dali-physics/third-party/bullet3/src/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.h

   1 #ifndef B3_GPU_SAP_BROADPHASE_H
   2 #define B3_GPU_SAP_BROADPHASE_H
   3
   4 #include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
   5 #include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h"  //b3Int2
   6 class b3Vector3;
   7 #include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
   8
   9 #include "b3SapAabb.h"
  10 #include "Bullet3Common/shared/b3Int2.h"
  11
  12 #include "b3GpuBroadphaseInterface.h"
  13
  14 class b3GpuSapBroadphase : public b3GpuBroadphaseInterface
  15 {
  16         cl_context m_context;
  17         cl_device_id m_device;
  18         cl_command_queue m_queue;
  19         cl_kernel m_flipFloatKernel;
  20         cl_kernel m_scatterKernel;
  21         cl_kernel m_copyAabbsKernel;
  22         cl_kernel m_sapKernel;
  23         cl_kernel m_sap2Kernel;
  24         cl_kernel m_prepareSumVarianceKernel;
  25
  26         class b3RadixSort32CL* m_sorter;
  27
  28         ///test for 3d SAP
  29         b3AlignedObjectArray<b3SortData> m_sortedAxisCPU[3][2];
  30         b3AlignedObjectArray<b3UnsignedInt2> m_objectMinMaxIndexCPU[3][2];
  31         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis0;
  32         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis1;
  33         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis2;
  34         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis0prev;
  35         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis1prev;
  36         b3OpenCLArray<b3UnsignedInt2> m_objectMinMaxIndexGPUaxis2prev;
  37
  38         b3OpenCLArray<b3SortData> m_sortedAxisGPU0;
  39         b3OpenCLArray<b3SortData> m_sortedAxisGPU1;
  40         b3OpenCLArray<b3SortData> m_sortedAxisGPU2;
  41         b3OpenCLArray<b3SortData> m_sortedAxisGPU0prev;
  42         b3OpenCLArray<b3SortData> m_sortedAxisGPU1prev;
  43         b3OpenCLArray<b3SortData> m_sortedAxisGPU2prev;
  44
  45         b3OpenCLArray<b3Int4> m_addedHostPairsGPU;
  46         b3OpenCLArray<b3Int4> m_removedHostPairsGPU;
  47         b3OpenCLArray<int> m_addedCountGPU;
  48         b3OpenCLArray<int> m_removedCountGPU;
  49
  50         int m_currentBuffer;
  51
  52 public:
  53         b3OpenCLArray<int> m_pairCount;
  54
  55         b3OpenCLArray<b3SapAabb> m_allAabbsGPU;
  56         b3AlignedObjectArray<b3SapAabb> m_allAabbsCPU;
  57
  58         virtual b3OpenCLArray<b3SapAabb>& getAllAabbsGPU()
  59         {
  60                 return m_allAabbsGPU;
  61         }
  62         virtual b3AlignedObjectArray<b3SapAabb>& getAllAabbsCPU()
  63         {
  64                 return m_allAabbsCPU;
  65         }
  66
  67         b3OpenCLArray<b3Vector3> m_sum;
  68         b3OpenCLArray<b3Vector3> m_sum2;
  69         b3OpenCLArray<b3Vector3> m_dst;
  70
  71         b3OpenCLArray<int> m_smallAabbsMappingGPU;
  72         b3AlignedObjectArray<int> m_smallAabbsMappingCPU;
  73
  74         b3OpenCLArray<int> m_largeAabbsMappingGPU;
  75         b3AlignedObjectArray<int> m_largeAabbsMappingCPU;
  76
  77         b3OpenCLArray<b3Int4> m_overlappingPairs;
  78
  79         //temporary gpu work memory
  80         b3OpenCLArray<b3SortData> m_gpuSmallSortData;
  81         b3OpenCLArray<b3SapAabb> m_gpuSmallSortedAabbs;
  82
  83         class b3PrefixScanFloat4CL* m_prefixScanFloat4;
  84
  85         enum b3GpuSapKernelType
  86         {
  87                 B3_GPU_SAP_KERNEL_BRUTE_FORCE_CPU = 1,
  88                 B3_GPU_SAP_KERNEL_BRUTE_FORCE_GPU,
  89                 B3_GPU_SAP_KERNEL_ORIGINAL,
  90                 B3_GPU_SAP_KERNEL_BARRIER,
  91                 B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY
  92         };
  93
  94         b3GpuSapBroadphase(cl_context ctx, cl_device_id device, cl_command_queue q, b3GpuSapKernelType kernelType = B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY);
  95         virtual ~b3GpuSapBroadphase();
  96
  97         static b3GpuBroadphaseInterface* CreateFuncBruteForceCpu(cl_context ctx, cl_device_id device, cl_command_queue q)
  98         {
  99                 return new b3GpuSapBroadphase(ctx, device, q, B3_GPU_SAP_KERNEL_BRUTE_FORCE_CPU);
 100         }
 101
 102         static b3GpuBroadphaseInterface* CreateFuncBruteForceGpu(cl_context ctx, cl_device_id device, cl_command_queue q)
 103         {
 104                 return new b3GpuSapBroadphase(ctx, device, q, B3_GPU_SAP_KERNEL_BRUTE_FORCE_GPU);
 105         }
 106
 107         static b3GpuBroadphaseInterface* CreateFuncOriginal(cl_context ctx, cl_device_id device, cl_command_queue q)
 108         {
 109                 return new b3GpuSapBroadphase(ctx, device, q, B3_GPU_SAP_KERNEL_ORIGINAL);
 110         }
 111         static b3GpuBroadphaseInterface* CreateFuncBarrier(cl_context ctx, cl_device_id device, cl_command_queue q)
 112         {
 113                 return new b3GpuSapBroadphase(ctx, device, q, B3_GPU_SAP_KERNEL_BARRIER);
 114         }
 115         static b3GpuBroadphaseInterface* CreateFuncLocalMemory(cl_context ctx, cl_device_id device, cl_command_queue q)
 116         {
 117                 return new b3GpuSapBroadphase(ctx, device, q, B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY);
 118         }
 119
 120         virtual void calculateOverlappingPairs(int maxPairs);
 121         virtual void calculateOverlappingPairsHost(int maxPairs);
 122
 123         void reset();
 124
 125         void init3dSap();
 126         virtual void calculateOverlappingPairsHostIncremental3Sap();
 127
 128         virtual void createProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr, int collisionFilterGroup, int collisionFilterMask);
 129         virtual void createLargeProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr, int collisionFilterGroup, int collisionFilterMask);
 130
 131         //call writeAabbsToGpu after done making all changes (createProxy etc)
 132         virtual void writeAabbsToGpu();
 133
 134         virtual cl_mem getAabbBufferWS();
 135         virtual int getNumOverlap();
 136         virtual cl_mem getOverlappingPairBuffer();
 137
 138         virtual b3OpenCLArray<b3Int4>& getOverlappingPairsGPU();
 139         virtual b3OpenCLArray<int>& getSmallAabbIndicesGPU();
 140         virtual b3OpenCLArray<int>& getLargeAabbIndicesGPU();
 141 };
 142
 143 #endif  //B3_GPU_SAP_BROADPHASE_H