00:00:00 2001
From: erwin coumans <erwin.coumans@gmail.com>
Date: Thu, 20 Jun 2013 11:50:19 -0700
Subject: [PATCH] fix OpenCL kernel: NVIDIA crashes in clBuildProgram and AMD
 reports an error 'irreducable flow detected" ?!? ray-convex works on GPU

---
 .../Raycast/kernels/rayCastKernels.cl         | 25 +++++++++++--------
 .../Raycast/kernels/rayCastKernels.h          | 25 +++++++++++--------
 2 files changed, 28 insertions(+), 22 deletions(-)

diff --git a/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl b/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
index 4cf493d1f..8bc21f51b 100644
--- a/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
+++ b/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
@@ -171,11 +171,12 @@ bool rayConvex(float4 rayFromLocal, float4 rayToLocal, int numFaces, int faceOff
 {
 	rayFromLocal.w = 0.f;
 	rayToLocal.w = 0.f;
-
+  bool result = true;
+  
 	float exitFraction = *hitFraction;
 	float enterFraction = -0.1f;
 	float4 curHitNormal = (float4)(0,0,0,0);
-	for (int i=0;i<numFaces;i++)
+	for (int i=0;i<numFaces && result;i++)
 	{
 		b3GpuFace face = faces[faceOffset+i];
 		float fromPlaneDist = dot(rayFromLocal,face.m_plane)+face.m_plane.w;
@@ -203,19 +204,21 @@ bool rayConvex(float4 rayFromLocal, float4 rayToLocal, int numFaces, int faceOff
 				}
 			} else
 			{
-				return false;
+				result = false;
 			}
 		}
 		if (exitFraction <= enterFraction)
-			return false;
+			result = false;
 	}
-
-	if (enterFraction < 0.f)
-		return false;
-
-	*hitFraction = enterFraction;
-	*hitNormal = curHitNormal;
-	return true;
+	
+	result = result && (enterFraction < 0.f);
+	
+	if (result)
+	{	
+		*hitFraction = enterFraction;
+		*hitNormal = curHitNormal;
+	}
+	return result;
 }
 
 
diff --git a/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h b/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
index 6f47345ef..5c2f65c89 100644
--- a/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
+++ b/src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
@@ -173,11 +173,12 @@ static const char* rayCastKernelCL= \
 "{\n"
 "	rayFromLocal.w = 0.f;\n"
 "	rayToLocal.w = 0.f;\n"
-"\n"
+"  bool result = true;\n"
+"  \n"
 "	float exitFraction = *hitFraction;\n"
 "	float enterFraction = -0.1f;\n"
 "	float4 curHitNormal = (float4)(0,0,0,0);\n"
-"	for (int i=0;i<numFaces;i++)\n"
+"	for (int i=0;i<numFaces && result;i++)\n"
 "	{\n"
 "		b3GpuFace face = faces[faceOffset+i];\n"
 "		float fromPlaneDist = dot(rayFromLocal,face.m_plane)+face.m_plane.w;\n"
@@ -205,19 +206,21 @@ static const char* rayCastKernelCL= \
 "				}\n"
 "			} else\n"
 "			{\n"
-"				return false;\n"
+"				result = false;\n"
 "			}\n"
 "		}\n"
 "		if (exitFraction <= enterFraction)\n"
-"			return false;\n"
+"			result = false;\n"
 "	}\n"
-"\n"
-"	if (enterFraction < 0.f)\n"
-"		return false;\n"
-"\n"
-"	*hitFraction = enterFraction;\n"
-"	*hitNormal = curHitNormal;\n"
-"	return true;\n"
+"	\n"
+"	result = result && (enterFraction < 0.f);\n"
+"	\n"
+"	if (result)\n"
+"	{	\n"
+"		*hitFraction = enterFraction;\n"
+"		*hitNormal = curHitNormal;\n"
+"	}\n"
+"	return result;\n"
 "}\n"
 "\n"
 "\n"