reorder files, in preparation for Bullet 3 -> Bullet 2 merge

src/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl

@@ -0,0 +1,344 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+#define WG_SIZE 64
+
+
+
+typedef struct 
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyA;//sign bit set for fixed objects
+	int m_bodyB;
+}Contact4;
+
+typedef struct 
+{
+	int m_n;
+	int m_start;
+	int m_staticIdx;
+	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct 
+{
+	int m_a;
+	int m_b;
+	u32 m_idx;
+}Elem;
+
+#define STACK_SIZE (WG_SIZE*10)
+//#define STACK_SIZE (WG_SIZE)
+#define RING_SIZE 1024
+#define RING_SIZE_MASK (RING_SIZE-1)
+#define CHECK_SIZE (WG_SIZE)
+
+
+#define GET_RING_CAPACITY (RING_SIZE - ldsRingEnd)
+#define RING_END ldsTmp
+
+u32 readBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	return buff[bufIdx] & (1<<bitIdx);
+}
+
+void writeBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+//	buff[bufIdx] |= (1<<bitIdx);
+	atom_or( &buff[bufIdx], (1<<bitIdx) );
+}
+
+u32 tryWrite(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );
+	return ((ans >> bitIdx)&1) == 0;
+}
+
+//	batching on the GPU
+__kernel void CreateBatches( __global const Contact4* gConstraints, __global Contact4* gConstraintsOut,
+		__global const u32* gN, __global const u32* gStart, 
+		int m_staticIdx )
+{
+	__local u32 ldsStackIdx[STACK_SIZE];
+	__local u32 ldsStackEnd;
+	__local Elem ldsRingElem[RING_SIZE];
+	__local u32 ldsRingEnd;
+	__local u32 ldsTmp;
+	__local u32 ldsCheckBuffer[CHECK_SIZE];
+	__local u32 ldsFixedBuffer[CHECK_SIZE];
+	__local u32 ldsGEnd;
+	__local u32 ldsDstEnd;
+
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	
+	const int m_n = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+		
+	if( lIdx == 0 )
+	{
+		ldsRingEnd = 0;
+		ldsGEnd = 0;
+		ldsStackEnd = 0;
+		ldsDstEnd = m_start;
+	}
+	
+//	while(1)
+//was 250
+	for(int ie=0; ie<50; ie++)
+	{
+		ldsFixedBuffer[lIdx] = 0;
+
+		for(int giter=0; giter<4; giter++)
+		{
+			int ringCap = GET_RING_CAPACITY;
+		
+			//	1. fill ring
+			if( ldsGEnd < m_n )
+			{
+				while( ringCap > WG_SIZE )
+				{
+					if( ldsGEnd >= m_n ) break;
+					if( lIdx < ringCap - WG_SIZE )
+					{
+						int srcIdx;
+						AtomInc1( ldsGEnd, srcIdx );
+						if( srcIdx < m_n )
+						{
+							int dstIdx;
+							AtomInc1( ldsRingEnd, dstIdx );
+							
+							int a = gConstraints[m_start+srcIdx].m_bodyA;
+							int b = gConstraints[m_start+srcIdx].m_bodyB;
+							ldsRingElem[dstIdx].m_a = (a>b)? b:a;
+							ldsRingElem[dstIdx].m_b = (a>b)? a:b;
+							ldsRingElem[dstIdx].m_idx = srcIdx;
+						}
+					}
+					ringCap = GET_RING_CAPACITY;
+				}
+			}
+
+			GROUP_LDS_BARRIER;
+	
+			//	2. fill stack
+			__local Elem* dst = ldsRingElem;
+			if( lIdx == 0 ) RING_END = 0;
+
+			int srcIdx=lIdx;
+			int end = ldsRingEnd;
+
+			{
+				for(int ii=0; ii<end; ii+=WG_SIZE, srcIdx+=WG_SIZE)
+				{
+					Elem e;
+					if(srcIdx<end) e = ldsRingElem[srcIdx];
+					bool done = (srcIdx<end)?false:true;
+
+					for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) ldsCheckBuffer[lIdx] = 0;
+					
+					if( !done )
+					{
+						int aUsed = readBuf( ldsFixedBuffer, abs(e.m_a));
+						int bUsed = readBuf( ldsFixedBuffer, abs(e.m_b));
+
+						if( aUsed==0 && bUsed==0 )
+						{
+							int aAvailable;
+							int bAvailable;
+							int ea = abs(e.m_a);
+							int eb = abs(e.m_b);
+
+							aAvailable = tryWrite( ldsCheckBuffer, ea );
+							bAvailable = tryWrite( ldsCheckBuffer, eb );
+
+							bool aStatic = (e.m_a<0) ||(ea==m_staticIdx);
+							bool bStatic = (e.m_b<0) ||(eb==m_staticIdx);
+							
+							aAvailable = aStatic? 1: aAvailable;
+							bAvailable = bStatic? 1: bAvailable;
+
+							bool success = (aAvailable && bAvailable);
+							if(success)
+							{
+							
+								if (!aStatic)
+									writeBuf( ldsFixedBuffer, ea );
+								if (!bStatic)
+									writeBuf( ldsFixedBuffer, eb );
+							}
+							done = success;
+						}
+					}
+
+					//	put it aside
+					if(srcIdx<end)
+					{
+						if( done )
+						{
+							int dstIdx; AtomInc1( ldsStackEnd, dstIdx );
+							if( dstIdx < STACK_SIZE )
+								ldsStackIdx[dstIdx] = e.m_idx;
+							else{
+								done = false;
+								AtomAdd( ldsStackEnd, -1 );
+							}
+						}
+						if( !done )
+						{
+							int dstIdx; AtomInc1( RING_END, dstIdx );
+							dst[dstIdx] = e;
+						}
+					}
+
+					//	if filled, flush
+					if( ldsStackEnd == STACK_SIZE )
+					{
+						for(int i=lIdx; i<STACK_SIZE; i+=WG_SIZE)
+						{
+							int idx = m_start + ldsStackIdx[i];
+							int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+							gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+							gConstraintsOut[ dstIdx ].m_batchIdx = ie;
+						}
+						if( lIdx == 0 ) ldsStackEnd = 0;
+
+						//for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) 
+						ldsFixedBuffer[lIdx] = 0;
+					}
+				}
+			}
+
+			if( lIdx == 0 ) ldsRingEnd = RING_END;
+		}
+
+		GROUP_LDS_BARRIER;
+
+		for(int i=lIdx; i<ldsStackEnd; i+=WG_SIZE)
+		{
+			int idx = m_start + ldsStackIdx[i];
+			int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+			gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+			gConstraintsOut[ dstIdx ].m_batchIdx = ie;
+		}
+
+		//	in case it couldn't consume any pair. Flush them
+		//	todo. Serial batch worth while?
+		if( ldsStackEnd == 0 )
+		{
+			for(int i=lIdx; i<ldsRingEnd; i+=WG_SIZE)
+			{
+				int idx = m_start + ldsRingElem[i].m_idx;
+				int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+				gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+				gConstraintsOut[ dstIdx ].m_batchIdx = 100+i;
+			}
+			GROUP_LDS_BARRIER;
+			if( lIdx == 0 ) ldsRingEnd = 0;
+		}
+
+		if( lIdx == 0 ) ldsStackEnd = 0;
+
+		GROUP_LDS_BARRIER;
+
+		//	termination
+		if( ldsGEnd == m_n && ldsRingEnd == 0 )
+			break;
+	}
+
+
+}
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

src/Bullet3OpenCL/RigidBody/kernels/batchingKernels.h

@@ -0,0 +1,348 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* batchingKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyA;//sign bit set for fixed objects\n"
+"	int m_bodyB;\n"
+"}Contact4;\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_n;\n"
+"	int m_start;\n"
+"	int m_staticIdx;\n"
+"	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_a;\n"
+"	int m_b;\n"
+"	u32 m_idx;\n"
+"}Elem;\n"
+"\n"
+"#define STACK_SIZE (WG_SIZE*10)\n"
+"//#define STACK_SIZE (WG_SIZE)\n"
+"#define RING_SIZE 1024\n"
+"#define RING_SIZE_MASK (RING_SIZE-1)\n"
+"#define CHECK_SIZE (WG_SIZE)\n"
+"\n"
+"\n"
+"#define GET_RING_CAPACITY (RING_SIZE - ldsRingEnd)\n"
+"#define RING_END ldsTmp\n"
+"\n"
+"u32 readBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	return buff[bufIdx] & (1<<bitIdx);\n"
+"}\n"
+"\n"
+"void writeBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"//	buff[bufIdx] |= (1<<bitIdx);\n"
+"	atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"}\n"
+"\n"
+"u32 tryWrite(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"	return ((ans >> bitIdx)&1) == 0;\n"
+"}\n"
+"\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatches( __global const Contact4* gConstraints, __global Contact4* gConstraintsOut,\n"
+"		__global const u32* gN, __global const u32* gStart, \n"
+"		int m_staticIdx )\n"
+"{\n"
+"	__local u32 ldsStackIdx[STACK_SIZE];\n"
+"	__local u32 ldsStackEnd;\n"
+"	__local Elem ldsRingElem[RING_SIZE];\n"
+"	__local u32 ldsRingEnd;\n"
+"	__local u32 ldsTmp;\n"
+"	__local u32 ldsCheckBuffer[CHECK_SIZE];\n"
+"	__local u32 ldsFixedBuffer[CHECK_SIZE];\n"
+"	__local u32 ldsGEnd;\n"
+"	__local u32 ldsDstEnd;\n"
+"\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	\n"
+"	const int m_n = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"		\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsRingEnd = 0;\n"
+"		ldsGEnd = 0;\n"
+"		ldsStackEnd = 0;\n"
+"		ldsDstEnd = m_start;\n"
+"	}\n"
+"	\n"
+"//	while(1)\n"
+"//was 250\n"
+"	for(int ie=0; ie<50; ie++)\n"
+"	{\n"
+"		ldsFixedBuffer[lIdx] = 0;\n"
+"\n"
+"		for(int giter=0; giter<4; giter++)\n"
+"		{\n"
+"			int ringCap = GET_RING_CAPACITY;\n"
+"		\n"
+"			//	1. fill ring\n"
+"			if( ldsGEnd < m_n )\n"
+"			{\n"
+"				while( ringCap > WG_SIZE )\n"
+"				{\n"
+"					if( ldsGEnd >= m_n ) break;\n"
+"					if( lIdx < ringCap - WG_SIZE )\n"
+"					{\n"
+"						int srcIdx;\n"
+"						AtomInc1( ldsGEnd, srcIdx );\n"
+"						if( srcIdx < m_n )\n"
+"						{\n"
+"							int dstIdx;\n"
+"							AtomInc1( ldsRingEnd, dstIdx );\n"
+"							\n"
+"							int a = gConstraints[m_start+srcIdx].m_bodyA;\n"
+"							int b = gConstraints[m_start+srcIdx].m_bodyB;\n"
+"							ldsRingElem[dstIdx].m_a = (a>b)? b:a;\n"
+"							ldsRingElem[dstIdx].m_b = (a>b)? a:b;\n"
+"							ldsRingElem[dstIdx].m_idx = srcIdx;\n"
+"						}\n"
+"					}\n"
+"					ringCap = GET_RING_CAPACITY;\n"
+"				}\n"
+"			}\n"
+"\n"
+"			GROUP_LDS_BARRIER;\n"
+"	\n"
+"			//	2. fill stack\n"
+"			__local Elem* dst = ldsRingElem;\n"
+"			if( lIdx == 0 ) RING_END = 0;\n"
+"\n"
+"			int srcIdx=lIdx;\n"
+"			int end = ldsRingEnd;\n"
+"\n"
+"			{\n"
+"				for(int ii=0; ii<end; ii+=WG_SIZE, srcIdx+=WG_SIZE)\n"
+"				{\n"
+"					Elem e;\n"
+"					if(srcIdx<end) e = ldsRingElem[srcIdx];\n"
+"					bool done = (srcIdx<end)?false:true;\n"
+"\n"
+"					for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) ldsCheckBuffer[lIdx] = 0;\n"
+"					\n"
+"					if( !done )\n"
+"					{\n"
+"						int aUsed = readBuf( ldsFixedBuffer, abs(e.m_a));\n"
+"						int bUsed = readBuf( ldsFixedBuffer, abs(e.m_b));\n"
+"\n"
+"						if( aUsed==0 && bUsed==0 )\n"
+"						{\n"
+"							int aAvailable;\n"
+"							int bAvailable;\n"
+"							int ea = abs(e.m_a);\n"
+"							int eb = abs(e.m_b);\n"
+"\n"
+"							aAvailable = tryWrite( ldsCheckBuffer, ea );\n"
+"							bAvailable = tryWrite( ldsCheckBuffer, eb );\n"
+"\n"
+"							bool aStatic = (e.m_a<0) ||(ea==m_staticIdx);\n"
+"							bool bStatic = (e.m_b<0) ||(eb==m_staticIdx);\n"
+"							\n"
+"							aAvailable = aStatic? 1: aAvailable;\n"
+"							bAvailable = bStatic? 1: bAvailable;\n"
+"\n"
+"							bool success = (aAvailable && bAvailable);\n"
+"							if(success)\n"
+"							{\n"
+"							\n"
+"								if (!aStatic)\n"
+"									writeBuf( ldsFixedBuffer, ea );\n"
+"								if (!bStatic)\n"
+"									writeBuf( ldsFixedBuffer, eb );\n"
+"							}\n"
+"							done = success;\n"
+"						}\n"
+"					}\n"
+"\n"
+"					//	put it aside\n"
+"					if(srcIdx<end)\n"
+"					{\n"
+"						if( done )\n"
+"						{\n"
+"							int dstIdx; AtomInc1( ldsStackEnd, dstIdx );\n"
+"							if( dstIdx < STACK_SIZE )\n"
+"								ldsStackIdx[dstIdx] = e.m_idx;\n"
+"							else{\n"
+"								done = false;\n"
+"								AtomAdd( ldsStackEnd, -1 );\n"
+"							}\n"
+"						}\n"
+"						if( !done )\n"
+"						{\n"
+"							int dstIdx; AtomInc1( RING_END, dstIdx );\n"
+"							dst[dstIdx] = e;\n"
+"						}\n"
+"					}\n"
+"\n"
+"					//	if filled, flush\n"
+"					if( ldsStackEnd == STACK_SIZE )\n"
+"					{\n"
+"						for(int i=lIdx; i<STACK_SIZE; i+=WG_SIZE)\n"
+"						{\n"
+"							int idx = m_start + ldsStackIdx[i];\n"
+"							int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"							gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"							gConstraintsOut[ dstIdx ].m_batchIdx = ie;\n"
+"						}\n"
+"						if( lIdx == 0 ) ldsStackEnd = 0;\n"
+"\n"
+"						//for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) \n"
+"						ldsFixedBuffer[lIdx] = 0;\n"
+"					}\n"
+"				}\n"
+"			}\n"
+"\n"
+"			if( lIdx == 0 ) ldsRingEnd = RING_END;\n"
+"		}\n"
+"\n"
+"		GROUP_LDS_BARRIER;\n"
+"\n"
+"		for(int i=lIdx; i<ldsStackEnd; i+=WG_SIZE)\n"
+"		{\n"
+"			int idx = m_start + ldsStackIdx[i];\n"
+"			int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"			gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"			gConstraintsOut[ dstIdx ].m_batchIdx = ie;\n"
+"		}\n"
+"\n"
+"		//	in case it couldn't consume any pair. Flush them\n"
+"		//	todo. Serial batch worth while?\n"
+"		if( ldsStackEnd == 0 )\n"
+"		{\n"
+"			for(int i=lIdx; i<ldsRingEnd; i+=WG_SIZE)\n"
+"			{\n"
+"				int idx = m_start + ldsRingElem[i].m_idx;\n"
+"				int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"				gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"				gConstraintsOut[ dstIdx ].m_batchIdx = 100+i;\n"
+"			}\n"
+"			GROUP_LDS_BARRIER;\n"
+"			if( lIdx == 0 ) ldsRingEnd = 0;\n"
+"		}\n"
+"\n"
+"		if( lIdx == 0 ) ldsStackEnd = 0;\n"
+"\n"
+"		GROUP_LDS_BARRIER;\n"
+"\n"
+"		//	termination\n"
+"		if( ldsGEnd == m_n && ldsRingEnd == 0 )\n"
+"			break;\n"
+"	}\n"
+"\n"
+"\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl

@@ -0,0 +1,236 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+#define SIMD_WIDTH 64
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+#define WG_SIZE 64
+
+
+
+typedef struct 
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;//sign bit set for fixed objects
+	int m_bodyBPtrAndSignBit;
+}Contact4;
+
+typedef struct 
+{
+	int m_n;
+	int m_start;
+	int m_staticIdx;
+	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct 
+{
+	int m_a;
+	int m_b;
+	u32 m_idx;
+}Elem;
+
+
+
+
+
+//	batching on the GPU
+__kernel void CreateBatchesBruteForce( __global Contact4* gConstraints, 	__global const u32* gN, __global const u32* gStart, int m_staticIdx )
+{
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	
+	const int m_n = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+		
+	if( lIdx == 0 )
+	{
+		for (int i=0;i<m_n;i++)
+		{
+			int srcIdx = i+m_start;
+			int batchIndex = i;
+			gConstraints[ srcIdx ].m_batchIdx = batchIndex;	
+		}
+	}
+}
+
+
+#define CHECK_SIZE (WG_SIZE)
+
+
+
+
+u32 readBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	return buff[bufIdx] & (1<<bitIdx);
+}
+
+void writeBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	buff[bufIdx] |= (1<<bitIdx);
+	//atom_or( &buff[bufIdx], (1<<bitIdx) );
+}
+
+u32 tryWrite(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );
+	return ((ans >> bitIdx)&1) == 0;
+}
+
+
+//	batching on the GPU
+__kernel void CreateBatchesNew( __global Contact4* gConstraints, __global const u32* gN, __global const u32* gStart, int staticIdx )
+{
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	const int numConstraints = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+		
+	
+	__local u32 ldsFixedBuffer[CHECK_SIZE];
+		
+	
+	
+	
+	
+	if( lIdx == 0 )
+	{
+	
+		
+		__global Contact4* cs = &gConstraints[m_start];	
+	
+		
+		int numValidConstraints = 0;
+		int batchIdx = 0;
+
+		while( numValidConstraints < numConstraints)
+		{
+			int nCurrentBatch = 0;
+			//	clear flag
+	
+			for(int i=0; i<CHECK_SIZE; i++) 
+				ldsFixedBuffer[i] = 0;		
+
+			for(int i=numValidConstraints; i<numConstraints; i++)
+			{
+
+				int bodyAS = cs[i].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[i].m_bodyBPtrAndSignBit;
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+				int aUnavailable = aIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyA);
+				int bUnavailable = bIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyB);
+				
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+					{
+						writeBuf( ldsFixedBuffer, bodyA );
+					}
+					if (!bIsStatic)
+					{
+						writeBuf( ldsFixedBuffer, bodyB );
+					}
+
+					cs[i].m_batchIdx = batchIdx;
+
+					if (i!=numValidConstraints)
+					{
+						//btSwap(cs[i],cs[numValidConstraints]);
+						
+						Contact4 tmp = cs[i];
+						cs[i] = cs[numValidConstraints];
+						cs[numValidConstraints] = tmp;
+						
+					}
+
+					numValidConstraints++;
+					
+					nCurrentBatch++;
+					if( nCurrentBatch == SIMD_WIDTH)
+					{
+						nCurrentBatch = 0;
+						for(int i=0; i<CHECK_SIZE; i++) 
+							ldsFixedBuffer[i] = 0;
+						
+					}
+				}
+			}//for
+			batchIdx ++;
+		}//while
+	}//if( lIdx == 0 )
+	
+	//return batchIdx;
+}

src/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.h

@@ -0,0 +1,240 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* batchingKernelsNewCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"\n"
+"#define SIMD_WIDTH 64\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;//sign bit set for fixed objects\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"}Contact4;\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_n;\n"
+"	int m_start;\n"
+"	int m_staticIdx;\n"
+"	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_a;\n"
+"	int m_b;\n"
+"	u32 m_idx;\n"
+"}Elem;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatchesBruteForce( __global Contact4* gConstraints, 	__global const u32* gN, __global const u32* gStart, int m_staticIdx )\n"
+"{\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	\n"
+"	const int m_n = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"		\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		for (int i=0;i<m_n;i++)\n"
+"		{\n"
+"			int srcIdx = i+m_start;\n"
+"			int batchIndex = i;\n"
+"			gConstraints[ srcIdx ].m_batchIdx = batchIndex;	\n"
+"		}\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"#define CHECK_SIZE (WG_SIZE)\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"u32 readBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	return buff[bufIdx] & (1<<bitIdx);\n"
+"}\n"
+"\n"
+"void writeBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	buff[bufIdx] |= (1<<bitIdx);\n"
+"	//atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"}\n"
+"\n"
+"u32 tryWrite(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"	return ((ans >> bitIdx)&1) == 0;\n"
+"}\n"
+"\n"
+"\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatchesNew( __global Contact4* gConstraints, __global const u32* gN, __global const u32* gStart, int staticIdx )\n"
+"{\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	const int numConstraints = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"		\n"
+"	\n"
+"	__local u32 ldsFixedBuffer[CHECK_SIZE];\n"
+"		\n"
+"	\n"
+"	\n"
+"	\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"	\n"
+"		\n"
+"		__global Contact4* cs = &gConstraints[m_start];	\n"
+"	\n"
+"		\n"
+"		int numValidConstraints = 0;\n"
+"		int batchIdx = 0;\n"
+"\n"
+"		while( numValidConstraints < numConstraints)\n"
+"		{\n"
+"			int nCurrentBatch = 0;\n"
+"			//	clear flag\n"
+"	\n"
+"			for(int i=0; i<CHECK_SIZE; i++) \n"
+"				ldsFixedBuffer[i] = 0;		\n"
+"\n"
+"			for(int i=numValidConstraints; i<numConstraints; i++)\n"
+"			{\n"
+"\n"
+"				int bodyAS = cs[i].m_bodyAPtrAndSignBit;\n"
+"				int bodyBS = cs[i].m_bodyBPtrAndSignBit;\n"
+"				int bodyA = abs(bodyAS);\n"
+"				int bodyB = abs(bodyBS);\n"
+"				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;\n"
+"				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;\n"
+"				int aUnavailable = aIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyA);\n"
+"				int bUnavailable = bIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyB);\n"
+"				\n"
+"				if( aUnavailable==0 && bUnavailable==0 ) // ok\n"
+"				{\n"
+"					if (!aIsStatic)\n"
+"					{\n"
+"						writeBuf( ldsFixedBuffer, bodyA );\n"
+"					}\n"
+"					if (!bIsStatic)\n"
+"					{\n"
+"						writeBuf( ldsFixedBuffer, bodyB );\n"
+"					}\n"
+"\n"
+"					cs[i].m_batchIdx = batchIdx;\n"
+"\n"
+"					if (i!=numValidConstraints)\n"
+"					{\n"
+"						//b3Swap(cs[i],cs[numValidConstraints]);\n"
+"						\n"
+"						Contact4 tmp = cs[i];\n"
+"						cs[i] = cs[numValidConstraints];\n"
+"						cs[numValidConstraints] = tmp;\n"
+"						\n"
+"					}\n"
+"\n"
+"					numValidConstraints++;\n"
+"					\n"
+"					nCurrentBatch++;\n"
+"					if( nCurrentBatch == SIMD_WIDTH)\n"
+"					{\n"
+"						nCurrentBatch = 0;\n"
+"						for(int i=0; i<CHECK_SIZE; i++) \n"
+"							ldsFixedBuffer[i] = 0;\n"
+"						\n"
+"					}\n"
+"				}\n"
+"			}//for\n"
+"			batchIdx ++;\n"
+"		}//while\n"
+"	}//if( lIdx == 0 )\n"
+"	\n"
+"	//return batchIdx;\n"
+"}\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/integrateKernel.cl

@@ -0,0 +1,92 @@
+
+float4 quatMult(float4 q1, float4 q2)
+{
+	float4 q;
+	q.x = q1.w * q2.x + q1.x * q2.w + q1.y * q2.z - q1.z * q2.y;
+	q.y = q1.w * q2.y + q1.y * q2.w + q1.z * q2.x - q1.x * q2.z;
+	q.z = q1.w * q2.z + q1.z * q2.w + q1.x * q2.y - q1.y * q2.x;
+	q.w = q1.w * q2.w - q1.x * q2.x - q1.y * q2.y - q1.z * q2.z; 
+	return q;
+}
+
+float4 quatNorm(float4 q)
+{
+	float len = native_sqrt(dot(q, q));
+	if(len > 0.f)
+	{
+		q *= 1.f / len;
+	}
+	else
+	{
+		q.x = q.y = q.z = 0.f;
+		q.w = 1.f;
+	}
+	return q;
+}
+
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	unsigned int m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+
+
+
+__kernel void 
+  integrateTransformsKernel( __global Body* bodies,const int numNodes, float timeStep, float angularDamping, float4 gravityAcceleration)
+{
+	int nodeID = get_global_id(0);
+	float BT_GPU_ANGULAR_MOTION_THRESHOLD = (0.25f * 3.14159254f);
+	if( nodeID < numNodes && (bodies[nodeID].m_invMass != 0.f))
+	{
+		//angular velocity
+		{
+			float4 axis;
+			//add some hardcoded angular damping
+			bodies[nodeID].m_angVel.x *= angularDamping;
+			bodies[nodeID].m_angVel.y *= angularDamping;
+			bodies[nodeID].m_angVel.z *= angularDamping;
+			
+			float4 angvel = bodies[nodeID].m_angVel;
+			float fAngle = native_sqrt(dot(angvel, angvel));
+			//limit the angular motion
+			if(fAngle*timeStep > BT_GPU_ANGULAR_MOTION_THRESHOLD)
+			{
+				fAngle = BT_GPU_ANGULAR_MOTION_THRESHOLD / timeStep;
+			}
+			if(fAngle < 0.001f)
+			{
+				// use Taylor's expansions of sync function
+				axis = angvel * (0.5f*timeStep-(timeStep*timeStep*timeStep)*0.020833333333f * fAngle * fAngle);
+			}
+			else
+			{
+				// sync(fAngle) = sin(c*fAngle)/t
+				axis = angvel * ( native_sin(0.5f * fAngle * timeStep) / fAngle);
+			}
+			float4 dorn = axis;
+			dorn.w = native_cos(fAngle * timeStep * 0.5f);
+			float4 orn0 = bodies[nodeID].m_quat;
+
+			float4 predictedOrn = quatMult(dorn, orn0);
+			predictedOrn = quatNorm(predictedOrn);
+			bodies[nodeID].m_quat=predictedOrn;
+		}
+
+		//linear velocity		
+		bodies[nodeID].m_pos +=  bodies[nodeID].m_linVel * timeStep;
+		
+		//apply gravity
+		bodies[nodeID].m_linVel += gravityAcceleration * timeStep;
+		
+	}
+}

src/Bullet3OpenCL/RigidBody/kernels/integrateKernel.h

@@ -0,0 +1,96 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* integrateKernelCL= \
+"\n"
+"float4 quatMult(float4 q1, float4 q2)\n"
+"{\n"
+"	float4 q;\n"
+"	q.x = q1.w * q2.x + q1.x * q2.w + q1.y * q2.z - q1.z * q2.y;\n"
+"	q.y = q1.w * q2.y + q1.y * q2.w + q1.z * q2.x - q1.x * q2.z;\n"
+"	q.z = q1.w * q2.z + q1.z * q2.w + q1.x * q2.y - q1.y * q2.x;\n"
+"	q.w = q1.w * q2.w - q1.x * q2.x - q1.y * q2.y - q1.z * q2.z; \n"
+"	return q;\n"
+"}\n"
+"\n"
+"float4 quatNorm(float4 q)\n"
+"{\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	unsigned int m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__kernel void \n"
+"  integrateTransformsKernel( __global Body* bodies,const int numNodes, float timeStep, float angularDamping, float4 gravityAcceleration)\n"
+"{\n"
+"	int nodeID = get_global_id(0);\n"
+"	float B3_GPU_ANGULAR_MOTION_THRESHOLD = (0.25f * 3.14159254f);\n"
+"	if( nodeID < numNodes && (bodies[nodeID].m_invMass != 0.f))\n"
+"	{\n"
+"		//angular velocity\n"
+"		{\n"
+"			float4 axis;\n"
+"			//add some hardcoded angular damping\n"
+"			bodies[nodeID].m_angVel.x *= angularDamping;\n"
+"			bodies[nodeID].m_angVel.y *= angularDamping;\n"
+"			bodies[nodeID].m_angVel.z *= angularDamping;\n"
+"			\n"
+"			float4 angvel = bodies[nodeID].m_angVel;\n"
+"			float fAngle = native_sqrt(dot(angvel, angvel));\n"
+"			//limit the angular motion\n"
+"			if(fAngle*timeStep > B3_GPU_ANGULAR_MOTION_THRESHOLD)\n"
+"			{\n"
+"				fAngle = B3_GPU_ANGULAR_MOTION_THRESHOLD / timeStep;\n"
+"			}\n"
+"			if(fAngle < 0.001f)\n"
+"			{\n"
+"				// use Taylor's expansions of sync function\n"
+"				axis = angvel * (0.5f*timeStep-(timeStep*timeStep*timeStep)*0.020833333333f * fAngle * fAngle);\n"
+"			}\n"
+"			else\n"
+"			{\n"
+"				// sync(fAngle) = sin(c*fAngle)/t\n"
+"				axis = angvel * ( native_sin(0.5f * fAngle * timeStep) / fAngle);\n"
+"			}\n"
+"			float4 dorn = axis;\n"
+"			dorn.w = native_cos(fAngle * timeStep * 0.5f);\n"
+"			float4 orn0 = bodies[nodeID].m_quat;\n"
+"\n"
+"			float4 predictedOrn = quatMult(dorn, orn0);\n"
+"			predictedOrn = quatNorm(predictedOrn);\n"
+"			bodies[nodeID].m_quat=predictedOrn;\n"
+"		}\n"
+"\n"
+"		//linear velocity		\n"
+"		bodies[nodeID].m_pos +=  bodies[nodeID].m_linVel * timeStep;\n"
+"		\n"
+"		//apply gravity\n"
+"		bodies[nodeID].m_linVel += gravityAcceleration * timeStep;\n"
+"		\n"
+"	}\n"
+"}\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/solveContact.cl

@@ -0,0 +1,476 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+//#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define mymake_float4 (float4)
+//#define make_float2 (float2)
+//#define make_uint4 (uint4)
+//#define make_int4 (int4)
+//#define make_uint2 (uint2)
+//#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+
+
+
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = mymake_float4(a.xyz,0.f);
+	float4 b1 = mymake_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+
+
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+
+
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+typedef struct
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+} Contact4;
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = mymake_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/(jmj0+jmj1+jmj2+jmj3);
+}
+
+
+void solveContact(__global Constraint4* cs,
+				  float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+				  float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB);
+
+void solveContact(__global Constraint4* cs,
+			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB)
+{
+	float minRambdaDt = 0;
+	float maxRambdaDt = FLT_MAX;
+
+	for(int ic=0; ic<4; ic++)
+	{
+		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;
+
+		float4 angular0, angular1, linear;
+		float4 r0 = cs->m_worldPos[ic] - posA;
+		float4 r1 = cs->m_worldPos[ic] - posB;
+		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );
+
+		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, 
+			*linVelA, *angVelA, *linVelB, *angVelB ) + cs->m_b[ic];
+		rambdaDt *= cs->m_jacCoeffInv[ic];
+
+		{
+			float prevSum = cs->m_appliedRambdaDt[ic];
+			float updated = prevSum;
+			updated += rambdaDt;
+			updated = max2( updated, minRambdaDt );
+			updated = min2( updated, maxRambdaDt );
+			rambdaDt = updated - prevSum;
+			cs->m_appliedRambdaDt[ic] = updated;
+		}
+
+		float4 linImp0 = invMassA*linear*rambdaDt;
+		float4 linImp1 = invMassB*(-linear)*rambdaDt;
+		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+
+		*linVelA += linImp0;
+		*angVelA += angImp0;
+		*linVelB += linImp1;
+		*angVelB += angImp1;
+	}
+}
+
+void btPlaneSpace1 (const float4* n, float4* p, float4* q);
+ void btPlaneSpace1 (const float4* n, float4* p, float4* q)
+{
+  if (fabs(n[0].z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n[0].y*n[0].y + n[0].z*n[0].z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n[0].z*k;
+	p[0].z = n[0].y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n[0].x*p[0].z;
+	q[0].z = n[0].x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n[0].x*n[0].x + n[0].y*n[0].y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n[0].y*k;
+	p[0].y = n[0].x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n[0].z*p[0].y;
+	q[0].y = n[0].z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)
+{
+	//float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+			posB, &linVelB, &angVelB, invMassB, invInertiaB );
+
+  if (gBodies[aIdx].m_invMass)
+  {
+		gBodies[aIdx].m_linVel = linVelA;
+		gBodies[aIdx].m_angVel = angVelA;
+	} else
+	{
+		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);
+	
+	}
+	if (gBodies[bIdx].m_invMass)
+  {
+		gBodies[bIdx].m_linVel = linVelB;
+		gBodies[bIdx].m_angVel = angVelB;
+	} else
+	{
+		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);
+	
+	}
+
+}
+
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void BatchSolveKernelContact(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                      __global int* gN,
+                      __global int* gOffsets,
+                       int maxBatch,
+                       int bIdx,
+                       int nSplit
+                      )
+{
+	//__local int ldsBatchIdx[WG_SIZE+1];
+	__local int ldsCurBatch;
+	__local int ldsNextBatch;
+	__local int ldsStart;
+
+	int lIdx = GET_LOCAL_IDX;
+	int wgIdx = GET_GROUP_IDX;
+
+//	int gIdx = GET_GLOBAL_IDX;
+//	debugInfo[gIdx].m_valInt0 = gIdx;
+	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
+
+
+	int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
+	int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
+	int cellIdx = xIdx+yIdx*nSplit;
+	
+	if( gN[cellIdx] == 0 ) 
+		return;
+
+	const int start = gOffsets[cellIdx];
+	const int end = start + gN[cellIdx];
+
+	
+	if( lIdx == 0 )
+	{
+		ldsCurBatch = 0;
+		ldsNextBatch = 0;
+		ldsStart = start;
+	}
+
+
+	GROUP_LDS_BARRIER;
+
+	int idx=ldsStart+lIdx;
+	while (ldsCurBatch < maxBatch)
+	{
+		for(; idx<end; )
+		{
+			if (gConstraints[idx].m_batchIdx == ldsCurBatch)
+			{
+					solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );
+
+				 idx+=64;
+			} else
+			{
+				break;
+			}
+		}
+		GROUP_LDS_BARRIER;
+		if( lIdx == 0 )
+		{
+			ldsCurBatch++;
+		}
+		GROUP_LDS_BARRIER;
+	}
+	
+    
+}

src/Bullet3OpenCL/RigidBody/kernels/solveContact.h

@@ -0,0 +1,480 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solveContactCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"\n"
+"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define mymake_float4 (float4)\n"
+"//#define make_float2 (float2)\n"
+"//#define make_uint4 (uint4)\n"
+"//#define make_int4 (int4)\n"
+"//#define make_uint2 (uint2)\n"
+"//#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = mymake_float4(a.xyz,0.f);\n"
+"	float4 b1 = mymake_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"} Contact4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = mymake_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"\n"
+"\n"
+"void solveContact(__global Constraint4* cs,\n"
+"				  float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"				  float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB);\n"
+"\n"
+"void solveContact(__global Constraint4* cs,\n"
+"			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB)\n"
+"{\n"
+"	float minRambdaDt = 0;\n"
+"	float maxRambdaDt = FLT_MAX;\n"
+"\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;\n"
+"\n"
+"		float4 angular0, angular1, linear;\n"
+"		float4 r0 = cs->m_worldPos[ic] - posA;\n"
+"		float4 r1 = cs->m_worldPos[ic] - posB;\n"
+"		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );\n"
+"\n"
+"		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, \n"
+"			*linVelA, *angVelA, *linVelB, *angVelB ) + cs->m_b[ic];\n"
+"		rambdaDt *= cs->m_jacCoeffInv[ic];\n"
+"\n"
+"		{\n"
+"			float prevSum = cs->m_appliedRambdaDt[ic];\n"
+"			float updated = prevSum;\n"
+"			updated += rambdaDt;\n"
+"			updated = max2( updated, minRambdaDt );\n"
+"			updated = min2( updated, maxRambdaDt );\n"
+"			rambdaDt = updated - prevSum;\n"
+"			cs->m_appliedRambdaDt[ic] = updated;\n"
+"		}\n"
+"\n"
+"		float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"		float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"\n"
+"		*linVelA += linImp0;\n"
+"		*angVelA += angImp0;\n"
+"		*linVelB += linImp1;\n"
+"		*angVelB += angImp1;\n"
+"	}\n"
+"}\n"
+"\n"
+"void b3PlaneSpace1 (const float4* n, float4* p, float4* q);\n"
+" void b3PlaneSpace1 (const float4* n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n[0].z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n[0].y*n[0].y + n[0].z*n[0].z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n[0].z*k;\n"
+"	p[0].z = n[0].y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n[0].x*p[0].z;\n"
+"	q[0].z = n[0].x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n[0].x*n[0].x + n[0].y*n[0].y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n[0].y*k;\n"
+"	p[0].y = n[0].x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n[0].z*p[0].y;\n"
+"	q[0].y = n[0].z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)\n"
+"{\n"
+"	//float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"\n"
+"	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"			posB, &linVelB, &angVelB, invMassB, invInertiaB );\n"
+"\n"
+"  if (gBodies[aIdx].m_invMass)\n"
+"  {\n"
+"		gBodies[aIdx].m_linVel = linVelA;\n"
+"		gBodies[aIdx].m_angVel = angVelA;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	\n"
+"	}\n"
+"	if (gBodies[bIdx].m_invMass)\n"
+"  {\n"
+"		gBodies[bIdx].m_linVel = linVelB;\n"
+"		gBodies[bIdx].m_angVel = angVelB;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	\n"
+"	}\n"
+"\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void BatchSolveKernelContact(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                      __global int* gN,\n"
+"                      __global int* gOffsets,\n"
+"                       int maxBatch,\n"
+"                       int bIdx,\n"
+"                       int nSplit\n"
+"                      )\n"
+"{\n"
+"	//__local int ldsBatchIdx[WG_SIZE+1];\n"
+"	__local int ldsCurBatch;\n"
+"	__local int ldsNextBatch;\n"
+"	__local int ldsStart;\n"
+"\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"\n"
+"//	int gIdx = GET_GLOBAL_IDX;\n"
+"//	debugInfo[gIdx].m_valInt0 = gIdx;\n"
+"	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
+"\n"
+"\n"
+"	int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
+"	int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
+"	int cellIdx = xIdx+yIdx*nSplit;\n"
+"	\n"
+"	if( gN[cellIdx] == 0 ) \n"
+"		return;\n"
+"\n"
+"	const int start = gOffsets[cellIdx];\n"
+"	const int end = start + gN[cellIdx];\n"
+"\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsCurBatch = 0;\n"
+"		ldsNextBatch = 0;\n"
+"		ldsStart = start;\n"
+"	}\n"
+"\n"
+"\n"
+"	GROUP_LDS_BARRIER;\n"
+"\n"
+"	int idx=ldsStart+lIdx;\n"
+"	while (ldsCurBatch < maxBatch)\n"
+"	{\n"
+"		for(; idx<end; )\n"
+"		{\n"
+"			if (gConstraints[idx].m_batchIdx == ldsCurBatch)\n"
+"			{\n"
+"					solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"\n"
+"				 idx+=64;\n"
+"			} else\n"
+"			{\n"
+"				break;\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		if( lIdx == 0 )\n"
+"		{\n"
+"			ldsCurBatch++;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"	\n"
+"    \n"
+"}\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl

@@ -0,0 +1,506 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+//#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define mymake_float4 (float4)
+//#define make_float2 (float2)
+//#define make_uint4 (uint4)
+//#define make_int4 (int4)
+//#define make_uint2 (uint2)
+//#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+
+
+
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = mymake_float4(a.xyz,0.f);
+	float4 b1 = mymake_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+
+
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+
+
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+typedef struct
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+} Contact4;
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = mymake_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/(jmj0+jmj1+jmj2+jmj3);
+}
+void btPlaneSpace1 (const float4* n, float4* p, float4* q);
+ void btPlaneSpace1 (const float4* n, float4* p, float4* q)
+{
+  if (fabs(n[0].z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n[0].y*n[0].y + n[0].z*n[0].z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n[0].z*k;
+	p[0].z = n[0].y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n[0].x*p[0].z;
+	q[0].z = n[0].x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n[0].x*n[0].x + n[0].y*n[0].y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n[0].y*k;
+	p[0].y = n[0].x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n[0].z*p[0].y;
+	q[0].y = n[0].z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)
+{
+	float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+	
+
+	{
+		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+		float sum = 0;
+		for(int j=0; j<4; j++)
+		{
+			sum +=ldsCs[0].m_appliedRambdaDt[j];
+		}
+		frictionCoeff = 0.7f;
+		for(int j=0; j<4; j++)
+		{
+			maxRambdaDt[j] = frictionCoeff*sum;
+			minRambdaDt[j] = -maxRambdaDt[j];
+		}
+
+		
+//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );
+		
+		
+		{
+			
+			__global Constraint4* cs = ldsCs;
+			
+			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;
+			const float4 center = cs->m_center;
+			
+			float4 n = -cs->m_linear;
+			
+			float4 tangent[2];
+			btPlaneSpace1(&n,&tangent[0],&tangent[1]);
+			float4 angular0, angular1, linear;
+			float4 r0 = center - posA;
+			float4 r1 = center - posB;
+			for(int i=0; i<2; i++)
+			{
+				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );
+				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+											linVelA, angVelA, linVelB, angVelB );
+				rambdaDt *= cs->m_fJacCoeffInv[i];
+				
+				{
+					float prevSum = cs->m_fAppliedRambdaDt[i];
+					float updated = prevSum;
+					updated += rambdaDt;
+					updated = max2( updated, minRambdaDt[i] );
+					updated = min2( updated, maxRambdaDt[i] );
+					rambdaDt = updated - prevSum;
+					cs->m_fAppliedRambdaDt[i] = updated;
+				}
+				
+				float4 linImp0 = invMassA*linear*rambdaDt;
+				float4 linImp1 = invMassB*(-linear)*rambdaDt;
+				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+				
+				linVelA += linImp0;
+				angVelA += angImp0;
+				linVelB += linImp1;
+				angVelB += angImp1;
+			}
+			{	//	angular damping for point constraint
+				float4 ab = normalize3( posB - posA );
+				float4 ac = normalize3( center - posA );
+				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))
+				{
+					float angNA = dot3F4( n, angVelA );
+					float angNB = dot3F4( n, angVelB );
+					
+					angVelA -= (angNA*0.1f)*n;
+					angVelB -= (angNB*0.1f)*n;
+				}
+			}
+		}
+
+		
+		
+	}
+
+	if (gBodies[aIdx].m_invMass)
+	{
+		gBodies[aIdx].m_linVel = linVelA;
+		gBodies[aIdx].m_angVel = angVelA;
+	} else
+	{
+		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);
+	}
+	if (gBodies[bIdx].m_invMass)
+	{
+		gBodies[bIdx].m_linVel = linVelB;
+		gBodies[bIdx].m_angVel = angVelB;
+	} else
+	{
+		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);
+	}
+ 
+
+}
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void BatchSolveKernelFriction(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                      __global int* gN,
+                      __global int* gOffsets,
+                       int maxBatch,
+                       int bIdx,
+                       int nSplit
+                      )
+{
+	//__local int ldsBatchIdx[WG_SIZE+1];
+	__local int ldsCurBatch;
+	__local int ldsNextBatch;
+	__local int ldsStart;
+
+	int lIdx = GET_LOCAL_IDX;
+	int wgIdx = GET_GROUP_IDX;
+
+//	int gIdx = GET_GLOBAL_IDX;
+//	debugInfo[gIdx].m_valInt0 = gIdx;
+	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
+
+
+	int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
+	int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
+	int cellIdx = xIdx+yIdx*nSplit;
+	
+	if( gN[cellIdx] == 0 ) 
+		return;
+
+	const int start = gOffsets[cellIdx];
+	const int end = start + gN[cellIdx];
+
+	
+	if( lIdx == 0 )
+	{
+		ldsCurBatch = 0;
+		ldsNextBatch = 0;
+		ldsStart = start;
+	}
+
+
+	GROUP_LDS_BARRIER;
+
+	int idx=ldsStart+lIdx;
+	while (ldsCurBatch < maxBatch)
+	{
+		for(; idx<end; )
+		{
+			if (gConstraints[idx].m_batchIdx == ldsCurBatch)
+			{
+
+					solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );
+
+				 idx+=64;
+			} else
+			{
+				break;
+			}
+		}
+		GROUP_LDS_BARRIER;
+		if( lIdx == 0 )
+		{
+			ldsCurBatch++;
+		}
+		GROUP_LDS_BARRIER;
+	}
+	
+    
+}

src/Bullet3OpenCL/RigidBody/kernels/solveFriction.h

@@ -0,0 +1,510 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solveFrictionCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"\n"
+"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define mymake_float4 (float4)\n"
+"//#define make_float2 (float2)\n"
+"//#define make_uint4 (uint4)\n"
+"//#define make_int4 (int4)\n"
+"//#define make_uint2 (uint2)\n"
+"//#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = mymake_float4(a.xyz,0.f);\n"
+"	float4 b1 = mymake_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"} Contact4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = mymake_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"void b3PlaneSpace1 (const float4* n, float4* p, float4* q);\n"
+" void b3PlaneSpace1 (const float4* n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n[0].z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n[0].y*n[0].y + n[0].z*n[0].z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n[0].z*k;\n"
+"	p[0].z = n[0].y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n[0].x*p[0].z;\n"
+"	q[0].z = n[0].x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n[0].x*n[0].x + n[0].y*n[0].y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n[0].y*k;\n"
+"	p[0].y = n[0].x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n[0].z*p[0].y;\n"
+"	q[0].y = n[0].z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"\n"
+"\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)\n"
+"{\n"
+"	float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"\n"
+"\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"	\n"
+"\n"
+"	{\n"
+"		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};\n"
+"		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};\n"
+"\n"
+"		float sum = 0;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			sum +=ldsCs[0].m_appliedRambdaDt[j];\n"
+"		}\n"
+"		frictionCoeff = 0.7f;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			maxRambdaDt[j] = frictionCoeff*sum;\n"
+"			minRambdaDt[j] = -maxRambdaDt[j];\n"
+"		}\n"
+"\n"
+"		\n"
+"//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );\n"
+"		\n"
+"		\n"
+"		{\n"
+"			\n"
+"			__global Constraint4* cs = ldsCs;\n"
+"			\n"
+"			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;\n"
+"			const float4 center = cs->m_center;\n"
+"			\n"
+"			float4 n = -cs->m_linear;\n"
+"			\n"
+"			float4 tangent[2];\n"
+"			b3PlaneSpace1(&n,&tangent[0],&tangent[1]);\n"
+"			float4 angular0, angular1, linear;\n"
+"			float4 r0 = center - posA;\n"
+"			float4 r1 = center - posB;\n"
+"			for(int i=0; i<2; i++)\n"
+"			{\n"
+"				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );\n"
+"				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,\n"
+"											linVelA, angVelA, linVelB, angVelB );\n"
+"				rambdaDt *= cs->m_fJacCoeffInv[i];\n"
+"				\n"
+"				{\n"
+"					float prevSum = cs->m_fAppliedRambdaDt[i];\n"
+"					float updated = prevSum;\n"
+"					updated += rambdaDt;\n"
+"					updated = max2( updated, minRambdaDt[i] );\n"
+"					updated = min2( updated, maxRambdaDt[i] );\n"
+"					rambdaDt = updated - prevSum;\n"
+"					cs->m_fAppliedRambdaDt[i] = updated;\n"
+"				}\n"
+"				\n"
+"				float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"				float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"				\n"
+"				linVelA += linImp0;\n"
+"				angVelA += angImp0;\n"
+"				linVelB += linImp1;\n"
+"				angVelB += angImp1;\n"
+"			}\n"
+"			{	//	angular damping for point constraint\n"
+"				float4 ab = normalize3( posB - posA );\n"
+"				float4 ac = normalize3( center - posA );\n"
+"				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))\n"
+"				{\n"
+"					float angNA = dot3F4( n, angVelA );\n"
+"					float angNB = dot3F4( n, angVelB );\n"
+"					\n"
+"					angVelA -= (angNA*0.1f)*n;\n"
+"					angVelB -= (angNB*0.1f)*n;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"\n"
+"		\n"
+"		\n"
+"	}\n"
+"\n"
+"	if (gBodies[aIdx].m_invMass)\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = linVelA;\n"
+"		gBodies[aIdx].m_angVel = angVelA;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	}\n"
+"	if (gBodies[bIdx].m_invMass)\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = linVelB;\n"
+"		gBodies[bIdx].m_angVel = angVelB;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	}\n"
+" \n"
+"\n"
+"}\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void BatchSolveKernelFriction(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                      __global int* gN,\n"
+"                      __global int* gOffsets,\n"
+"                       int maxBatch,\n"
+"                       int bIdx,\n"
+"                       int nSplit\n"
+"                      )\n"
+"{\n"
+"	//__local int ldsBatchIdx[WG_SIZE+1];\n"
+"	__local int ldsCurBatch;\n"
+"	__local int ldsNextBatch;\n"
+"	__local int ldsStart;\n"
+"\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"\n"
+"//	int gIdx = GET_GLOBAL_IDX;\n"
+"//	debugInfo[gIdx].m_valInt0 = gIdx;\n"
+"	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
+"\n"
+"\n"
+"	int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
+"	int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
+"	int cellIdx = xIdx+yIdx*nSplit;\n"
+"	\n"
+"	if( gN[cellIdx] == 0 ) \n"
+"		return;\n"
+"\n"
+"	const int start = gOffsets[cellIdx];\n"
+"	const int end = start + gN[cellIdx];\n"
+"\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsCurBatch = 0;\n"
+"		ldsNextBatch = 0;\n"
+"		ldsStart = start;\n"
+"	}\n"
+"\n"
+"\n"
+"	GROUP_LDS_BARRIER;\n"
+"\n"
+"	int idx=ldsStart+lIdx;\n"
+"	while (ldsCurBatch < maxBatch)\n"
+"	{\n"
+"		for(; idx<end; )\n"
+"		{\n"
+"			if (gConstraints[idx].m_batchIdx == ldsCurBatch)\n"
+"			{\n"
+"\n"
+"					solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"\n"
+"				 idx+=64;\n"
+"			} else\n"
+"			{\n"
+"				break;\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		if( lIdx == 0 )\n"
+"		{\n"
+"			ldsCurBatch++;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"	\n"
+"    \n"
+"}\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl

@@ -0,0 +1,660 @@
+
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+Matrix3x3 mtZero();
+
+__inline
+Matrix3x3 mtIdentity();
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m);
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+__inline
+Matrix3x3 mtZero()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(0.f);
+	m.m_row[1] = (float4)(0.f);
+	m.m_row[2] = (float4)(0.f);
+	return m;
+}
+
+__inline
+Matrix3x3 mtIdentity()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(1,0,0,0);
+	m.m_row[1] = (float4)(0,1,0,0);
+	m.m_row[2] = (float4)(0,0,1,0);
+	return m;
+}
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+__inline
+Matrix3x3 qtGetRotationMatrix(Quaternion q);
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+__inline
+Matrix3x3 qtGetRotationMatrix(Quaternion quat)
+{
+	float4 quat2 = (float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);
+	Matrix3x3 out;
+
+	out.m_row[0].x=1-2*quat2.y-2*quat2.z;
+	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;
+	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;
+	out.m_row[0].w = 0.f;
+
+	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;
+	out.m_row[1].y=1-2*quat2.x-2*quat2.z;
+	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;
+	out.m_row[1].w = 0.f;
+
+	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;
+	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;
+	out.m_row[2].z=1-2*quat2.x-2*quat2.y;
+	out.m_row[2].w = 0.f;
+
+	return out;
+}
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+typedef struct
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+} Contact4;
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = make_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/(jmj0+jmj1+jmj2+jmj3);
+}
+
+
+
+ 
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+typedef struct
+{
+	int m_nContacts;
+	int m_staticIdx;
+	float m_scale;
+	int m_nSplit;
+} ConstBufferSSD;
+
+
+void btPlaneSpace1 (float4 n, float4* p, float4* q);
+ void btPlaneSpace1 (float4 n, float4* p, float4* q)
+{
+  if (fabs(n.z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n.y*n.y + n.z*n.z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n.z*k;
+	p[0].z = n.y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n.x*p[0].z;
+	q[0].z = n.x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n.x*n.x + n.y*n.y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n.y*k;
+	p[0].y = n.x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n.z*p[0].y;
+	q[0].y = n.z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+
+void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,
+	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, 
+	__global Contact4* src, float dt, float positionDrift, float positionConstraintCoeff,
+	Constraint4* dstC )
+{
+	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);
+	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);
+
+	float dtInv = 1.f/dt;
+	for(int ic=0; ic<4; ic++)
+	{
+		dstC->m_appliedRambdaDt[ic] = 0.f;
+	}
+	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;
+
+
+	dstC->m_linear = -src->m_worldNormal;
+	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );
+	for(int ic=0; ic<4; ic++)
+	{
+		float4 r0 = src->m_worldPos[ic] - posA;
+		float4 r1 = src->m_worldPos[ic] - posB;
+
+		if( ic >= src->m_worldNormal.w )//npoints
+		{
+			dstC->m_jacCoeffInv[ic] = 0.f;
+			continue;
+		}
+
+		float relVelN;
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(src->m_worldNormal, r0, r1, &linear, &angular0, &angular1);
+
+			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB );
+
+			relVelN = calcRelVel(linear, -linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB);
+
+			float e = 0.f;//src->getRestituitionCoeff();
+			if( relVelN*relVelN < 0.004f ) e = 0.f;
+
+			dstC->m_b[ic] = e*relVelN;
+			//float penetration = src->m_worldPos[ic].w;
+			dstC->m_b[ic] += (src->m_worldPos[ic].w + positionDrift)*positionConstraintCoeff*dtInv;
+			dstC->m_appliedRambdaDt[ic] = 0.f;
+		}
+	}
+
+	if( src->m_worldNormal.w > 0 )//npoints
+	{	//	prepare friction
+		float4 center = make_float4(0.f);
+		for(int i=0; i<src->m_worldNormal.w; i++) 
+			center += src->m_worldPos[i];
+		center /= (float)src->m_worldNormal.w;
+
+		float4 tangent[2];
+		btPlaneSpace1(src->m_worldNormal,&tangent[0],&tangent[1]);
+		
+		float4 r[2];
+		r[0] = center - posA;
+		r[1] = center - posB;
+
+		for(int i=0; i<2; i++)
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);
+
+			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB );
+			dstC->m_fAppliedRambdaDt[i] = 0.f;
+		}
+		dstC->m_center = center;
+	}
+
+	for(int i=0; i<4; i++)
+	{
+		if( i<src->m_worldNormal.w )
+		{
+			dstC->m_worldPos[i] = src->m_worldPos[i];
+		}
+		else
+		{
+			dstC->m_worldPos[i] = make_float4(0.f);
+		}
+	}
+}
+
+typedef struct
+{
+	int m_nContacts;
+	float m_dt;
+	float m_positionDrift;
+	float m_positionConstraintCoeff;
+} ConstBufferCTC;
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ContactToConstraintKernel(__global Contact4* gContact, __global Body* gBodies, __global Shape* gShapes, __global Constraint4* gConstraintOut, 
+int nContacts,
+float dt,
+float positionDrift,
+float positionConstraintCoeff
+)
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		float4 posA = gBodies[aIdx].m_pos;
+		float4 linVelA = gBodies[aIdx].m_linVel;
+		float4 angVelA = gBodies[aIdx].m_angVel;
+		float invMassA = gBodies[aIdx].m_invMass;
+		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+		float4 posB = gBodies[bIdx].m_pos;
+		float4 linVelB = gBodies[bIdx].m_linVel;
+		float4 angVelB = gBodies[bIdx].m_angVel;
+		float invMassB = gBodies[bIdx].m_invMass;
+		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+		Constraint4 cs;
+
+    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,
+			&cs );
+		
+		cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+		gConstraintOut[gIdx] = cs;
+	}
+}
+
+
+
+
+

src/Bullet3OpenCL/RigidBody/kernels/solverSetup.h

@@ -0,0 +1,664 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverSetupCL= \
+"\n"
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"\n"
+"\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(0.f);\n"
+"	m.m_row[1] = (float4)(0.f);\n"
+"	m.m_row[2] = (float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(1,0,0,0);\n"
+"	m.m_row[1] = (float4)(0,1,0,0);\n"
+"	m.m_row[2] = (float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 qtGetRotationMatrix(Quaternion q);\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 qtGetRotationMatrix(Quaternion quat)\n"
+"{\n"
+"	float4 quat2 = (float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	Matrix3x3 out;\n"
+"\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"\n"
+"	return out;\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"} Contact4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = make_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+" \n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"\n"
+"\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nContacts;\n"
+"	int m_staticIdx;\n"
+"	float m_scale;\n"
+"	int m_nSplit;\n"
+"} ConstBufferSSD;\n"
+"\n"
+"\n"
+"void b3PlaneSpace1 (float4 n, float4* p, float4* q);\n"
+" void b3PlaneSpace1 (float4 n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n.z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n.y*n.y + n.z*n.z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n.z*k;\n"
+"	p[0].z = n.y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n.x*p[0].z;\n"
+"	q[0].z = n.x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n.x*n.x + n.y*n.y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n.y*k;\n"
+"	p[0].y = n.x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n.z*p[0].y;\n"
+"	q[0].y = n.z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"\n"
+"\n"
+"void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,\n"
+"	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, \n"
+"	__global Contact4* src, float dt, float positionDrift, float positionConstraintCoeff,\n"
+"	Constraint4* dstC )\n"
+"{\n"
+"	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);\n"
+"	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);\n"
+"\n"
+"	float dtInv = 1.f/dt;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"	}\n"
+"	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;\n"
+"\n"
+"\n"
+"	dstC->m_linear = -src->m_worldNormal;\n"
+"	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		float4 r0 = src->m_worldPos[ic] - posA;\n"
+"		float4 r1 = src->m_worldPos[ic] - posB;\n"
+"\n"
+"		if( ic >= src->m_worldNormal.w )//npoints\n"
+"		{\n"
+"			dstC->m_jacCoeffInv[ic] = 0.f;\n"
+"			continue;\n"
+"		}\n"
+"\n"
+"		float relVelN;\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(src->m_worldNormal, r0, r1, &linear, &angular0, &angular1);\n"
+"\n"
+"			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB );\n"
+"\n"
+"			relVelN = calcRelVel(linear, -linear, angular0, angular1,\n"
+"				linVelA, angVelA, linVelB, angVelB);\n"
+"\n"
+"			float e = 0.f;//src->getRestituitionCoeff();\n"
+"			if( relVelN*relVelN < 0.004f ) e = 0.f;\n"
+"\n"
+"			dstC->m_b[ic] = e*relVelN;\n"
+"			//float penetration = src->m_worldPos[ic].w;\n"
+"			dstC->m_b[ic] += (src->m_worldPos[ic].w + positionDrift)*positionConstraintCoeff*dtInv;\n"
+"			dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"		}\n"
+"	}\n"
+"\n"
+"	if( src->m_worldNormal.w > 0 )//npoints\n"
+"	{	//	prepare friction\n"
+"		float4 center = make_float4(0.f);\n"
+"		for(int i=0; i<src->m_worldNormal.w; i++) \n"
+"			center += src->m_worldPos[i];\n"
+"		center /= (float)src->m_worldNormal.w;\n"
+"\n"
+"		float4 tangent[2];\n"
+"		b3PlaneSpace1(src->m_worldNormal,&tangent[0],&tangent[1]);\n"
+"		\n"
+"		float4 r[2];\n"
+"		r[0] = center - posA;\n"
+"		r[1] = center - posB;\n"
+"\n"
+"		for(int i=0; i<2; i++)\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);\n"
+"\n"
+"			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB );\n"
+"			dstC->m_fAppliedRambdaDt[i] = 0.f;\n"
+"		}\n"
+"		dstC->m_center = center;\n"
+"	}\n"
+"\n"
+"	for(int i=0; i<4; i++)\n"
+"	{\n"
+"		if( i<src->m_worldNormal.w )\n"
+"		{\n"
+"			dstC->m_worldPos[i] = src->m_worldPos[i];\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			dstC->m_worldPos[i] = make_float4(0.f);\n"
+"		}\n"
+"	}\n"
+"}\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nContacts;\n"
+"	float m_dt;\n"
+"	float m_positionDrift;\n"
+"	float m_positionConstraintCoeff;\n"
+"} ConstBufferCTC;\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ContactToConstraintKernel(__global Contact4* gContact, __global Body* gBodies, __global Shape* gShapes, __global Constraint4* gConstraintOut, \n"
+"int nContacts,\n"
+"float dt,\n"
+"float positionDrift,\n"
+"float positionConstraintCoeff\n"
+")\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);\n"
+"		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);\n"
+"\n"
+"		float4 posA = gBodies[aIdx].m_pos;\n"
+"		float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"		float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"		float invMassA = gBodies[aIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"		float4 posB = gBodies[bIdx].m_pos;\n"
+"		float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"		float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"		float invMassB = gBodies[bIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"\n"
+"		Constraint4 cs;\n"
+"\n"
+"    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,\n"
+"			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,\n"
+"			&cs );\n"
+"		\n"
+"		cs.m_batchIdx = gContact[gIdx].m_batchIdx;\n"
+"\n"
+"		gConstraintOut[gIdx] = cs;\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl

@@ -0,0 +1,494 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+Matrix3x3 mtZero();
+
+__inline
+Matrix3x3 mtIdentity();
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m);
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+__inline
+Matrix3x3 mtZero()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(0.f);
+	m.m_row[1] = (float4)(0.f);
+	m.m_row[2] = (float4)(0.f);
+	return m;
+}
+
+__inline
+Matrix3x3 mtIdentity()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(1,0,0,0);
+	m.m_row[1] = (float4)(0,1,0,0);
+	m.m_row[2] = (float4)(0,0,1,0);
+	return m;
+}
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+typedef struct
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+} Contact4;
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+
+ 
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+//	others
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ReorderContactKernel(__global Contact4* in, __global Contact4* out, __global int2* sortData, int4 cb )
+{
+	int nContacts = cb.x;
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int srcIdx = sortData[gIdx].y;
+		out[gIdx] = in[srcIdx];
+	}
+}
+
+typedef struct
+{
+	int m_nContacts;
+	int m_staticIdx;
+	float m_scale;
+	int m_nSplit;
+} ConstBufferSSD;
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetSortDataKernel(__global Contact4* gContact, __global Body* gBodies, __global int2* gSortDataOut, 
+int nContacts,
+float scale,
+int N_SPLIT
+)
+
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		int idx = (gContact[gIdx].m_bodyAPtrAndSignBit<0)? bIdx: aIdx;
+		float4 p = gBodies[idx].m_pos;
+		int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);
+		int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);
+
+		gSortDataOut[gIdx].x = (xIdx+zIdx*N_SPLIT);
+		gSortDataOut[gIdx].y = gIdx;
+	}
+	else
+	{
+		gSortDataOut[gIdx].x = 0xffffffff;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void CopyConstraintKernel(__global Contact4* gIn, __global Contact4* gOut, int4 cb )
+{
+	int gIdx = GET_GLOBAL_IDX;
+	if( gIdx < cb.x )
+	{
+		gOut[gIdx] = gIn[gIdx];
+	}
+}
+
+
+

src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.h

@@ -0,0 +1,498 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverSetup2CL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"\n"
+"\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(0.f);\n"
+"	m.m_row[1] = (float4)(0.f);\n"
+"	m.m_row[2] = (float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(1,0,0,0);\n"
+"	m.m_row[1] = (float4)(0,1,0,0);\n"
+"	m.m_row[2] = (float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"} Contact4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"\n"
+"\n"
+" \n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"//	others\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ReorderContactKernel(__global Contact4* in, __global Contact4* out, __global int2* sortData, int4 cb )\n"
+"{\n"
+"	int nContacts = cb.x;\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int srcIdx = sortData[gIdx].y;\n"
+"		out[gIdx] = in[srcIdx];\n"
+"	}\n"
+"}\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_nContacts;\n"
+"	int m_staticIdx;\n"
+"	float m_scale;\n"
+"	int m_nSplit;\n"
+"} ConstBufferSSD;\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetSortDataKernel(__global Contact4* gContact, __global Body* gBodies, __global int2* gSortDataOut, \n"
+"int nContacts,\n"
+"float scale,\n"
+"int N_SPLIT\n"
+")\n"
+"\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);\n"
+"		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);\n"
+"\n"
+"		int idx = (gContact[gIdx].m_bodyAPtrAndSignBit<0)? bIdx: aIdx;\n"
+"		float4 p = gBodies[idx].m_pos;\n"
+"		int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);\n"
+"		int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (N_SPLIT-1);\n"
+"\n"
+"		gSortDataOut[gIdx].x = (xIdx+zIdx*N_SPLIT);\n"
+"		gSortDataOut[gIdx].y = gIdx;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		gSortDataOut[gIdx].x = 0xffffffff;\n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void CopyConstraintKernel(__global Contact4* gIn, __global Contact4* gOut, int4 cb )\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < cb.x )\n"
+"	{\n"
+"		gOut[gIdx] = gIn[gIdx];\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/solverUtils.cl

@@ -0,0 +1,971 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a1, float4 b1)
+{
+
+	float4 	a=make_float4(a1.xyz,0.f);
+	float4 	b=make_float4(b1.xyz,0.f);
+	//float4 	a=a1;
+	//float4 	b=b1;
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+Matrix3x3 mtZero();
+
+__inline
+Matrix3x3 mtIdentity();
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m);
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+__inline
+Matrix3x3 mtZero()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(0.f);
+	m.m_row[1] = (float4)(0.f);
+	m.m_row[2] = (float4)(0.f);
+	return m;
+}
+
+__inline
+Matrix3x3 mtIdentity()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(1,0,0,0);
+	m.m_row[1] = (float4)(0,1,0,0);
+	m.m_row[2] = (float4)(0,0,1,0);
+	return m;
+}
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+	int m_batchIdx;
+	u32 m_paddings;
+} Constraint4;
+
+typedef struct
+{
+	float4 m_worldPos[4];
+	float4 m_worldNormal;
+	u32 m_coeffs;
+	int m_batchIdx;
+
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+} Contact4;
+
+
+__kernel void CountBodiesKernel(__global Contact4* manifoldPtr, __global unsigned int* bodyCount, __global int2* contactConstraintOffsets, int numContactManifolds, int fixedBodyIndex)
+{
+	int i = GET_GLOBAL_IDX;
+	
+	if( i < numContactManifolds)
+	{
+		int pa = manifoldPtr[i].m_bodyAPtrAndSignBit;
+		bool isFixedA = (pa <0) || (pa == fixedBodyIndex);
+		int bodyIndexA = abs(pa);
+		if (!isFixedA)
+		{
+			 AtomInc1(bodyCount[bodyIndexA],contactConstraintOffsets[i].x);
+		}
+		barrier(CLK_GLOBAL_MEM_FENCE);
+		int pb = manifoldPtr[i].m_bodyBPtrAndSignBit;
+		bool isFixedB = (pb <0) || (pb == fixedBodyIndex);
+		int bodyIndexB = abs(pb);
+		if (!isFixedB)
+		{
+			AtomInc1(bodyCount[bodyIndexB],contactConstraintOffsets[i].y);
+		} 
+	}
+}
+
+__kernel void ClearVelocitiesKernel(__global float4* linearVelocities,__global float4* angularVelocities, int numSplitBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	
+	if( i < numSplitBodies)
+	{
+		linearVelocities[i] = make_float4(0);
+		angularVelocities[i] = make_float4(0);
+	}
+}
+
+
+__kernel void AverageVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,
+__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numBodies)
+	{
+		if (gBodies[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			float factor = 1.f/((float)count);
+			float4 averageLinVel = make_float4(0.f);
+			float4 averageAngVel = make_float4(0.f);
+			
+			for (int j=0;j<count;j++)
+			{
+				averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+				averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+			}
+			
+			for (int j=0;j<count;j++)
+			{
+				deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+				deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+			}
+			
+		}//bodies[i].m_invMass
+	}//i<numBodies
+}
+
+
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = make_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1, float countA, float countB)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/((jmj0+jmj1)*countA+(jmj2+jmj3)*countB);
+}
+
+
+void btPlaneSpace1 (float4 n, float4* p, float4* q);
+ void btPlaneSpace1 (float4 n, float4* p, float4* q)
+{
+  if (fabs(n.z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n.y*n.y + n.z*n.z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n.z*k;
+	p[0].z = n.y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n.x*p[0].z;
+	q[0].z = n.x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n.x*n.x + n.y*n.y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n.y*k;
+	p[0].y = n.x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n.z*p[0].y;
+	q[0].y = n.z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+
+
+
+
+void solveContact(__global Constraint4* cs,
+			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB,
+			float4* dLinVelA, float4* dAngVelA, float4* dLinVelB, float4* dAngVelB)
+{
+	float minRambdaDt = 0;
+	float maxRambdaDt = FLT_MAX;
+
+	for(int ic=0; ic<4; ic++)
+	{
+		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;
+
+		float4 angular0, angular1, linear;
+		float4 r0 = cs->m_worldPos[ic] - posA;
+		float4 r1 = cs->m_worldPos[ic] - posB;
+		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );
+	
+
+
+		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, 
+			*linVelA+*dLinVelA, *angVelA+*dAngVelA, *linVelB+*dLinVelB, *angVelB+*dAngVelB ) + cs->m_b[ic];
+		rambdaDt *= cs->m_jacCoeffInv[ic];
+
+		
+		{
+			float prevSum = cs->m_appliedRambdaDt[ic];
+			float updated = prevSum;
+			updated += rambdaDt;
+			updated = max2( updated, minRambdaDt );
+			updated = min2( updated, maxRambdaDt );
+			rambdaDt = updated - prevSum;
+			cs->m_appliedRambdaDt[ic] = updated;
+		}
+
+			
+		float4 linImp0 = invMassA*linear*rambdaDt;
+		float4 linImp1 = invMassB*(-linear)*rambdaDt;
+		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+
+		
+		if (invMassA)
+		{
+			*dLinVelA += linImp0;
+			*dAngVelA += angImp0;
+		}
+		if (invMassB)
+		{
+			*dLinVelB += linImp1;
+			*dAngVelB += angImp1;
+		}
+	}
+}
+
+
+//	solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,contactConstraintOffsets,offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+
+
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs, 
+__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)
+{
+
+	//float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+			
+	float4 dLinVelA = make_float4(0,0,0,0);
+	float4 dAngVelA = make_float4(0,0,0,0);
+	float4 dLinVelB = make_float4(0,0,0,0);
+	float4 dAngVelB = make_float4(0,0,0,0);
+			
+	int bodyOffsetA = offsetSplitBodies[aIdx];
+	int constraintOffsetA = contactConstraintOffsets[0].x;
+	int splitIndexA = bodyOffsetA+constraintOffsetA;
+	
+	if (invMassA)
+	{
+		dLinVelA = deltaLinearVelocities[splitIndexA];
+		dAngVelA = deltaAngularVelocities[splitIndexA];
+	}
+
+	int bodyOffsetB = offsetSplitBodies[bIdx];
+	int constraintOffsetB = contactConstraintOffsets[0].y;
+	int splitIndexB= bodyOffsetB+constraintOffsetB;
+
+	if (invMassB)
+	{
+		dLinVelB = deltaLinearVelocities[splitIndexB];
+		dAngVelB = deltaAngularVelocities[splitIndexB];
+	}
+
+	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+			posB, &linVelB, &angVelB, invMassB, invInertiaB ,&dLinVelA, &dAngVelA, &dLinVelB, &dAngVelB);
+
+	if (invMassA)
+	{
+		deltaLinearVelocities[splitIndexA] = dLinVelA;
+		deltaAngularVelocities[splitIndexA] = dAngVelA;
+	} 
+	if (invMassB)
+	{
+		deltaLinearVelocities[splitIndexB] = dLinVelB;
+		deltaAngularVelocities[splitIndexB] = dAngVelB;
+	}
+
+}
+
+
+__kernel void SolveContactJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,
+__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,
+float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds
+)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numManifolds)
+	{
+		solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+	}
+}
+
+
+
+
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs,
+							__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+							__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)
+{
+	float frictionCoeff = 0.7f;//ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+	
+
+	float4 dLinVelA = make_float4(0,0,0,0);
+	float4 dAngVelA = make_float4(0,0,0,0);
+	float4 dLinVelB = make_float4(0,0,0,0);
+	float4 dAngVelB = make_float4(0,0,0,0);
+			
+	int bodyOffsetA = offsetSplitBodies[aIdx];
+	int constraintOffsetA = contactConstraintOffsets[0].x;
+	int splitIndexA = bodyOffsetA+constraintOffsetA;
+	
+	if (invMassA)
+	{
+		dLinVelA = deltaLinearVelocities[splitIndexA];
+		dAngVelA = deltaAngularVelocities[splitIndexA];
+	}
+
+	int bodyOffsetB = offsetSplitBodies[bIdx];
+	int constraintOffsetB = contactConstraintOffsets[0].y;
+	int splitIndexB= bodyOffsetB+constraintOffsetB;
+
+	if (invMassB)
+	{
+		dLinVelB = deltaLinearVelocities[splitIndexB];
+		dAngVelB = deltaAngularVelocities[splitIndexB];
+	}
+
+
+
+
+	{
+		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+		float sum = 0;
+		for(int j=0; j<4; j++)
+		{
+			sum +=ldsCs[0].m_appliedRambdaDt[j];
+		}
+		frictionCoeff = 0.7f;
+		for(int j=0; j<4; j++)
+		{
+			maxRambdaDt[j] = frictionCoeff*sum;
+			minRambdaDt[j] = -maxRambdaDt[j];
+		}
+
+		
+//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );
+		
+		
+		{
+			
+			__global Constraint4* cs = ldsCs;
+			
+			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;
+			const float4 center = cs->m_center;
+			
+			float4 n = -cs->m_linear;
+			
+			float4 tangent[2];
+			btPlaneSpace1(n,&tangent[0],&tangent[1]);
+			float4 angular0, angular1, linear;
+			float4 r0 = center - posA;
+			float4 r1 = center - posB;
+			for(int i=0; i<2; i++)
+			{
+				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );
+				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+											linVelA+dLinVelA, angVelA+dAngVelA, linVelB+dLinVelB, angVelB+dAngVelB );
+				rambdaDt *= cs->m_fJacCoeffInv[i];
+				
+				{
+					float prevSum = cs->m_fAppliedRambdaDt[i];
+					float updated = prevSum;
+					updated += rambdaDt;
+					updated = max2( updated, minRambdaDt[i] );
+					updated = min2( updated, maxRambdaDt[i] );
+					rambdaDt = updated - prevSum;
+					cs->m_fAppliedRambdaDt[i] = updated;
+				}
+				
+				float4 linImp0 = invMassA*linear*rambdaDt;
+				float4 linImp1 = invMassB*(-linear)*rambdaDt;
+				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+				
+				dLinVelA += linImp0;
+				dAngVelA += angImp0;
+				dLinVelB += linImp1;
+				dAngVelB += angImp1;
+			}
+			{	//	angular damping for point constraint
+				float4 ab = normalize3( posB - posA );
+				float4 ac = normalize3( center - posA );
+				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))
+				{
+					float angNA = dot3F4( n, angVelA );
+					float angNB = dot3F4( n, angVelB );
+					
+					dAngVelA -= (angNA*0.1f)*n;
+					dAngVelB -= (angNB*0.1f)*n;
+				}
+			}
+		}
+
+		
+		
+	}
+
+	if (invMassA)
+	{
+		deltaLinearVelocities[splitIndexA] = dLinVelA;
+		deltaAngularVelocities[splitIndexA] = dAngVelA;
+	} 
+	if (invMassB)
+	{
+		deltaLinearVelocities[splitIndexB] = dLinVelB;
+		deltaAngularVelocities[splitIndexB] = dAngVelB;
+	}
+ 
+
+}
+
+
+__kernel void SolveFrictionJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,
+										__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+										__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,
+										float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds
+)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numManifolds)
+	{
+		solveFrictionConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+	}
+}
+
+
+__kernel void UpdateBodyVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,
+									__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numBodies)
+	{
+		if (gBodies[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			if (count)
+			{
+				gBodies[i].m_linVel += deltaLinearVelocities[bodyOffset];
+				gBodies[i].m_angVel += deltaAngularVelocities[bodyOffset];
+			}
+		}
+	}
+}
+
+
+
+void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,
+	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, 
+	__global Contact4* src, float dt, float positionDrift, float positionConstraintCoeff,float countA, float countB,
+	Constraint4* dstC )
+{
+	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);
+	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);
+
+	float dtInv = 1.f/dt;
+	for(int ic=0; ic<4; ic++)
+	{
+		dstC->m_appliedRambdaDt[ic] = 0.f;
+	}
+	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;
+
+
+	dstC->m_linear = -src->m_worldNormal;
+	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );
+	for(int ic=0; ic<4; ic++)
+	{
+		float4 r0 = src->m_worldPos[ic] - posA;
+		float4 r1 = src->m_worldPos[ic] - posB;
+
+		if( ic >= src->m_worldNormal.w )//npoints
+		{
+			dstC->m_jacCoeffInv[ic] = 0.f;
+			continue;
+		}
+
+		float relVelN;
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(src->m_worldNormal, r0, r1, &linear, &angular0, &angular1);
+
+			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB , countA, countB);
+
+			relVelN = calcRelVel(linear, -linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB);
+
+			float e = 0.f;//src->getRestituitionCoeff();
+			if( relVelN*relVelN < 0.004f ) e = 0.f;
+
+			dstC->m_b[ic] = e*relVelN;
+			//float penetration = src->m_worldPos[ic].w;
+			dstC->m_b[ic] += (src->m_worldPos[ic].w + positionDrift)*positionConstraintCoeff*dtInv;
+			dstC->m_appliedRambdaDt[ic] = 0.f;
+		}
+	}
+
+	if( src->m_worldNormal.w > 0 )//npoints
+	{	//	prepare friction
+		float4 center = make_float4(0.f);
+		for(int i=0; i<src->m_worldNormal.w; i++) 
+			center += src->m_worldPos[i];
+		center /= (float)src->m_worldNormal.w;
+
+		float4 tangent[2];
+		btPlaneSpace1(src->m_worldNormal,&tangent[0],&tangent[1]);
+		
+		float4 r[2];
+		r[0] = center - posA;
+		r[1] = center - posB;
+
+		for(int i=0; i<2; i++)
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);
+
+			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA, countB);
+			dstC->m_fAppliedRambdaDt[i] = 0.f;
+		}
+		dstC->m_center = center;
+	}
+
+	for(int i=0; i<4; i++)
+	{
+		if( i<src->m_worldNormal.w )
+		{
+			dstC->m_worldPos[i] = src->m_worldPos[i];
+		}
+		else
+		{
+			dstC->m_worldPos[i] = make_float4(0.f);
+		}
+	}
+}
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ContactToConstraintSplitKernel(__global const Contact4* gContact, __global const Body* gBodies, __global const Shape* gShapes, __global Constraint4* gConstraintOut, 
+__global const unsigned int* bodyCount,
+int nContacts,
+float dt,
+float positionDrift,
+float positionConstraintCoeff
+)
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		float4 posA = gBodies[aIdx].m_pos;
+		float4 linVelA = gBodies[aIdx].m_linVel;
+		float4 angVelA = gBodies[aIdx].m_angVel;
+		float invMassA = gBodies[aIdx].m_invMass;
+		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+		float4 posB = gBodies[bIdx].m_pos;
+		float4 linVelB = gBodies[bIdx].m_linVel;
+		float4 angVelB = gBodies[bIdx].m_angVel;
+		float invMassB = gBodies[bIdx].m_invMass;
+		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+		Constraint4 cs;
+
+		float countA = invMassA ? (float)bodyCount[aIdx] : 1;
+		float countB = invMassB ? (float)bodyCount[bIdx] : 1;
+
+    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,countA,countB,
+			&cs  );
+		
+		cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+		gConstraintOut[gIdx] = cs;
+	}
+}

src/Bullet3OpenCL/RigidBody/kernels/solverUtils.h

@@ -0,0 +1,974 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverUtilsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"\n"
+"\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"\n"
+"\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"\n"
+"\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a1, float4 b1)\n"
+"{\n"
+"\n"
+"	float4 	a=make_float4(a1.xyz,0.f);\n"
+"	float4 	b=make_float4(b1.xyz,0.f);\n"
+"	//float4 	a=a1;\n"
+"	//float4 	b=b1;\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity();\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtZero()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(0.f);\n"
+"	m.m_row[1] = (float4)(0.f);\n"
+"	m.m_row[2] = (float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtIdentity()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(1,0,0,0);\n"
+"	m.m_row[1] = (float4)(0,1,0,0);\n"
+"	m.m_row[2] = (float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"#define WG_SIZE 64\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings;\n"
+"} Constraint4;\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_worldNormal;\n"
+"	u32 m_coeffs;\n"
+"	int m_batchIdx;\n"
+"\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"} Contact4;\n"
+"\n"
+"\n"
+"__kernel void CountBodiesKernel(__global Contact4* manifoldPtr, __global unsigned int* bodyCount, __global int2* contactConstraintOffsets, int numContactManifolds, int fixedBodyIndex)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( i < numContactManifolds)\n"
+"	{\n"
+"		int pa = manifoldPtr[i].m_bodyAPtrAndSignBit;\n"
+"		bool isFixedA = (pa <0) || (pa == fixedBodyIndex);\n"
+"		int bodyIndexA = abs(pa);\n"
+"		if (!isFixedA)\n"
+"		{\n"
+"			 AtomInc1(bodyCount[bodyIndexA],contactConstraintOffsets[i].x);\n"
+"		}\n"
+"		barrier(CLK_GLOBAL_MEM_FENCE);\n"
+"		int pb = manifoldPtr[i].m_bodyBPtrAndSignBit;\n"
+"		bool isFixedB = (pb <0) || (pb == fixedBodyIndex);\n"
+"		int bodyIndexB = abs(pb);\n"
+"		if (!isFixedB)\n"
+"		{\n"
+"			AtomInc1(bodyCount[bodyIndexB],contactConstraintOffsets[i].y);\n"
+"		} \n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel void ClearVelocitiesKernel(__global float4* linearVelocities,__global float4* angularVelocities, int numSplitBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( i < numSplitBodies)\n"
+"	{\n"
+"		linearVelocities[i] = make_float4(0);\n"
+"		angularVelocities[i] = make_float4(0);\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"__kernel void AverageVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,\n"
+"__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numBodies)\n"
+"	{\n"
+"		if (gBodies[i].m_invMass)\n"
+"		{\n"
+"			int bodyOffset = offsetSplitBodies[i];\n"
+"			int count = bodyCount[i];\n"
+"			float factor = 1.f/((float)count);\n"
+"			float4 averageLinVel = make_float4(0.f);\n"
+"			float4 averageAngVel = make_float4(0.f);\n"
+"			\n"
+"			for (int j=0;j<count;j++)\n"
+"			{\n"
+"				averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;\n"
+"				averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;\n"
+"			}\n"
+"			\n"
+"			for (int j=0;j<count;j++)\n"
+"			{\n"
+"				deltaLinearVelocities[bodyOffset+j] = averageLinVel;\n"
+"				deltaAngularVelocities[bodyOffset+j] = averageAngVel;\n"
+"			}\n"
+"			\n"
+"		}//bodies[i].m_invMass\n"
+"	}//i<numBodies\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = make_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"\n"
+"\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"\n"
+"\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1, float countA, float countB)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/((jmj0+jmj1)*countA+(jmj2+jmj3)*countB);\n"
+"}\n"
+"\n"
+"\n"
+"void b3PlaneSpace1 (float4 n, float4* p, float4* q);\n"
+" void b3PlaneSpace1 (float4 n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n.z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n.y*n.y + n.z*n.z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n.z*k;\n"
+"	p[0].z = n.y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n.x*p[0].z;\n"
+"	q[0].z = n.x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n.x*n.x + n.y*n.y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n.y*k;\n"
+"	p[0].y = n.x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n.z*p[0].y;\n"
+"	q[0].y = n.z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"void solveContact(__global Constraint4* cs,\n"
+"			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB,\n"
+"			float4* dLinVelA, float4* dAngVelA, float4* dLinVelB, float4* dAngVelB)\n"
+"{\n"
+"	float minRambdaDt = 0;\n"
+"	float maxRambdaDt = FLT_MAX;\n"
+"\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;\n"
+"\n"
+"		float4 angular0, angular1, linear;\n"
+"		float4 r0 = cs->m_worldPos[ic] - posA;\n"
+"		float4 r1 = cs->m_worldPos[ic] - posB;\n"
+"		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );\n"
+"	\n"
+"\n"
+"\n"
+"		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, \n"
+"			*linVelA+*dLinVelA, *angVelA+*dAngVelA, *linVelB+*dLinVelB, *angVelB+*dAngVelB ) + cs->m_b[ic];\n"
+"		rambdaDt *= cs->m_jacCoeffInv[ic];\n"
+"\n"
+"		\n"
+"		{\n"
+"			float prevSum = cs->m_appliedRambdaDt[ic];\n"
+"			float updated = prevSum;\n"
+"			updated += rambdaDt;\n"
+"			updated = max2( updated, minRambdaDt );\n"
+"			updated = min2( updated, maxRambdaDt );\n"
+"			rambdaDt = updated - prevSum;\n"
+"			cs->m_appliedRambdaDt[ic] = updated;\n"
+"		}\n"
+"\n"
+"			\n"
+"		float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"		float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"\n"
+"		\n"
+"		if (invMassA)\n"
+"		{\n"
+"			*dLinVelA += linImp0;\n"
+"			*dAngVelA += angImp0;\n"
+"		}\n"
+"		if (invMassB)\n"
+"		{\n"
+"			*dLinVelB += linImp1;\n"
+"			*dAngVelB += angImp1;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"//	solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,contactConstraintOffsets,offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"\n"
+"\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs, \n"
+"__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)\n"
+"{\n"
+"\n"
+"	//float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"\n"
+"			\n"
+"	float4 dLinVelA = make_float4(0,0,0,0);\n"
+"	float4 dAngVelA = make_float4(0,0,0,0);\n"
+"	float4 dLinVelB = make_float4(0,0,0,0);\n"
+"	float4 dAngVelB = make_float4(0,0,0,0);\n"
+"			\n"
+"	int bodyOffsetA = offsetSplitBodies[aIdx];\n"
+"	int constraintOffsetA = contactConstraintOffsets[0].x;\n"
+"	int splitIndexA = bodyOffsetA+constraintOffsetA;\n"
+"	\n"
+"	if (invMassA)\n"
+"	{\n"
+"		dLinVelA = deltaLinearVelocities[splitIndexA];\n"
+"		dAngVelA = deltaAngularVelocities[splitIndexA];\n"
+"	}\n"
+"\n"
+"	int bodyOffsetB = offsetSplitBodies[bIdx];\n"
+"	int constraintOffsetB = contactConstraintOffsets[0].y;\n"
+"	int splitIndexB= bodyOffsetB+constraintOffsetB;\n"
+"\n"
+"	if (invMassB)\n"
+"	{\n"
+"		dLinVelB = deltaLinearVelocities[splitIndexB];\n"
+"		dAngVelB = deltaAngularVelocities[splitIndexB];\n"
+"	}\n"
+"\n"
+"	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"			posB, &linVelB, &angVelB, invMassB, invInertiaB ,&dLinVelA, &dAngVelA, &dLinVelB, &dAngVelB);\n"
+"\n"
+"	if (invMassA)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexA] = dLinVelA;\n"
+"		deltaAngularVelocities[splitIndexA] = dAngVelA;\n"
+"	} \n"
+"	if (invMassB)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexB] = dLinVelB;\n"
+"		deltaAngularVelocities[splitIndexB] = dAngVelB;\n"
+"	}\n"
+"\n"
+"}\n"
+"\n"
+"\n"
+"__kernel void SolveContactJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,\n"
+"__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,\n"
+"float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds\n"
+")\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numManifolds)\n"
+"	{\n"
+"		solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs,\n"
+"							__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"							__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)\n"
+"{\n"
+"	float frictionCoeff = 0.7f;//ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"\n"
+"\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"	\n"
+"\n"
+"	float4 dLinVelA = make_float4(0,0,0,0);\n"
+"	float4 dAngVelA = make_float4(0,0,0,0);\n"
+"	float4 dLinVelB = make_float4(0,0,0,0);\n"
+"	float4 dAngVelB = make_float4(0,0,0,0);\n"
+"			\n"
+"	int bodyOffsetA = offsetSplitBodies[aIdx];\n"
+"	int constraintOffsetA = contactConstraintOffsets[0].x;\n"
+"	int splitIndexA = bodyOffsetA+constraintOffsetA;\n"
+"	\n"
+"	if (invMassA)\n"
+"	{\n"
+"		dLinVelA = deltaLinearVelocities[splitIndexA];\n"
+"		dAngVelA = deltaAngularVelocities[splitIndexA];\n"
+"	}\n"
+"\n"
+"	int bodyOffsetB = offsetSplitBodies[bIdx];\n"
+"	int constraintOffsetB = contactConstraintOffsets[0].y;\n"
+"	int splitIndexB= bodyOffsetB+constraintOffsetB;\n"
+"\n"
+"	if (invMassB)\n"
+"	{\n"
+"		dLinVelB = deltaLinearVelocities[splitIndexB];\n"
+"		dAngVelB = deltaAngularVelocities[splitIndexB];\n"
+"	}\n"
+"\n"
+"\n"
+"\n"
+"\n"
+"	{\n"
+"		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};\n"
+"		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};\n"
+"\n"
+"		float sum = 0;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			sum +=ldsCs[0].m_appliedRambdaDt[j];\n"
+"		}\n"
+"		frictionCoeff = 0.7f;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			maxRambdaDt[j] = frictionCoeff*sum;\n"
+"			minRambdaDt[j] = -maxRambdaDt[j];\n"
+"		}\n"
+"\n"
+"		\n"
+"//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );\n"
+"		\n"
+"		\n"
+"		{\n"
+"			\n"
+"			__global Constraint4* cs = ldsCs;\n"
+"			\n"
+"			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;\n"
+"			const float4 center = cs->m_center;\n"
+"			\n"
+"			float4 n = -cs->m_linear;\n"
+"			\n"
+"			float4 tangent[2];\n"
+"			b3PlaneSpace1(n,&tangent[0],&tangent[1]);\n"
+"			float4 angular0, angular1, linear;\n"
+"			float4 r0 = center - posA;\n"
+"			float4 r1 = center - posB;\n"
+"			for(int i=0; i<2; i++)\n"
+"			{\n"
+"				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );\n"
+"				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,\n"
+"											linVelA+dLinVelA, angVelA+dAngVelA, linVelB+dLinVelB, angVelB+dAngVelB );\n"
+"				rambdaDt *= cs->m_fJacCoeffInv[i];\n"
+"				\n"
+"				{\n"
+"					float prevSum = cs->m_fAppliedRambdaDt[i];\n"
+"					float updated = prevSum;\n"
+"					updated += rambdaDt;\n"
+"					updated = max2( updated, minRambdaDt[i] );\n"
+"					updated = min2( updated, maxRambdaDt[i] );\n"
+"					rambdaDt = updated - prevSum;\n"
+"					cs->m_fAppliedRambdaDt[i] = updated;\n"
+"				}\n"
+"				\n"
+"				float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"				float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"				\n"
+"				dLinVelA += linImp0;\n"
+"				dAngVelA += angImp0;\n"
+"				dLinVelB += linImp1;\n"
+"				dAngVelB += angImp1;\n"
+"			}\n"
+"			{	//	angular damping for point constraint\n"
+"				float4 ab = normalize3( posB - posA );\n"
+"				float4 ac = normalize3( center - posA );\n"
+"				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))\n"
+"				{\n"
+"					float angNA = dot3F4( n, angVelA );\n"
+"					float angNB = dot3F4( n, angVelB );\n"
+"					\n"
+"					dAngVelA -= (angNA*0.1f)*n;\n"
+"					dAngVelB -= (angNB*0.1f)*n;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"\n"
+"		\n"
+"		\n"
+"	}\n"
+"\n"
+"	if (invMassA)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexA] = dLinVelA;\n"
+"		deltaAngularVelocities[splitIndexA] = dAngVelA;\n"
+"	} \n"
+"	if (invMassB)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexB] = dLinVelB;\n"
+"		deltaAngularVelocities[splitIndexB] = dAngVelB;\n"
+"	}\n"
+" \n"
+"\n"
+"}\n"
+"\n"
+"\n"
+"__kernel void SolveFrictionJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,\n"
+"										__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"										__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,\n"
+"										float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds\n"
+")\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numManifolds)\n"
+"	{\n"
+"		solveFrictionConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"__kernel void UpdateBodyVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,\n"
+"									__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numBodies)\n"
+"	{\n"
+"		if (gBodies[i].m_invMass)\n"
+"		{\n"
+"			int bodyOffset = offsetSplitBodies[i];\n"
+"			int count = bodyCount[i];\n"
+"			if (count)\n"
+"			{\n"
+"				gBodies[i].m_linVel += deltaLinearVelocities[bodyOffset];\n"
+"				gBodies[i].m_angVel += deltaAngularVelocities[bodyOffset];\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,\n"
+"	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, \n"
+"	__global Contact4* src, float dt, float positionDrift, float positionConstraintCoeff,float countA, float countB,\n"
+"	Constraint4* dstC )\n"
+"{\n"
+"	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);\n"
+"	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);\n"
+"\n"
+"	float dtInv = 1.f/dt;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"	}\n"
+"	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;\n"
+"\n"
+"\n"
+"	dstC->m_linear = -src->m_worldNormal;\n"
+"	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		float4 r0 = src->m_worldPos[ic] - posA;\n"
+"		float4 r1 = src->m_worldPos[ic] - posB;\n"
+"\n"
+"		if( ic >= src->m_worldNormal.w )//npoints\n"
+"		{\n"
+"			dstC->m_jacCoeffInv[ic] = 0.f;\n"
+"			continue;\n"
+"		}\n"
+"\n"
+"		float relVelN;\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(src->m_worldNormal, r0, r1, &linear, &angular0, &angular1);\n"
+"\n"
+"			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB , countA, countB);\n"
+"\n"
+"			relVelN = calcRelVel(linear, -linear, angular0, angular1,\n"
+"				linVelA, angVelA, linVelB, angVelB);\n"
+"\n"
+"			float e = 0.f;//src->getRestituitionCoeff();\n"
+"			if( relVelN*relVelN < 0.004f ) e = 0.f;\n"
+"\n"
+"			dstC->m_b[ic] = e*relVelN;\n"
+"			//float penetration = src->m_worldPos[ic].w;\n"
+"			dstC->m_b[ic] += (src->m_worldPos[ic].w + positionDrift)*positionConstraintCoeff*dtInv;\n"
+"			dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"		}\n"
+"	}\n"
+"\n"
+"	if( src->m_worldNormal.w > 0 )//npoints\n"
+"	{	//	prepare friction\n"
+"		float4 center = make_float4(0.f);\n"
+"		for(int i=0; i<src->m_worldNormal.w; i++) \n"
+"			center += src->m_worldPos[i];\n"
+"		center /= (float)src->m_worldNormal.w;\n"
+"\n"
+"		float4 tangent[2];\n"
+"		b3PlaneSpace1(src->m_worldNormal,&tangent[0],&tangent[1]);\n"
+"		\n"
+"		float4 r[2];\n"
+"		r[0] = center - posA;\n"
+"		r[1] = center - posB;\n"
+"\n"
+"		for(int i=0; i<2; i++)\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);\n"
+"\n"
+"			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA, countB);\n"
+"			dstC->m_fAppliedRambdaDt[i] = 0.f;\n"
+"		}\n"
+"		dstC->m_center = center;\n"
+"	}\n"
+"\n"
+"	for(int i=0; i<4; i++)\n"
+"	{\n"
+"		if( i<src->m_worldNormal.w )\n"
+"		{\n"
+"			dstC->m_worldPos[i] = src->m_worldPos[i];\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			dstC->m_worldPos[i] = make_float4(0.f);\n"
+"		}\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ContactToConstraintSplitKernel(__global const Contact4* gContact, __global const Body* gBodies, __global const Shape* gShapes, __global Constraint4* gConstraintOut, \n"
+"__global const unsigned int* bodyCount,\n"
+"int nContacts,\n"
+"float dt,\n"
+"float positionDrift,\n"
+"float positionConstraintCoeff\n"
+")\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);\n"
+"		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);\n"
+"\n"
+"		float4 posA = gBodies[aIdx].m_pos;\n"
+"		float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"		float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"		float invMassA = gBodies[aIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"\n"
+"		float4 posB = gBodies[bIdx].m_pos;\n"
+"		float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"		float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"		float invMassB = gBodies[bIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"\n"
+"		Constraint4 cs;\n"
+"\n"
+"		float countA = invMassA ? (float)bodyCount[aIdx] : 1;\n"
+"		float countB = invMassB ? (float)bodyCount[bIdx] : 1;\n"
+"\n"
+"    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,\n"
+"			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,countA,countB,\n"
+"			&cs  );\n"
+"		\n"
+"		cs.m_batchIdx = gContact[gIdx].m_batchIdx;\n"
+"\n"
+"		gConstraintOut[gIdx] = cs;\n"
+"	}\n"
+"}\n"
+;

src/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.cl

@@ -0,0 +1,195 @@
+
+#define SHAPE_CONVEX_HULL 3
+
+typedef float4 Quaternion;
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = (float4)(a.xyz,0.f);
+	float4 b1 = (float4)(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+typedef struct
+{
+	float4	m_row[3];
+} Matrix3x3;
+
+typedef unsigned int u32;
+
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct Collidable
+{
+	int m_unused1;
+	int m_unused2;
+	int m_shapeType;
+	int m_shapeIndex;
+} Collidable;
+
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+
+__inline
+Matrix3x3 qtGetRotationMatrix(float4 quat)
+{
+	float4 quat2 = (float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);
+	Matrix3x3 out;
+
+	out.m_row[0].x=fabs(1-2*quat2.y-2*quat2.z);
+	out.m_row[0].y=fabs(2*quat.x*quat.y-2*quat.w*quat.z);
+	out.m_row[0].z=fabs(2*quat.x*quat.z+2*quat.w*quat.y);
+	out.m_row[0].w = 0.f;
+
+	out.m_row[1].x=fabs(2*quat.x*quat.y+2*quat.w*quat.z);
+	out.m_row[1].y=fabs(1-2*quat2.x-2*quat2.z);
+	out.m_row[1].z=fabs(2*quat.y*quat.z-2*quat.w*quat.x);
+	out.m_row[1].w = 0.f;
+
+	out.m_row[2].x=fabs(2*quat.x*quat.z-2*quat.w*quat.y);
+	out.m_row[2].y=fabs(2*quat.y*quat.z+2*quat.w*quat.x);
+	out.m_row[2].z=fabs(1-2*quat2.x-2*quat2.y);
+	out.m_row[2].w = 0.f;
+
+	return out;
+}
+
+
+typedef struct 
+{
+	float			fx;
+	float			fy;
+	float			fz;
+	int	uw;
+} btAABBCL;
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+
+__kernel void initializeGpuAabbsFull(  const int numNodes, __global Body* gBodies,__global Collidable* collidables, __global btAABBCL* plocalShapeAABB, __global btAABBCL* pAABB)
+{
+	int nodeID = get_global_id(0);
+		
+	if( nodeID < numNodes )
+	{
+		float4 position = gBodies[nodeID].m_pos;
+		float4 orientation = gBodies[nodeID].m_quat;
+		
+			
+		int collidableIndex = gBodies[nodeID].m_collidableIdx;
+		int shapeIndex = collidables[collidableIndex].m_shapeIndex;
+			
+		if (shapeIndex>=0)
+		{
+			btAABBCL minAabb = plocalShapeAABB[collidableIndex*2];
+			btAABBCL maxAabb = plocalShapeAABB[collidableIndex*2+1];
+				
+			float4 halfExtents = ((float4)(maxAabb.fx - minAabb.fx,maxAabb.fy - minAabb.fy,maxAabb.fz - minAabb.fz,0.f))*0.5f;
+			float4 localCenter = ((float4)(maxAabb.fx + minAabb.fx,maxAabb.fy + minAabb.fy,maxAabb.fz + minAabb.fz,0.f))*0.5f;
+				
+			float4 worldCenter = transform(&localCenter,&position,&orientation);
+				
+			Matrix3x3 abs_b = qtGetRotationMatrix(orientation);
+			float4 extent = (float4) (	dot(abs_b.m_row[0],halfExtents),dot(abs_b.m_row[1],halfExtents),dot(abs_b.m_row[2],halfExtents),0.f);
+			
+	
+			pAABB[nodeID*2].fx = worldCenter.x-extent.x;
+			pAABB[nodeID*2].fy = worldCenter.y-extent.y;
+			pAABB[nodeID*2].fz = worldCenter.z-extent.z;
+			pAABB[nodeID*2].uw = nodeID;
+	
+			pAABB[nodeID*2+1].fx = worldCenter.x+extent.x;
+			pAABB[nodeID*2+1].fy = worldCenter.y+extent.y;
+			pAABB[nodeID*2+1].fz = worldCenter.z+extent.z;
+			pAABB[nodeID*2+1].uw = gBodies[nodeID].m_invMass==0.f? 0 : 1;
+		}
+	} 
+}

src/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.h

@@ -0,0 +1,199 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* updateAabbsKernelCL= \
+"\n"
+"#define SHAPE_CONVEX_HULL 3\n"
+"\n"
+"typedef float4 Quaternion;\n"
+"\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = (float4)(a.xyz,0.f);\n"
+"	float4 b1 = (float4)(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"\n"
+"\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4	m_row[3];\n"
+"} Matrix3x3;\n"
+"\n"
+"typedef unsigned int u32;\n"
+"\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"\n"
+"	u32 m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"\n"
+"typedef struct Collidable\n"
+"{\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"} Collidable;\n"
+"\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"\n"
+"\n"
+"__inline\n"
+"Matrix3x3 qtGetRotationMatrix(float4 quat)\n"
+"{\n"
+"	float4 quat2 = (float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	Matrix3x3 out;\n"
+"\n"
+"	out.m_row[0].x=fabs(1-2*quat2.y-2*quat2.z);\n"
+"	out.m_row[0].y=fabs(2*quat.x*quat.y-2*quat.w*quat.z);\n"
+"	out.m_row[0].z=fabs(2*quat.x*quat.z+2*quat.w*quat.y);\n"
+"	out.m_row[0].w = 0.f;\n"
+"\n"
+"	out.m_row[1].x=fabs(2*quat.x*quat.y+2*quat.w*quat.z);\n"
+"	out.m_row[1].y=fabs(1-2*quat2.x-2*quat2.z);\n"
+"	out.m_row[1].z=fabs(2*quat.y*quat.z-2*quat.w*quat.x);\n"
+"	out.m_row[1].w = 0.f;\n"
+"\n"
+"	out.m_row[2].x=fabs(2*quat.x*quat.z-2*quat.w*quat.y);\n"
+"	out.m_row[2].y=fabs(2*quat.y*quat.z+2*quat.w*quat.x);\n"
+"	out.m_row[2].z=fabs(1-2*quat2.x-2*quat2.y);\n"
+"	out.m_row[2].w = 0.f;\n"
+"\n"
+"	return out;\n"
+"}\n"
+"\n"
+"\n"
+"typedef struct \n"
+"{\n"
+"	float			fx;\n"
+"	float			fy;\n"
+"	float			fz;\n"
+"	int	uw;\n"
+"} b3AABBCL;\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"\n"
+"\n"
+"\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"\n"
+"\n"
+"__kernel void initializeGpuAabbsFull(  const int numNodes, __global Body* gBodies,__global Collidable* collidables, __global b3AABBCL* plocalShapeAABB, __global b3AABBCL* pAABB)\n"
+"{\n"
+"	int nodeID = get_global_id(0);\n"
+"		\n"
+"	if( nodeID < numNodes )\n"
+"	{\n"
+"		float4 position = gBodies[nodeID].m_pos;\n"
+"		float4 orientation = gBodies[nodeID].m_quat;\n"
+"		\n"
+"			\n"
+"		int collidableIndex = gBodies[nodeID].m_collidableIdx;\n"
+"		int shapeIndex = collidables[collidableIndex].m_shapeIndex;\n"
+"			\n"
+"		if (shapeIndex>=0)\n"
+"		{\n"
+"			b3AABBCL minAabb = plocalShapeAABB[collidableIndex*2];\n"
+"			b3AABBCL maxAabb = plocalShapeAABB[collidableIndex*2+1];\n"
+"				\n"
+"			float4 halfExtents = ((float4)(maxAabb.fx - minAabb.fx,maxAabb.fy - minAabb.fy,maxAabb.fz - minAabb.fz,0.f))*0.5f;\n"
+"			float4 localCenter = ((float4)(maxAabb.fx + minAabb.fx,maxAabb.fy + minAabb.fy,maxAabb.fz + minAabb.fz,0.f))*0.5f;\n"
+"				\n"
+"			float4 worldCenter = transform(&localCenter,&position,&orientation);\n"
+"				\n"
+"			Matrix3x3 abs_b = qtGetRotationMatrix(orientation);\n"
+"			float4 extent = (float4) (	dot(abs_b.m_row[0],halfExtents),dot(abs_b.m_row[1],halfExtents),dot(abs_b.m_row[2],halfExtents),0.f);\n"
+"			\n"
+"	\n"
+"			pAABB[nodeID*2].fx = worldCenter.x-extent.x;\n"
+"			pAABB[nodeID*2].fy = worldCenter.y-extent.y;\n"
+"			pAABB[nodeID*2].fz = worldCenter.z-extent.z;\n"
+"			pAABB[nodeID*2].uw = nodeID;\n"
+"	\n"
+"			pAABB[nodeID*2+1].fx = worldCenter.x+extent.x;\n"
+"			pAABB[nodeID*2+1].fy = worldCenter.y+extent.y;\n"
+"			pAABB[nodeID*2+1].fz = worldCenter.z+extent.z;\n"
+"			pAABB[nodeID*2+1].uw = gBodies[nodeID].m_invMass==0.f? 0 : 1;\n"
+"		}\n"
+"	} \n"
+"}\n"
+"\n"
+;