/// <summary>
/// This function initializes a previously created RNN descriptor object.
/// </summary>
/// <param name="hiddenSize">Size of the internal hidden state for each layer.</param>
/// <param name="seqLength">Number of iterations to unroll over.</param>
/// <param name="numLayers">Number of layers.</param>
/// <param name="dropoutDesc">Handle to a previously created and initialized dropout descriptor.</param>
/// <param name="inputMode">Specifies the behavior at the input to the first layer.</param>
/// <param name="direction">Specifies the recurrence pattern. (eg. bidirectional)</param>
/// <param name="mode">The type of RNN to compute.</param>
/// <param name="dataType">Math precision.</param>
public void SetRNNDescriptor(
int hiddenSize,
int seqLength,
int numLayers,
DropoutDescriptor dropoutDesc, // Between layers, not between recurrent steps.
cudnnRNNInputMode inputMode,
cudnnDirectionMode direction,
cudnnRNNMode mode,
cudnnDataType dataType)
{
res = CudaDNNNativeMethods.cudnnSetRNNDescriptor(_desc, hiddenSize, seqLength, numLayers, dropoutDesc.Desc, inputMode, direction, mode, dataType);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnSetRNNDescriptor", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/// <summary>
/// This function initializes a previously created generic Tensor descriptor object.
/// </summary>
/// <param name="dataType">Data type.</param>
/// <param name="nbDims">Dimension of the tensor.</param>
/// <param name="dimA">Array of dimension nbDims that contain the size of the tensor for every dimension.</param>
/// <param name="strideA">Array of dimension nbDims that contain the stride of the tensor for every dimension.</param>
public void SetTensorNdDescriptor(cudnnDataType dataType,
int nbDims,
int[] dimA,
int[] strideA
)
{
res = CudaDNNNativeMethods.cudnnSetTensorNdDescriptor(_desc, dataType, nbDims, dimA, strideA);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnSetTensorNdDescriptor", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/// <summary>
/// This function queries the parameters of the previouly initialized Tensor4D descriptor object.
/// </summary>
/// <param name="dataType">Data type.</param>
/// <param name="n">Number of images.</param>
/// <param name="c">Number of feature maps per image.</param>
/// <param name="h">Height of each feature map.</param>
/// <param name="w">Width of each feature map.</param>
/// <param name="nStride">Stride between two consecutive images.</param>
/// <param name="cStride">Stride between two consecutive feature maps.</param>
/// <param name="hStride">Stride between two consecutive rows.</param>
/// <param name="wStride">Stride between two consecutive columns.</param>
public void GetTensor4dDescriptor(ref cudnnDataType dataType, // image data type
ref int n, // number of inputs (batch size)
ref int c, // number of input feature maps
ref int h, // height of input section
ref int w, // width of input section
ref int nStride,
ref int cStride,
ref int hStride,
ref int wStride
)
{
res = CudaDNNNativeMethods.cudnnGetTensor4dDescriptor(_desc, ref dataType, ref n, ref c, ref h, ref w, ref nStride, ref cStride, ref hStride, ref wStride);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnGetTensor4dDescriptor", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/// <summary>
/// This routine executes the recurrent neural network described by rnnDesc with inputs x, hx, cx, weights w
/// and outputs y, hy, cy. workspace is required for intermediate storage. reserveSpace stores data required
/// for training. The same reserveSpace data must be used for future calls to cudnnRNNBackwardData and
/// cudnnRNNBackwardWeights if these execute on the same input data.
/// </summary>
/// <param name="xDesc">An array of tensor descriptors describing the input to each recurrent iteration. Each
/// tensor descriptor must have the same first dimension. The second dimension of the tensors may decrease
/// from element n to element n+1 but may not increase. The tensor must be fully packed.</param>
/// <param name="x">Data pointer to GPU memory associated with the tensor descriptors in the array xDesc.</param>
/// <param name="hxDesc">Handle to a previously initialized tensor descriptor describing the initial hidden state
/// of the RNN. The first dimension of the tensor must match the hiddenSize argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second
/// dimension of the first tensor described in xDesc. The third dimension must match the numLayers argument
/// passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="hx">Data pointer to GPU memory associated with the tensor descriptor hxDesc. If a NULL pointer
/// is passed, the initial hidden state of the network will be initialized to zero.</param>
/// <param name="cxDesc">Handle to a previously initialized tensor descriptor describing the initial
/// cell state for LSTM networks. The first dimension of the tensor must match the hiddenSize argument
/// passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match
/// the second dimension of the first tensor described in xDesc. The third dimension must match the numLayers
/// argument passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully
/// packed.</param>
/// <param name="cx">Data pointer to GPU memory associated with the tensor descriptor cxDesc. If a NULL pointer is
/// passed, the initial cell state of the network will be initialized to zero.</param>
/// <param name="wDesc">Handle to a previously initialized filter descriptor describing the weights for the RNN.</param>
/// <param name="w">Data pointer to GPU memory associated with the filter descriptor wDesc.</param>
/// <param name="yDesc">An array of tensor descriptors describing the output from each recurrent iteration. The first
/// dimension of the tensor depends on the direction argument passed to the cudnnSetRNNDescriptor
/// call used to initialize rnnDesc:
/// * If direction is CUDNN_UNIDIRECTIONAL the first dimension should match the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// * If direction is CUDNN_BIDIRECTIONAL the first dimension should match double the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// The second dimension of the tensor n must match the second dimension of the tensor
/// n in xDesc. The tensor must be fully packed.</param>
/// <param name="y">Data pointer to GPU memory associated with the output tensor descriptor yDesc.</param>
/// <param name="hyDesc">Handle to a previously initialized tensor descriptor describing the final
/// hidden state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second dimension
/// of the first tensor described in xDesc. The third dimension must match the numLayers argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="hy">Data pointer to GPU memory associated with the tensor descriptor hyDesc. If a
/// NULL pointer is passed, the final hidden state of the network will not be saved.</param>
/// <param name="cyDesc">Handle to a previously initialized tensor descriptor describing the final cell state
/// for LSTM networks. The first dimension of the tensor must match the hiddenSize argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second dimension
/// of the first tensor described in xDesc. The third dimension must match the numLayers argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="cy">Data pointer to GPU memory associated with the tensor descriptor cyDesc. If a NULL pointer is
/// passed, the final cell state of the network will be not be saved.</param>
/// <param name="workspace">Data pointer to GPU memory to be used as a workspace for this call.</param>
/// <param name="workSpaceSizeInBytes">Specifies the size in bytes of the provided workspace.</param>
/// <param name="reserveSpace">Data pointer to GPU memory to be used as a reserve space for this call.</param>
/// <param name="reserveSpaceSizeInBytes">Specifies the size in bytes of the provided reserveSpace.</param>
public void RNNForwardTraining(
TensorDescriptor[] xDesc,
CudaDeviceVariable<double> x,
TensorDescriptor hxDesc,
CudaDeviceVariable<double> hx,
TensorDescriptor cxDesc,
CudaDeviceVariable<double> cx,
FilterDescriptor wDesc,
CudaDeviceVariable<double> w,
TensorDescriptor[] yDesc,
CudaDeviceVariable<double> y,
TensorDescriptor hyDesc,
CudaDeviceVariable<double> hy,
TensorDescriptor cyDesc,
CudaDeviceVariable<double> cy,
CudaDeviceVariable<byte> workspace,
SizeT workSpaceSizeInBytes,
CudaDeviceVariable<byte> reserveSpace,
SizeT reserveSpaceSizeInBytes)
{
var a1 = xDesc.Select(q => q.Desc).ToArray();
var a2 = yDesc.Select(q => q.Desc).ToArray();
res = CudaDNNNativeMethods.cudnnRNNForwardTraining(
_handle, _desc, a1, x.DevicePointer, hxDesc.Desc, hx.DevicePointer, cxDesc.Desc, cx.DevicePointer, wDesc.Desc, w.DevicePointer,
a2, y.DevicePointer, hyDesc.Desc, hy.DevicePointer, cyDesc.Desc, cy.DevicePointer, workspace.DevicePointer, workSpaceSizeInBytes, reserveSpace.DevicePointer, reserveSpaceSizeInBytes);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnRNNForwardTraining", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/// <summary>
/// This routine accumulates weight gradients dw from the recurrent neural network described
/// by rnnDesc with inputs x, hx, and outputs y. The mode of operation in this case is additive,
/// the weight gradients calculated will be added to those already existing in dw. workspace
/// is required for intermediate storage. The data in reserveSpace must have previously been
/// generated by cudnnRNNBackwardData.
/// </summary>
/// <param name="xDesc">An array of tensor descriptors describing the input to each recurrent iteration.
/// Each tensor descriptor must have the same first dimension. The second dimension of the tensors may
/// decrease from element n to element n+1 but may not increase. The tensor must be fully packed.</param>
/// <param name="x">Data pointer to GPU memory associated with the tensor descriptors in the array xDesc.</param>
/// <param name="hxDesc">Handle to a previously initialized tensor descriptor describing the initial hidden
/// state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second dimension
/// of the first tensor described in xDesc. The third dimension must match the numLayers argument passed to
/// the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed. </param>
/// <param name="hx">Data pointer to GPU memory associated with the tensor descriptor hxDesc. If
/// a NULL pointer is passed, the initial hidden state of the network will be initialized to zero.</param>
/// <param name="yDesc">An array of tensor descriptors describing the output from each
/// recurrent iteration. The first dimension of the tensor depends on the direction
/// argument passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc:
/// * If direction is CUDNN_UNIDIRECTIONAL the first dimension should match the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// * If direction is CUDNN_BIDIRECTIONAL the first dimension should match double the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// The second dimension of the tensor n must match the second dimension of the tensor n in dyDesc.
/// The tensor must be fully packed.</param>
/// <param name="y">Data pointer to GPU memory associated with the output tensor descriptor yDesc.</param>
/// <param name="workspace">Data pointer to GPU memory to be used as a workspace for this call.</param>
/// <param name="workSpaceSizeInBytes">Specifies the size in bytes of the provided workspace.</param>
/// <param name="dwDesc">Handle to a previously initialized filter descriptor describing the gradients of the weights for the RNN.</param>
/// <param name="dw">Data pointer to GPU memory associated with the filter descriptor dwDesc.</param>
/// <param name="reserveSpace">Data pointer to GPU memory to be used as a reserve space for this call.</param>
/// <param name="reserveSpaceSizeInBytes">Specifies the size in bytes of the provided reserveSpace.</param>
public void RNNBackwardWeights(
TensorDescriptor[] xDesc,
CudaDeviceVariable<float> x,
TensorDescriptor hxDesc,
CudaDeviceVariable<float> hx,
TensorDescriptor[] yDesc,
CudaDeviceVariable<float> y,
CudaDeviceVariable<byte> workspace,
SizeT workSpaceSizeInBytes,
FilterDescriptor dwDesc,
CudaDeviceVariable<float> dw,
CudaDeviceVariable<byte> reserveSpace,
SizeT reserveSpaceSizeInBytes)
{
var a1 = xDesc.Select(q => q.Desc).ToArray();
var a2 = yDesc.Select(q => q.Desc).ToArray();
res = CudaDNNNativeMethods.cudnnRNNBackwardWeights(
_handle, _desc, a1, x.DevicePointer, hxDesc.Desc, hx.DevicePointer, a2, y.DevicePointer, workspace.DevicePointer, workSpaceSizeInBytes, dwDesc.Desc, dw.DevicePointer, reserveSpace.DevicePointer, reserveSpaceSizeInBytes);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnRNNBackwardWeights", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
// LRN uses a window [center-lookBehind, center+lookAhead], where
// lookBehind = floor( (lrnN-1)/2 ), lookAhead = lrnN-lookBehind-1.
// So for n=10, the window is [k-4...k...k+5] with a total of 10 samples.
// Values of double parameters will be cast down to tensor data type.
public void SetLRNDescriptor(uint lrnN,
double lrnAlpha,
double lrnBeta,
double lrnK
)
{
res = CudaDNNNativeMethods.cudnnSetLRNDescriptor(_desc, lrnN, lrnAlpha, lrnBeta, lrnK);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnSetLRNDescriptor", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/// <summary>
/// This function is used to obtain a pointer and descriptor for the matrix parameters in layer within
/// the RNN described by rnnDesc with inputs dimensions defined by xDesc.
/// </summary>
/// <param name="layer">The layer to query.</param>
/// <param name="xDesc">An array of tensor descriptors describing the input to each recurrent iteration.</param>
/// <param name="wDesc">Handle to a previously initialized filter descriptor describing the weights for the RNN.</param>
/// <param name="w">Data pointer to GPU memory associated with the filter descriptor wDesc.</param>
/// <param name="linLayerID">
/// The linear layer to obtain information about:
/// * If mode in rnnDesc was set to CUDNN_RNN_RELU or CUDNN_RNN_TANH a value of 0 references the matrix multiplication
/// applied to the input from the previous layer, a value of 1 references the matrix multiplication applied to the recurrent input.
/// * If mode in rnnDesc was set to CUDNN_LSTM values of 0-3 reference matrix multiplications applied to the input from the
/// previous layer, value of 4-7 reference matrix multiplications applied to the recurrent input.
/// ‣ Values 0 and 4 reference the input gate.
/// ‣ Values 1 and 5 reference the forget gate.
/// ‣ Values 2 and 6 reference the new memory gate.
/// ‣ Values 3 and 7 reference the output gate.
/// * If mode in rnnDesc was set to CUDNN_GRU values of 0-2 reference matrix multiplications applied to the input
/// from the previous layer, value of 3-5 reference matrix multiplications applied to the recurrent input.
/// ‣ Values 0 and 3 reference the reset gate.
/// ‣ Values 1 and 4 reference the update gate.
/// ‣ Values 2 and 5 reference the new memory gate.
/// </param>
/// <param name="linLayerMatDesc">Handle to a previously created filter descriptor.</param>
/// <param name="linLayerMat">Data pointer to GPU memory associated with the filter descriptor linLayerMatDesc.</param>
public void GetRNNLinLayerMatrixParams(
int layer,
TensorDescriptor[] xDesc,
FilterDescriptor wDesc,
CudaDeviceVariable<double> w,
int linLayerID,
FilterDescriptor linLayerMatDesc,
CudaDeviceVariable<SizeT> linLayerMat // void **
)
{
var a1 = xDesc.Select(x => x.Desc).ToArray();
res = CudaDNNNativeMethods.cudnnGetRNNLinLayerMatrixParams(_handle, _desc, layer, a1, wDesc.Desc, w.DevicePointer, linLayerID, linLayerMatDesc.Desc, linLayerMat.DevicePointer);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnGetRNNLinLayerMatrixParams", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
/* Helper function to return the dimensions of the output tensor given a convolution descriptor */
public void GetConvolution2dForwardOutputDim(TensorDescriptor inputTensorDesc,
FilterDescriptor filterDesc,
ref int n,
ref int c,
ref int h,
ref int w
)
{
res = CudaDNNNativeMethods.cudnnGetConvolution2dForwardOutputDim(_desc, inputTensorDesc.Desc, filterDesc.Desc, ref n, ref c, ref h, ref w);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnGetConvolution2dForwardOutputDim", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
//.........这里部分代码省略.........
/// <param name="dyDesc">An array of tensor descriptors describing the gradient at the output from each
/// recurrent iteration. The first dimension of the tensor depends on the direction argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc:
/// * If direction is CUDNN_UNIDIRECTIONAL the first dimension should match the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// * If direction is CUDNN_BIDIRECTIONAL the first dimension should match double the hiddenSize
/// argument passed to cudnnSetRNNDescriptor.
/// The second dimension of the tensor n must match the second dimension of the tensor n in dxDesc. The
/// tensor must be fully packed.</param>
/// <param name="dy">Data pointer to GPU memory associated with the tensor descriptors in the array dyDesc.</param>
/// <param name="dhyDesc">Handle to a previously initialized tensor descriptor describing the gradients at the
/// final hidden state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed
/// to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second
/// dimension of the first tensor described in dyDesc. The third dimension must match the numLayers argument
/// passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="dhy">Data pointer to GPU memory associated with the tensor descriptor dhyDesc. If a NULL pointer
/// is passed, the gradients at the final hidden state of the network will be initialized to zero.</param>
/// <param name="dcyDesc">Handle to a previously initialized tensor descriptor describing the gradients at
/// the final cell state of the RNN. The first dimension of the tensor must match the hiddenSize argument
/// passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the
/// second dimension of the first tensor described in dyDesc. The third dimension must match the numLayers argument
/// passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="dcy">Data pointer to GPU memory associated with the tensor descriptor dcyDesc. If a NULL pointer
/// is passed, the gradients at the final cell state of the network will be initialized to zero.</param>
/// <param name="wDesc">Handle to a previously initialized filter descriptor describing the weights for the RNN.</param>
/// <param name="w">Data pointer to GPU memory associated with the filter descriptor wDesc.</param>
/// <param name="hxDesc">Handle to a previously initialized tensor descriptor describing the initial hidden
/// state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second
/// dimension of the first tensor described in xDesc. The third dimension must match the numLayers
/// argument passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be
/// fully packed.</param>
/// <param name="hx">Data pointer to GPU memory associated with the tensor descriptor hxDesc. If a NULL pointer is
/// passed, the initial hidden state of the network will be initialized to zero.</param>
/// <param name="cxDesc">Handle to a previously initialized tensor descriptor describing the
/// initial cell state for LSTM networks. The first dimension of the tensor must match the
/// hiddenSize argument passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The
/// second dimension must match the second dimension of the first tensor described in xDesc. The
/// third dimension must match the numLayers argument passed to the cudnnSetRNNDescriptor call
/// used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="cx">Data pointer to GPU memory associated with the tensor descriptor cxDesc.
/// If a NULL pointer is passed, the initial cell state of the network will be initialized to zero.</param>
/// <param name="dxDesc">An array of tensor descriptors describing the gradient at the input of each recurrent iteration.
/// Each tensor descriptor must have the same first dimension. The second dimension of the tensors may decrease from
/// element n to element n+1 but may not increase. The tensor must be fully packed.</param>
/// <param name="dx">Data pointer to GPU memory associated with the tensor descriptors in the array dxDesc. </param>
/// <param name="dhxDesc">Handle to a previously initialized tensor descriptor describing the gradient at the initial hidden
/// state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed to the cudnnSetRNNDescriptor
/// call used to initialize rnnDesc. The second dimension must match the second dimension of the first tensor described in xDesc.
/// The third dimension must match the numLayers argument passed to the cudnnSetRNNDescriptor call used to initialize rnnDesc.
/// The tensor must be fully packed.</param>
/// <param name="dhx">Data pointer to GPU memory associated with the tensor descriptor dhxDesc. If a NULL pointer is passed, the
/// gradient at the hidden input of the network will not be set.</param>
/// <param name="dcxDesc">Handle to a previously initialized tensor descriptor describing the gradient
/// at the initial cell state of the RNN. The first dimension of the tensor must match the hiddenSize argument passed
/// to the cudnnSetRNNDescriptor call used to initialize rnnDesc. The second dimension must match the second dimension
/// of the first tensor described in xDesc. The third dimension must match the numLayers argument passed to the
/// cudnnSetRNNDescriptor call used to initialize rnnDesc. The tensor must be fully packed.</param>
/// <param name="dcx">Data pointer to GPU memory associated with the tensor descriptor dcxDesc. If
/// a NULL pointer is passed, the gradient at the cell input of the network will not be set.</param>
/// <param name="workspace">Data pointer to GPU memory to be used as a workspace for this call.</param>
/// <param name="workSpaceSizeInBytes">Specifies the size in bytes of the provided workspace.</param>
/// <param name="reserveSpace">Data pointer to GPU memory to be used as a reserve space for this call.</param>
/// <param name="reserveSpaceSizeInBytes">Specifies the size in bytes of the provided reserveSpace.</param>
public void RNNBackwardData(
TensorDescriptor[] yDesc,
CudaDeviceVariable<float> y,
TensorDescriptor[] dyDesc,
CudaDeviceVariable<float> dy,
TensorDescriptor dhyDesc,
CudaDeviceVariable<float> dhy,
TensorDescriptor dcyDesc,
CudaDeviceVariable<float> dcy,
FilterDescriptor wDesc,
CudaDeviceVariable<float> w,
TensorDescriptor hxDesc,
CudaDeviceVariable<float> hx,
TensorDescriptor cxDesc,
CudaDeviceVariable<float> cx,
TensorDescriptor[] dxDesc,
CudaDeviceVariable<float> dx,
TensorDescriptor dhxDesc,
CudaDeviceVariable<float> dhx,
TensorDescriptor dcxDesc,
CudaDeviceVariable<float> dcx,
CudaDeviceVariable<byte> workspace,
SizeT workSpaceSizeInBytes,
CudaDeviceVariable<byte> reserveSpace,
SizeT reserveSpaceSizeInBytes)
{
var a1 = yDesc.Select(q => q.Desc).ToArray();
var a2 = dyDesc.Select(q => q.Desc).ToArray();
var a3 = dxDesc.Select(q => q.Desc).ToArray();
res = CudaDNNNativeMethods.cudnnRNNBackwardData(
_handle, _desc, a1, y.DevicePointer, a2, dy.DevicePointer, dhyDesc.Desc, dhy.DevicePointer, dcyDesc.Desc, dcy.DevicePointer, wDesc.Desc, w.DevicePointer,
hxDesc.Desc, hx.DevicePointer, cxDesc.Desc, cx.DevicePointer, a3, dx.DevicePointer, dhxDesc.Desc, dhx.DevicePointer, dcxDesc.Desc, dcx.DevicePointer,
workspace.DevicePointer, workSpaceSizeInBytes, reserveSpace.DevicePointer, reserveSpaceSizeInBytes);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnRNNBackwardData", res));
if (res != cudnnStatus.Success) throw new CudaDNNException(res);
}
客服电话
APP下载
官方微信
请发表评论