1/*
2Copyright (c) 2012 Advanced Micro Devices, Inc.
3
4This software is provided 'as-is', without any express or implied warranty.
5In no event will the authors be held liable for any damages arising from the use of this software.
6Permission is granted to anyone to use this software for any purpose,
7including commercial applications, and to alter it and redistribute it freely,
8subject to the following restrictions:
9
101. 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.
112. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
123. This notice may not be removed or altered from any source distribution.
13*/
14//Originally written by Erwin Coumans
15
16#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
17
18#pragma OPENCL EXTENSION cl_amd_printf : enable
19#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
20#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
21#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
22#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
23
24#ifdef cl_ext_atomic_counters_32
25#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
26#else
27#define counter32_t volatile __global int*
28#endif
29
30#define SIMD_WIDTH 64
31
32typedef unsigned int u32;
33typedef unsigned short u16;
34typedef unsigned char u8;
35
36#define GET_GROUP_IDX get_group_id(0)
37#define GET_LOCAL_IDX get_local_id(0)
38#define GET_GLOBAL_IDX get_global_id(0)
39#define GET_GROUP_SIZE get_local_size(0)
40#define GET_NUM_GROUPS get_num_groups(0)
41#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
42#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
43#define AtomInc(x) atom_inc(&(x))
44#define AtomInc1(x, out) out = atom_inc(&(x))
45#define AppendInc(x, out) out = atomic_inc(x)
46#define AtomAdd(x, value) atom_add(&(x), value)
47#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
48#define AtomXhg(x, value) atom_xchg ( &(x), value )
49
50
51#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
52
53#define make_float4 (float4)
54#define make_float2 (float2)
55#define make_uint4 (uint4)
56#define make_int4 (int4)
57#define make_uint2 (uint2)
58#define make_int2 (int2)
59
60
61#define max2 max
62#define min2 min
63
64
65#define WG_SIZE 64
66
67
68
69
70
71typedef struct
72{
73	int m_n;
74	int m_start;
75	int m_staticIdx;
76	int m_paddings[1];
77} ConstBuffer;
78
79typedef struct
80{
81	int m_a;
82	int m_b;
83	u32 m_idx;
84}Elem;
85
86
87
88
89
90//	batching on the GPU
91__kernel void CreateBatchesBruteForce( __global struct b3Contact4Data* gConstraints, 	__global const u32* gN, __global const u32* gStart, int m_staticIdx )
92{
93	int wgIdx = GET_GROUP_IDX;
94	int lIdx = GET_LOCAL_IDX;
95
96	const int m_n = gN[wgIdx];
97	const int m_start = gStart[wgIdx];
98
99	if( lIdx == 0 )
100	{
101		for (int i=0;i<m_n;i++)
102		{
103			int srcIdx = i+m_start;
104			int batchIndex = i;
105			gConstraints[ srcIdx ].m_batchIdx = batchIndex;
106		}
107	}
108}
109
110
111#define CHECK_SIZE (WG_SIZE)
112
113
114
115
116u32 readBuf(__local u32* buff, int idx)
117{
118	idx = idx % (32*CHECK_SIZE);
119	int bitIdx = idx%32;
120	int bufIdx = idx/32;
121	return buff[bufIdx] & (1<<bitIdx);
122}
123
124void writeBuf(__local u32* buff, int idx)
125{
126	idx = idx % (32*CHECK_SIZE);
127	int bitIdx = idx%32;
128	int bufIdx = idx/32;
129	buff[bufIdx] |= (1<<bitIdx);
130	//atom_or( &buff[bufIdx], (1<<bitIdx) );
131}
132
133u32 tryWrite(__local u32* buff, int idx)
134{
135	idx = idx % (32*CHECK_SIZE);
136	int bitIdx = idx%32;
137	int bufIdx = idx/32;
138	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );
139	return ((ans >> bitIdx)&1) == 0;
140}
141
142
143//	batching on the GPU
144__kernel void CreateBatchesNew( __global struct b3Contact4Data* gConstraints, __global const u32* gN, __global const u32* gStart, __global int* batchSizes, int staticIdx )
145{
146	int wgIdx = GET_GROUP_IDX;
147	int lIdx = GET_LOCAL_IDX;
148	const int numConstraints = gN[wgIdx];
149	const int m_start = gStart[wgIdx];
150	b3Contact4Data_t tmp;
151
152	__local u32 ldsFixedBuffer[CHECK_SIZE];
153
154
155
156
157
158	if( lIdx == 0 )
159	{
160
161
162		__global struct b3Contact4Data* cs = &gConstraints[m_start];
163
164
165		int numValidConstraints = 0;
166		int batchIdx = 0;
167
168		while( numValidConstraints < numConstraints)
169		{
170			int nCurrentBatch = 0;
171			//	clear flag
172
173			for(int i=0; i<CHECK_SIZE; i++)
174				ldsFixedBuffer[i] = 0;
175
176			for(int i=numValidConstraints; i<numConstraints; i++)
177			{
178
179				int bodyAS = cs[i].m_bodyAPtrAndSignBit;
180				int bodyBS = cs[i].m_bodyBPtrAndSignBit;
181				int bodyA = abs(bodyAS);
182				int bodyB = abs(bodyBS);
183				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
184				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
185				int aUnavailable = aIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyA);
186				int bUnavailable = bIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyB);
187
188				if( aUnavailable==0 && bUnavailable==0 ) // ok
189				{
190					if (!aIsStatic)
191					{
192						writeBuf( ldsFixedBuffer, bodyA );
193					}
194					if (!bIsStatic)
195					{
196						writeBuf( ldsFixedBuffer, bodyB );
197					}
198
199					cs[i].m_batchIdx = batchIdx;
200
201					if (i!=numValidConstraints)
202					{
203
204						tmp = cs[i];
205						cs[i] = cs[numValidConstraints];
206						cs[numValidConstraints]  = tmp;
207
208
209					}
210
211					numValidConstraints++;
212
213					nCurrentBatch++;
214					if( nCurrentBatch == SIMD_WIDTH)
215					{
216						nCurrentBatch = 0;
217						for(int i=0; i<CHECK_SIZE; i++)
218							ldsFixedBuffer[i] = 0;
219
220					}
221				}
222			}//for
223			batchIdx ++;
224		}//while
225
226		batchSizes[wgIdx] = batchIdx;
227
228	}//if( lIdx == 0 )
229
230	//return batchIdx;
231}
232