/// <summary>
/// see <see cref="MatrixBase.AssembleFinalFormat"/>
/// </summary>
protected override MatrixBase.FormatBase AssembleFinalFormat(MatrixBase.TempCSR tmp)
{
using (new FuncTrace()) {
ManualCacheELLPACK ret = null;
bool TooBig = true;
int maxSharesSize;
cu.DeviceGetAttribute(out maxSharesSize, CUdevice_attribute.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK, m_CudaEnv.CUDAdev);
while (TooBig)
{
ret = new ManualCacheELLPACK(tmp, 1, blocksize, 32);
int ReqSharedSize = ret.BlockSubVectorLength.Max() * sizeof(double);
if (ReqSharedSize > maxSharesSize)
{
blocksize /= 2;
if (blocksize < 16)
{
throw new ApplicationException("unable to create ELLPACKcache - matrix; not enough shared memory available;");
}
}
else
{
TooBig = false;
}
}
return(ret);
}
}
/// <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);
}
}
public clELLPACKcacheMatrix(MsrMatrix M, clDevice device)
: base(M, device, "mcellMultiply")
{
m_internalData = (ManualCacheELLPACK)m_LocalMtx;
size = m_internalData.NoOfRows;
colCount = m_internalData.NoOfPackedCols;
valStride = m_internalData.MtxEntries.ColStride;
colStride = m_internalData.ColIndBlock.ColStride;
localsize = 256;
globalsize = size;
int m = size % localsize;
if (m > 0)
{
globalsize += localsize - m;
}
}
/// <summary>
/// the poor Man's test unit
/// </summary>
void TestFormats(MatrixBase.TempCSR _tmp)
{
MatrixBase.CSR comp1 = new CSR(_tmp);
//MatrixBase.CCBCSR comp2 = new CCBCSR(_tmp, 128, CellSize);
//MatrixBase.ELLPACKmod comp2 = new ELLPACKmod(_tmp, 40, 4);
MatrixBase.ManualCacheELLPACK comp2 = new ManualCacheELLPACK(_tmp, 1, 4, 2);
double[] x = new double[this.ColPartition.LocalLength];
double[] y = new double[this.ColPartition.LocalLength];
double alpha = 1.234;
double beta = 432.1;
Random r = new Random(0);
for (int i = 0; i < x.Length; i++)
{
x[i] = r.NextDouble();
}
for (int i = 0; i < y.Length; i++)
{
y[i] = r.NextDouble();
}
double[] res1 = (double[])y.Clone();
comp1.RefSpMv(alpha, (double[])x.Clone(), beta, res1);
double[] res2 = (double[])y.Clone();
comp2.RefSpMv(alpha, (double[])x.Clone(), beta, res2);
double err = 0;
for (int i = 0; i < y.Length; i++)
{
err += Math.Abs(res1[i] - res2[i]);
}
Console.WriteLine("err = " + err);
Console.Write("");
}
