public (I3DTensor Result, I3DTensor Indices) MaxPool(int filterWidth, int filterHeight,
int xStride, int yStride, bool saveIndices)
{
Debug.Assert(IsValid);
var maxPool = _cuda.TensorMaxPool(Memory, _rows, _columns, _depth, 1, filterWidth,
filterHeight, xStride, yStride, saveIndices);
var ret = new Gpu3DTensor(_cuda, maxPool.Rows, maxPool.Columns, _depth, maxPool.Data, true);
var indices = saveIndices
? new Gpu3DTensor(_cuda, maxPool.Rows, maxPool.Columns, _depth, maxPool.Indices, true)
: null;
return(ret, indices);
}
public I3DTensor TransposeThisAndMultiply(I4DTensor tensor)
{
var other = (Gpu4DTensor)tensor;
#if DEBUG
Debug.Assert(tensor.Count == Depth && IsValid && other.IsValid);
#endif
var ptr = Memory.DevicePointer;
var ptr2 = other.Memory.DevicePointer;
int rowsA = _rows, columnsA = _columns, columnsB = other.Depth,
rowsB = other.RowCount * other.ColumnCount, blockSize2 = columnsB * rowsB;
float alpha = 1.0f, beta = 0.0f;
var output = new Gpu3DTensor(_cuda, _columns, columnsB, _depth,
_cuda.Allocate(_columns * columnsB * _depth), true);
var status = CudaBlasNativeMethods.cublasSgemmStridedBatched(_cuda.Blas.CublasHandle,
Operation.Transpose, Operation.NonTranspose, columnsA, columnsB, rowsB, ref alpha, ptr,
rowsA, _blockSize, ptr2, rowsB, blockSize2, ref beta, output.Memory.DevicePointer, columnsA,
_columns * columnsB, _depth);
if (status != CublasStatus.Success)
{
throw new CudaBlasException(status);
}
return(output);
//var output = Enumerable.Range(0, _depth).Select(i => new GpuMatrix(_cuda, _columns, columnsB, _cuda.Allocate(_columns * 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 => ptr2 + i * blockSize2 * CudaProvider.FLOAT_SIZE).ToArray()))
//using (var cPtrs = new PtrToDeviceMemoryList(output.Select(m => m.Memory.DevicePointer).ToArray())) {
// var status = CudaBlasNativeMethods.cublasSgemmBatched(_cuda.Blas.CublasHandle,
// Operation.Transpose,
// Operation.NonTranspose,
// columnsA,
// columnsB,
// rowsB,
// ref alpha,
// aPtrs.DevicePointer,
// rowsA,
// bPtrs.DevicePointer,
// rowsB,
// ref beta,
// cPtrs.DevicePointer,
// columnsA,
// _depth
// );
// if (status != CublasStatus.Success)
// throw new CudaBlasException(status);
//}
//return _cuda.Create3DTensor(output);
}
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);
}