public static void ConvolutionBackwardData(DNNConvolutionBwdDataAlgo algo, Cpu.ConvolutionDesc2d cd, CudaStorage workspace, Tensor w, Tensor dy, Tensor dx)
{
using (var dnn = CudaHelpers.TSContextForTensor(w).DNNForTensor(w))
{
var convDesc = GetConvDescriptor(cd, w.ElementType);
using (var workspacePtr = new CudaDeviceVariable <byte>(workspace.DevicePtrAtElement(0), false, workspace.ByteLength))
using (var wPtr = GetDeviceVar(w))
using (var dxPtr = GetDeviceVar(dx))
using (var dyPtr = GetDeviceVar(dy))
using (var wDesc = GetFilterDescriptor(w))
using (var dxDesc = GetDescriptor(dx))
using (var dyDesc = GetDescriptor(dy))
{
dnn.Value.ConvolutionBackwardData(1,
wDesc, wPtr,
dyDesc, dyPtr,
convDesc,
0,
(cudnnConvolutionBwdDataAlgo)algo,
workspacePtr,
dxDesc, dxPtr);
}
}
}
public static long GetConvolutionBackwardDataWorkspaceSize(IAllocator allocator, DNNConvolutionBwdDataAlgo algo, Cpu.ConvolutionDesc2d cd, TensorShape w, TensorShape dy, TensorShape dx)
{
if (!(allocator is CudaAllocator))
{
throw new InvalidOperationException("allocator must be a CUDA allocator");
}
var cudaAllocator = (CudaAllocator)allocator;
using (var dnn = cudaAllocator.Context.DNNForDevice(cudaAllocator.DeviceId))
{
var convDesc = GetConvDescriptor(cd, w.ElementType);
using (var wDesc = GetFilterDescriptor(w))
using (var dyDesc = GetDescriptor(dy))
using (var dxDesc = GetDescriptor(dx))
{
return(dnn.Value.GetConvolutionBackwardDataWorkspaceSize(
wDesc,
dyDesc,
convDesc,
dxDesc,
(cudnnConvolutionBwdDataAlgo)algo));
}
}
}
