//[/CudnnMnistFCF]
//[CudnnMnistCF]
public void ConvoluteForward(Layer conv, nchw_t nchw, DeviceMemory <float> srcData, ref DeviceMemory <float> dstData)
{
_srcTensorDesc.Set4D(TensorFormat, DataType, nchw.N, nchw.C, nchw.H, nchw.W);
_filterDesc.Set4D(DataType, conv.Outputs, conv.Inputs, conv.KernelDim, conv.KernelDim);
_convDesc.Set2D(0, 0, 1, 1, 1, 1, CUDNNInterop.cudnnConvolutionMode_t.CUDNN_CROSS_CORRELATION);
// find dimension of convoltion output
// outputDim = 1 + (inputDim + 2*pad - filterDim) / convolutionStride
int n, c, h, w;
_convDesc.Get2DForwardOutputDim(_srcTensorDesc, _filterDesc, out n, out c, out h, out w);
nchw.N = n;
nchw.C = c;
nchw.H = h;
nchw.W = w;
_dstTensorDesc.Set4D(TensorFormat, DataType, nchw.N, nchw.C, nchw.H, nchw.W);
var algo = _cudnn.GetConvolutionForwardAlgorithm(_srcTensorDesc, _filterDesc, _convDesc, _dstTensorDesc,
CUDNNInterop.cudnnConvolutionFwdPreference_t.CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, (IntPtr)0);
Resize(ref dstData, nchw.N * nchw.C * nchw.H * nchw.W);
var sizeInBytes = _cudnn.GetConvolutionForwardWorkspaceSize(_srcTensorDesc, _filterDesc, _convDesc, _dstTensorDesc, algo);
using (var workSpace = _worker.Malloc <byte>(sizeInBytes.ToInt32()))
{
const float alpha = 1.0f;
const float beta = 0.0f;
_cudnn.ConvolutionForward(alpha, _srcTensorDesc, srcData.Ptr, _filterDesc, conv.DataD.Ptr, _convDesc, algo, workSpace.Ptr, sizeInBytes, beta, _dstTensorDesc, dstData.Ptr);
AddBias(_dstTensorDesc, conv, c, dstData);
}
}
