oclQuantize::oclQuantize(oclContext& iContext, oclProgram* iParent)
: oclProgram(iContext, "oclQuantize", iParent)
// kernels
, clQuantizeLAB(*this, "clQuantizeLAB")
{
addSourceFile("color/oclQuantize.cl");
}
oclColor::oclColor(oclContext& iContext, oclProgram* iParent)
: oclProgram(iContext, "oclColor", iParent)
// kernels
, clHSVtoRGB(*this, "clHSVtoRGB")
, clRGBtoHSV(*this, "clRGBtoHSV")
, clRGBtoXYZ(*this, "clRGBtoXYZ")
, clXYZtoRGB(*this, "clXYZtoRGB")
, clRGBtoLAB(*this, "clRGBtoLAB")
, clLABtoRGB(*this, "clLABtoRGB")
{
addSourceFile("color/oclColor.cl");
}
oclConvolute::oclConvolute(oclContext& iContext)
: oclProgram(iContext, "oclConvolute")
// kernels
, clIso2D(*this)
, clIso2Dsep(*this)
, clAniso2Dtang(*this)
, clAniso2Dorth(*this)
{
exportKernel(clIso2D);
exportKernel(clIso2Dsep);
exportKernel(clAniso2Dtang);
exportKernel(clAniso2Dorth);
addSourceFile("filter/oclConvolute.cl");
}
oclBloom::oclBloom(oclContext& iContext, oclProgram* iParent)
: oclProgram(iContext, "oclBloom", iParent)
, mGaussian(iContext, this)
// buffers
, bfTempA(iContext, "bfTemp0")
, bfTempB(iContext, "bfTemp1")
// kernels
, clFilter(*this, "clFilter")
, clCombine(*this, "clCombine")
{
bfTempA.create(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sDefaultFormat, 256, 256);
bfTempB.create(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sDefaultFormat, 256, 256);
addSourceFile("image/oclBloom.cl");
}
oclBilinearPyramid::oclBilinearPyramid(oclContext& iContext)
: oclProgram(iContext, "oclBilinearPyramid")
// kernels
, clUpsample(*this)
, clDownsample(*this)
, mLevel(1)
{
addSourceFile("filter/oclBilinearPyramid.cl");
exportKernel(clUpsample);
exportKernel(clDownsample);
cl_image_format lFormat = { CL_RGBA, CL_HALF_FLOAT };
mLevel[0] = new oclImage2D(mContext, "Level");
mLevel[0]->create(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, lFormat, 256, 256);
}
oclToneMapping::oclToneMapping(oclContext& iContext, oclProgram* iParent)
: oclProgram(iContext, "oclToneMapping", iParent)
// buffers
, bfTempA(iContext, "bfTempA")
, bfTempB(iContext, "bfTempB")
// kernels
, clCombine(*this, "clCombine")
// programs
, mColor(iContext, this)
, mPyramid(iContext, this)
, mMemory(iContext, this)
{
cl_image_format lFormat = { CL_RGBA, CL_HALF_FLOAT };
bfTempA.create(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, lFormat, 256, 256);
bfTempB.create(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, lFormat, 256, 256);
addSourceFile("image/oclToneMapping.cl");
}
oclBvhTrimesh::oclBvhTrimesh(oclContext& iContext)
: oclProgram(iContext, "oclBvhTrimesh")
// buffers
, bfAABB(iContext, "bfAABB")
, bfMortonKey(iContext, "bfMortonKey")
, bfMortonVal(iContext, "bfMortonVal")
, bfBvhRoot(iContext, "bfBvhRoot")
, bfBvhNode(iContext, "bfBvhNode")
// kernels
, clAABB(*this)
, clMorton(*this)
, clCreateNodes(*this)
, clLinkNodes(*this)
, clCreateLeaves(*this)
, clComputeAABBs(*this)
// programs
, mRadixSort(iContext)
// members
, mRootNode(0)
{
bfAABB.create<srtAABB>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfMortonKey.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 256);
bfMortonVal.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 256);
bfBvhNode.create<cl_char>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfBvhRoot.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR, 1,
&mRootNode);
addSourceFile("geom/oclBvhTrimesh.cl");
exportKernel(clAABB);
exportKernel(clMorton);
exportKernel(clCreateNodes);
exportKernel(clLinkNodes);
exportKernel(clCreateLeaves);
exportKernel(clComputeAABBs);
}
oclFluid3D::oclFluid3D(oclContext& iContext)
: oclProgram(iContext, "oclFluid3D")
// buffers
, bfCell(iContext, "bfCell")
, bfCellStart(iContext, "bfCellStart")
, bfCellEnd(iContext, "bfCellEnd")
, bfIndex(iContext, "bfIndex")
, bfSortedPosition(iContext, "bfSortedPosition")
, bfSortedVelocity(iContext, "bfSortedVelocity")
, bfParams(iContext, "bfParams")
, bfPosition(0)
, bfVelocity(0)
, bfForce(0)
// kernels
, clIntegrateForce(*this)
, clIntegrateVelocity(*this)
, clHash(*this)
, clReorder(*this)
, clInitBounds(*this)
, clFindBounds(*this)
, clCalculateDensity(*this)
, clCalculateForces(*this)
, clInitFluid(*this)
, clClipBox(*this)
, clGravity(*this)
// programs
, mRadixSort(iContext)
// members
, mParticleCount(cLocalSize)
, mIntegrateCb(0)
{
bfCell.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfCellStart.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, cBucketCount);
bfCellEnd.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, cBucketCount);
bfIndex.create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfSortedPosition.create<cl_float4>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfSortedVelocity.create<cl_float4>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfParams.create<Params>(CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR, 1, &mParams);
/*
oclBuffer* ll = new oclBuffer(iContext, "dsdsdsdsd");
for (int i=0; i<100; i++)
{
ll->create<cl_uint>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, cBucketCount);
delete ll;
ll = new oclBuffer(iContext, "dsdsdsdsd");
}
*/
bfPosition = new oclBuffer(iContext, "bfPosition");
bfPosition->setOwner(this);
bfVelocity = new oclBuffer(iContext, "bfVelocity");
bfVelocity->setOwner(this);
bfForce = new oclBuffer(iContext, "bfForce");
bfForce->setOwner(this);
bfPosition->create<cl_float4>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfVelocity->create<cl_float4>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
bfForce->create<cl_float4>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mParticleCount);
addSourceFile("phys/oclFluid3D.cl");
exportKernel(clClipBox);
exportKernel(clGravity);
exportKernel(clIntegrateForce);
exportKernel(clIntegrateVelocity);
}
oclRbm::oclRbm(oclContext& iContext, oclProgram* iParent)
: oclProgram(iContext, "oclRbm", iParent)
// buffers
, bfVis0(iContext, "bfVis0")
, bfVisN(iContext, "bfVisN")
, bfError(iContext, "bfError", oclBuffer::_float)
, bfC(iContext, "bfC")
, bfWtemp(iContext, "bfWtemp")
, bfTempA(iContext, "bfTempA", oclBuffer::_float)
, bfTTT(iContext, "bfTTT", oclBuffer::_float)
, bfMaps(0)
, bfBernoulli(iContext, "bfBernoulli", oclBuffer::_float)
// kernels
, clGetMap(*this, "clGetMap")
, clGetVis(*this, "clGetVis")
, clGetWeight(*this, "clGetWeight")
, clLoadImage(*this, "clLoadImage")
, clGetImage(*this, "clGetImage")
, clGibbsUpA(*this, "clGibbsUpA")
, clGibbsUpB(*this, "clGibbsUpB")
, clGibbsDnA(*this, "clGibbsDnA")
, clGibbsDnB0(*this, "clGibbsDnB0")
, clGibbsDnB1(*this, "clGibbsDnB1")
, clGibbsDnC(*this, "clGibbsDnC")
, clError(*this, "clError")
, clDwA(*this, "clDwA")
, clDwB(*this, "clDwB")
, clDc(*this, "clDc")
, clDb(*this, "clDb")
, clLearn(*this, "clLearn")
, clSparsity(*this, "clSparsity")
// programs
, mMemory(iContext, this)
, mKernelW(5)
, mKernelH(5)
, mInputW(0)
, mInputH(0)
, mRate(0.01f)
, mMomentum(0.0f)
, mPenalty(0.05f)
, mSparsity(0.02f)
{
//cl_image_format lFormat = { CL_RGBA, CL_HALF_FLOAT };
bfVis0.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 256 * 256);
bfVisN.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 256 * 256);
bfBernoulli.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 256 * 256);
bfError.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfTTT.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfWtemp.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfC.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfDc[0] = new oclBuffer(iContext, "bfDc0");
bfDc[0]->create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfDc[1] = new oclBuffer(iContext, "bfDc1");
bfDc[1]->create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, 1);
bfTempA.create<cl_float>(CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, mKernelW*mKernelH);
addSourceFile("learning/oclRbm.cl");
}
static void doSetCompilerPath(const char * path, const char * includes, const char * libs, const char * tmpdir, unsigned targetCompiler, bool verbose)
{
if (!includes)
includes = INCLUDEPATH[targetCompiler];
if (!libs)
libs = LIB_DIR[targetCompiler];
if (verbose)
{
PrintLog("Include directory set to %s", includes);
PrintLog("Library directory set to %s", libs);
}
compilerRoot.set(path ? path : targetCompiler==GccCppCompiler ? "/usr" : ".\\CL");
stdIncludes.set(includes);
stdLibs.clear();
for (;;)
{
StringBuffer thislib;
while (*libs && *libs != ';')
thislib.append(*libs++);
if (thislib.length())
{
stdLibs.append(" ").append(USE_LIBPATH_FLAG[targetCompiler]).append(thislib).append(USE_LIBPATH_TAIL[targetCompiler]);
if (USE_LIBRPATH_FLAG[targetCompiler])
stdLibs.append(" ").append(USE_LIBRPATH_FLAG[targetCompiler]).append(thislib);
}
if (!*libs)
break;
libs++;
}
StringBuffer fname;
if (path)
{
const char *finger = CC_NAME[targetCompiler];
while (*finger)
{
if (*finger == '#')
fname.append(path);
else
fname.append(*finger);
finger++;
}
#if defined(__linux__)
StringBuffer clbin_dir;
const char* dir_end = strrchr(fname, '/');
if(dir_end == NULL)
clbin_dir.append(".");
else
clbin_dir.append((dir_end - fname.str()) + 1, fname.str());
StringBuffer pathenv(clbin_dir.str());
const char* oldpath = getenv("PATH");
if(oldpath != NULL && *oldpath != '\0')
pathenv.append(":").append(oldpath);
setenv("PATH", pathenv.str(), 1);
#endif
}
else
{
fname.append(compilerRoot).append(CC_NAME[targetCompiler]);
fname.replaceString("#",NULL);
}
if (verbose)
PrintLog("Compiler path set to %s", fname.str());
dequote(fname);
#ifdef _WIN32
if (_access(fname.str(), 4))
{
#else
#if defined(__linux__) || defined(__FreeBSD__) || defined(__APPLE__)
struct stat filestatus;
int r = stat(fname.str(), &filestatus);
if ( (r != 0)
|| (!S_ISREG(filestatus.st_mode))
|| (filestatus.st_mode&(S_IXOTH|S_IXGRP|S_IXUSR)==0))
{
if (r == -1) errno = ENOENT;
#endif
#endif
if (verbose)
PrintLog("SetCompilerPath - no compiler found");
throw MakeOsException(GetLastError(), "setCompilerPath could not locate compiler %s", fname.str());
}
if(tmpdir && *tmpdir)
{
//MORE: this should be done for the child process instead of the parent but invoke does not let me do it
#if defined(__linux__)
setenv("TMPDIR", tmpdir, 1);
#endif
#ifdef _WIN32
StringBuffer tmpbuf;
tmpbuf.append("TMP=").append(tmpdir);
_putenv(tmpbuf.str());
#endif
}
}
//===========================================================================
class CCompilerThreadParam : CInterface
{
public:
IMPLEMENT_IINTERFACE;
CCompilerThreadParam(const StringBuffer & _cmdline, Semaphore & _finishedCompiling, const StringBuffer & _logfile) : cmdline(_cmdline), finishedCompiling(_finishedCompiling), logfile(_logfile) {};
StringBuffer cmdline;
StringBuffer logfile;
Semaphore & finishedCompiling;
};
//===========================================================================
static void setDirectoryPrefix(StringAttr & target, const char * source)
{
if (source && *source)
{
StringBuffer temp;
target.set(addDirectoryPrefix(temp, source));
}
}
CppCompiler::CppCompiler(const char * _coreName, const char * _sourceDir, const char * _targetDir, unsigned _targetCompiler, bool _verbose)
{
#define CORE_NAME allSources.item(0)
if (_coreName)
addSourceFile(_coreName);
targetCompiler = _targetCompiler;
createDLL = true;
#ifdef _DEBUG
setDebug(true);
setDebugLibrary(true);
#else
setDebug(false);
setDebugLibrary(false);
#endif
setDirectoryPrefix(sourceDir, _sourceDir);
setDirectoryPrefix(targetDir, _targetDir);
maxCompileThreads = 1;
onlyCompile = false;
verbose = _verbose;
saveTemps = false;
abortChecker = NULL;
}
void CppCompiler::addCompileOption(const char * option)
{
compilerOptions.append(' ').append(option);
}
void QMakeSourceFileInfo::addSourceFiles(const QStringList &l, uchar seek,
QMakeSourceFileInfo::SourceFileType type)
{
for(int i=0; i<l.size(); ++i)
addSourceFile(l.at(i), seek, type);
}