dali-physics/third-party/bullet3/src/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.cpp

   1 #include "b3PrefixScanFloat4CL.h"
   2 #include "b3FillCL.h"
   3 #define B3_PREFIXSCAN_FLOAT4_PROG_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanFloat4Kernels.cl"
   4
   5 #include "b3LauncherCL.h"
   6 #include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
   7 #include "kernels/PrefixScanKernelsFloat4CL.h"
   8
   9 b3PrefixScanFloat4CL::b3PrefixScanFloat4CL(cl_context ctx, cl_device_id device, cl_command_queue queue, int size)
  10         : m_commandQueue(queue)
  11 {
  12         const char* scanKernelSource = prefixScanKernelsFloat4CL;
  13         cl_int pErrNum;
  14         char* additionalMacros = 0;
  15
  16         m_workBuffer = new b3OpenCLArray<b3Vector3>(ctx, queue, size);
  17         cl_program scanProg = b3OpenCLUtils::compileCLProgramFromString(ctx, device, scanKernelSource, &pErrNum, additionalMacros, B3_PREFIXSCAN_FLOAT4_PROG_PATH);
  18         b3Assert(scanProg);
  19
  20         m_localScanKernel = b3OpenCLUtils::compileCLKernelFromString(ctx, device, scanKernelSource, "LocalScanKernel", &pErrNum, scanProg, additionalMacros);
  21         b3Assert(m_localScanKernel);
  22         m_blockSumKernel = b3OpenCLUtils::compileCLKernelFromString(ctx, device, scanKernelSource, "TopLevelScanKernel", &pErrNum, scanProg, additionalMacros);
  23         b3Assert(m_blockSumKernel);
  24         m_propagationKernel = b3OpenCLUtils::compileCLKernelFromString(ctx, device, scanKernelSource, "AddOffsetKernel", &pErrNum, scanProg, additionalMacros);
  25         b3Assert(m_propagationKernel);
  26 }
  27
  28 b3PrefixScanFloat4CL::~b3PrefixScanFloat4CL()
  29 {
  30         delete m_workBuffer;
  31         clReleaseKernel(m_localScanKernel);
  32         clReleaseKernel(m_blockSumKernel);
  33         clReleaseKernel(m_propagationKernel);
  34 }
  35
  36 template <class T>
  37 T b3NextPowerOf2(T n)
  38 {
  39         n -= 1;
  40         for (int i = 0; i < sizeof(T) * 8; i++)
  41                 n = n | (n >> i);
  42         return n + 1;
  43 }
  44
  45 void b3PrefixScanFloat4CL::execute(b3OpenCLArray<b3Vector3>& src, b3OpenCLArray<b3Vector3>& dst, int n, b3Vector3* sum)
  46 {
  47         //      b3Assert( data->m_option == EXCLUSIVE );
  48         const unsigned int numBlocks = (const unsigned int)((n + BLOCK_SIZE * 2 - 1) / (BLOCK_SIZE * 2));
  49
  50         dst.resize(src.size());
  51         m_workBuffer->resize(src.size());
  52
  53         b3Int4 constBuffer;
  54         constBuffer.x = n;
  55         constBuffer.y = numBlocks;
  56         constBuffer.z = (int)b3NextPowerOf2(numBlocks);
  57
  58         b3OpenCLArray<b3Vector3>* srcNative = &src;
  59         b3OpenCLArray<b3Vector3>* dstNative = &dst;
  60
  61         {
  62                 b3BufferInfoCL bInfo[] = {b3BufferInfoCL(dstNative->getBufferCL()), b3BufferInfoCL(srcNative->getBufferCL()), b3BufferInfoCL(m_workBuffer->getBufferCL())};
  63
  64                 b3LauncherCL launcher(m_commandQueue, m_localScanKernel, "m_localScanKernel");
  65                 launcher.setBuffers(bInfo, sizeof(bInfo) / sizeof(b3BufferInfoCL));
  66                 launcher.setConst(constBuffer);
  67                 launcher.launch1D(numBlocks * BLOCK_SIZE, BLOCK_SIZE);
  68         }
  69
  70         {
  71                 b3BufferInfoCL bInfo[] = {b3BufferInfoCL(m_workBuffer->getBufferCL())};
  72
  73                 b3LauncherCL launcher(m_commandQueue, m_blockSumKernel, "m_blockSumKernel");
  74                 launcher.setBuffers(bInfo, sizeof(bInfo) / sizeof(b3BufferInfoCL));
  75                 launcher.setConst(constBuffer);
  76                 launcher.launch1D(BLOCK_SIZE, BLOCK_SIZE);
  77         }
  78
  79         if (numBlocks > 1)
  80         {
  81                 b3BufferInfoCL bInfo[] = {b3BufferInfoCL(dstNative->getBufferCL()), b3BufferInfoCL(m_workBuffer->getBufferCL())};
  82                 b3LauncherCL launcher(m_commandQueue, m_propagationKernel, "m_propagationKernel");
  83                 launcher.setBuffers(bInfo, sizeof(bInfo) / sizeof(b3BufferInfoCL));
  84                 launcher.setConst(constBuffer);
  85                 launcher.launch1D((numBlocks - 1) * BLOCK_SIZE, BLOCK_SIZE);
  86         }
  87
  88         if (sum)
  89         {
  90                 clFinish(m_commandQueue);
  91                 dstNative->copyToHostPointer(sum, 1, n - 1, true);
  92         }
  93 }
  94
  95 void b3PrefixScanFloat4CL::executeHost(b3AlignedObjectArray<b3Vector3>& src, b3AlignedObjectArray<b3Vector3>& dst, int n, b3Vector3* sum)
  96 {
  97         b3Vector3 s = b3MakeVector3(0, 0, 0);
  98         //if( data->m_option == EXCLUSIVE )
  99         {
 100                 for (int i = 0; i < n; i++)
 101                 {
 102                         dst[i] = s;
 103                         s += src[i];
 104                 }
 105         }
 106         /*else
 107         {
 108                 for(int i=0; i<n; i++)
 109                 {
 110                         s += hSrc[i];
 111                         hDst[i] = s;
 112                 }
 113         }
 114         */
 115
 116         if (sum)
 117         {
 118                 *sum = dst[n - 1];
 119         }
 120 }