/// <summary>
/// ctor
/// </summary>
public CudaMatrix(MsrMatrix M, string funcName, CudaEnviroment CudaEnv)
: base(M)
{
m_CudaEnv = CudaEnv;
base.PackMatrix(M);
cu.StreamCreate(out intStream, 0);
cu.StreamCreate(out extStream, 0);
disposed = false;
sparseMultiply = CudaEnv.Get_CudaMatrixKernelDP_Function(funcName);
cuaccext = CudaEnv.Get_CudaMatrixKernelDP_Function("accumulateExternal");
//int numreg;
//cu.FuncGetAttribute(out numreg, CUfunction_attribute.CU_FUNC_ATTRIBUTE_NUM_REGS, sparseMultiply);
//int version;
//cu.FuncGetAttribute(out version, CUfunction_attribute.CU_FUNC_ATTRIBUTE_BINARY_VERSION, sparseMultiply);
//System.Console.WriteLine("Number of registers: " + numreg + ", version: " + version);
LMAA();
if (extSize > 0)
{
// allocate page-locked mem
cu.MemHostAlloc(out h_ElementsToAcc, sizeof(double) * (uint)extSize, CUmem_host_alloc.CU_MEMHOSTALLOC_DEVICEMAP);
//test_ext = new double[totLen];
cu.MemHostGetDevicePointer(out d_ElementsToAcc, h_ElementsToAcc, 0);
cu.MemAlloc(out d_IndicesToAccumulate, (uint)extSize * sizeof(int));
// Copy indices for combining external and internal part to GPU as they don't change over execution
cu.MemcpyHtoD(d_IndicesToAccumulate, h_IndicesToAccumulate, (uint)extSize * sizeof(int));
}
}
/// <summary>
/// constructor which uses memory that is allocated elsewhere
/// </summary>
/// <param name="P"></param>
/// <param name="content">
/// used to initialize <see cref="h_data"/>
/// </param>
/// <param name="env"></param>
public CudaVector(IPartitioning P, double[] content, CudaEnviroment env) : base(P)
{
if (P.LocalLength > content.Length)
{
throw new ArgumentException("vector content must match local length of partition", "content");
}
h_data = content;
ConstructorCommon(env);
}
/// <summary>
/// Constructor
/// </summary>
/// <param name="M">Sparse matrix in MSR format</param>
/// <param name="CudaEnv"></param>
public CudaELLPACKmodMatrix(MsrMatrix M, CudaEnviroment CudaEnv)
: base(M, "ellMultiply", CudaEnv)
{
m_internalData = (ELLPACKmod)m_LocalMtx;
size = m_internalData.NoOfRows;
colCount = m_internalData.NoOfPackedCols;
valStride = m_internalData.MtxEntries.ColStride;
colStride = m_internalData.ColInd.ColStride;
blockcount = (int)Math.Ceiling((Decimal)size / blocksize);
}
/// <summary>
/// Constructor
/// </summary>
/// <param name="M">Sparse matrix in MSR format</param>
/// <param name="CudaEnv"></param>
public CudaELLPACKcacheMatrix(MsrMatrix M, CudaEnviroment CudaEnv)
: base(M, "mcellMultiply", CudaEnv)
{
using (new FuncTrace()) {
m_internalData = (ManualCacheELLPACK)m_LocalMtx;
size = m_internalData.NoOfRows;
colCount = m_internalData.NoOfPackedCols;
valStride = m_internalData.MtxEntries.ColStride;
colStride = m_internalData.ColIndBlock.ColStride;
blockcount = (int)Math.Ceiling((Decimal)size / blocksize);
}
}
private void ConstructorCommon(CudaEnviroment env)
{
m_env = env;
cuscale = m_env.Get_CudaVectorKernelDP_Function("scale");
cuacc = m_env.Get_CudaVectorKernelDP_Function("acc");
cudnrm2 = m_env.Get_CudaVectorKernelDP_Function("dnrm2");
cuinnerprod = m_env.Get_CudaVectorKernelDP_Function("innerprod");
cumew = m_env.Get_CudaVectorKernelDP_Function("mew");
size = this.Part.LocalLength;
blockcountfull = (int)Math.Ceiling((decimal)size / blocksize);
blockcounthalf = (int)Math.Ceiling((decimal)size / (2 * blocksize));
}
/// <summary>
///
/// </summary>
internal CudaCCBCSRMatrix(MsrMatrix M, CudaEnviroment CudaEnv)
: base(M, "blockMultiply2", CudaEnv)
{
m_internalData = (CCBCSR)m_LocalMtx;
rowcount = base.RowPartitioning.LocalLength;
cellsize = m_internalData.CellSize;
// Number of cells per block, choose so that it is around 128 threads per block
cellrowsperblock = (int)Math.Ceiling(128.0 / cellsize);
cellsperrow = m_internalData.NoOfCellsPerRow;
stride = m_internalData.CellStride;
// Number of threads per block
blocksize = cellsize * cellrowsperblock;
// Number of blocks
blockcount = (int)Math.Ceiling((Decimal)rowcount / blocksize);
}
/// <summary>
/// ctor
/// </summary>
/// <param name="cuEnv">Distribution of processes and CUDA devices</param>
public CudaDevice(CudaEnviroment cuEnv)
{
m_Env = cuEnv;
}
/// <summary>
/// constructor that allocates its own memory
/// </summary>
/// <param name="p">vector partition among MPI processes</param>
/// <param name="env"></param>
public CudaVector(IPartitioning p, CudaEnviroment env)
: base(p)
{
h_data = new double[p.LocalLength];
ConstructorCommon(env);
}
/// <summary>
/// Constructor
/// </summary>
public CudaCSRMatrix(MsrMatrix M, CudaEnviroment CudaEnv)
: base(M, "sparseMultiply", CudaEnv)
{
size = base.RowPartitioning.LocalLength;
blockcount = (int)Math.Ceiling((decimal)size / blocksize);
}
