public IVector Add(IVector vector)
{
Debug.Assert(IsValid && vector.IsValid);
var other = (GpuVector)vector;
Debug.Assert(other.Count == Count);
var ret = _cuda.Allocate(other._data.Size);
ret.CopyToDevice(other._data);
_cuda.Blas.Axpy(1.0f, _data.DeviceVariable, 1, ret.DeviceVariable, 1);
return(new GpuVector(_cuda, ret, true));
}
public GpuMatrix(CudaProvider cuda, int rows, int columns, Func <int, int, float> init)
{
_cuda = cuda;
_rows = rows;
_columns = columns;
var count = rows * columns;
var data = new float[count];
for (var j = 0; j < columns; j++)
{
for (var i = 0; i < rows; i++)
{
data[j * rows + i] = init(i, j);
}
}
_data = cuda.Allocate(count);
_data.CopyToDevice(data);
cuda.Register(this);
#if DEBUG
if (_id == _badAlloc)
{
Debugger.Break();
}
#endif
}
public IMatrix ConvertToMatrix()
{
Debug.Assert(IsValid);
var rows = ColumnCount * RowCount;
var columns = Depth;
var ret = _cuda.Allocate(rows * columns);
_cuda.TensorConvertToMatrix(_tensorInfo.Value.Single(), ColumnCount, RowCount, rows, columns, ret);
return(new GpuMatrix(_cuda, rows, columns, ret));
}
public GpuVector(CudaProvider cuda, int size, Func <int, float> init)
{
_cuda = cuda;
var data = new float[size];
for (var i = 0; i < size; i++)
{
data[i] = init(i);
}
_data = cuda.Allocate(size);
_data.CopyToDevice(data);
cuda.Register(this);
#if DEBUG
if (_id == _badAlloc)
{
Debugger.Break();
}
#endif
}
public I3DTensor Multiply(IMatrix matrix)
{
var other = (GpuMatrix)matrix;
var ptr = _data.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._rows == _rows && other._columns == _columns);
var ret = _cuda.Allocate(other._data.Size);
ret.CopyToDevice(other._data);
_cuda.Blas.Axpy(1.0f, _data.DeviceVariable, 1, ret.DeviceVariable, 1);
return(new GpuMatrix(_cuda, _rows, _columns, ret));
}
