public IVector Add(IVector vector)
{
Debug.Assert(IsValid && vector.IsValid);
var other = (GpuVector)vector;
Debug.Assert(other.Count == Count);
var ret = _cuda.Allocate(other.Memory.Size);
ret.CopyToDevice(other.Memory);
_cuda.Blas.Axpy(1.0f, Memory.DeviceVariable, 1, ret.DeviceVariable, 1);
return(new GpuVector(_cuda, ret, true));
}
public I3DTensor Multiply(IMatrix matrix)
{
var other = (GpuMatrix)matrix;
var ptr = Memory.DevicePointer;
int rowsA = _rows, columnsArowsB = _columns, columnsB = matrix.ColumnCount;
float alpha = 1.0f, beta = 0.0f;
var output = new Gpu3DTensor(_cuda, _rows, columnsB, _depth,
_cuda.Allocate(_rows * columnsB * _depth), true);
var status = CudaBlasNativeMethods.cublasSgemmStridedBatched(_cuda.Blas.CublasHandle,
Operation.NonTranspose, Operation.NonTranspose, rowsA, columnsB, columnsArowsB, ref alpha,
ptr, rowsA, _blockSize, other.Memory.DevicePointer, columnsArowsB, 0, ref beta,
output.Memory.DevicePointer, rowsA, _rows * columnsB, _depth);
if (status != CublasStatus.Success)
{
throw new CudaBlasException(status);
}
return(output);
//var output = Enumerable.Range(0, _depth).Select(i => new GpuMatrix(_cuda, _rows, columnsB, _cuda.Allocate(_rows * columnsB), true)).ToList();
//using (var aPtrs = new PtrToDeviceMemoryList(Enumerable.Range(0, _depth).Select(i => ptr + i * _blockSize * CudaProvider.FLOAT_SIZE).ToArray()))
//using (var bPtrs = new PtrToDeviceMemoryList(Enumerable.Range(0, _depth).Select(i => other.Memory.DevicePointer).ToArray()))
//using (var cPtrs = new PtrToDeviceMemoryList(output.Select(m => m.Memory.DevicePointer).ToArray())) {
// var status = CudaBlasNativeMethods.cublasSgemmBatched(_cuda.Blas.CublasHandle,
// Operation.NonTranspose,
// Operation.NonTranspose,
// rowsA,
// columnsB,
// columnsArowsB,
// ref alpha,
// aPtrs.DevicePointer,
// rowsA,
// bPtrs.DevicePointer,
// columnsArowsB,
// ref beta,
// cPtrs.DevicePointer,
// rowsA,
// _depth
// );
// if (status != CublasStatus.Success)
// throw new CudaBlasException(status);
//}
//return _cuda.Create3DTensor(output);
}
public IMatrix Add(IMatrix matrix)
{
Debug.Assert(IsValid && matrix.IsValid);
var other = (GpuMatrix)matrix;
Debug.Assert(other.RowCount == RowCount && other.ColumnCount == ColumnCount);
var ret = _cuda.Allocate(other.Memory.Size);
ret.CopyToDevice(other.Memory);
_cuda.Blas.Axpy(1.0f, Memory.DeviceVariable, 1, ret.DeviceVariable, 1);
return(new GpuMatrix(_cuda, RowCount, ColumnCount, ret, true));
}