public ConvolutionalLayerNPP(int widthIn, int heightIn, int channelsIn, int widthOut, int heightOut, int channelsOut, int batch, int filterWidth, int filterHeight, Activation activation, CudaContext ctx, CUmodule module)
: base(widthIn, heightIn, channelsIn, widthOut, heightOut, channelsOut, batch)
{
_activation = activation;
_filterX = filterWidth;
_filterY = filterHeight;
_weights = new CudaDeviceVariable <float>(filterWidth * filterHeight * channelsIn * channelsOut);
_bias = new CudaDeviceVariable <float>(channelsOut);
_y = new CudaDeviceVariable <float>(widthOut * heightOut * channelsOut * batch);
_z = new CudaDeviceVariable <float>(widthOut * heightOut * channelsOut * batch);
_tempConvolution = new CudaDeviceVariable <float>(widthOut * heightOut);
if (_activation == Activation.PRelu || _activation == Activation.LeakyRelu)
{
_aRelu = new CudaDeviceVariable <float>(channelsOut);
_KernelPReluForward = new PReluForwardKernel(module, ctx);
_KernelPReluForward.SetComputeSize((uint)widthOut * (uint)heightOut, (uint)channelsOut, (uint)batch);
}
else
if (_activation == Activation.Relu)
{
_aRelu = new CudaDeviceVariable <float>(channelsOut);
_aRelu.Memset(0); //use a fixed alpha of 0 for Relu...
_KernelPReluForward = new PReluForwardKernel(module, ctx);
_KernelPReluForward.SetComputeSize((uint)widthOut * (uint)heightOut, (uint)channelsOut, (uint)batch);
}
}
public ConvolutionalLayer(int widthIn, int heightIn, int channelsIn, int widthOut, int heightOut, int channelsOut, int batch, int filterWidth, int filterHeight, Activation activation, CudaBlas blasCtx, CudaDNNContext cudnnCtx, CudaContext ctx, CUmodule moduleBorder, CUmodule modulePrelu)
: base(widthIn, heightIn, channelsIn, widthOut, heightOut, channelsOut, batch)
{
_activation = activation;
_filterX = filterWidth;
_filterY = filterHeight;
_weights = new CudaDeviceVariable <float>(filterWidth * filterHeight * channelsIn * channelsOut);
_d_weights = new CudaDeviceVariable <float>(filterWidth * filterHeight * channelsIn * channelsOut);
_bias = new CudaDeviceVariable <float>(channelsOut);
_d_bias = new CudaDeviceVariable <float>(channelsOut);
_dx = new CudaDeviceVariable <float>(widthIn * heightIn * channelsIn * batch);
_y = new CudaDeviceVariable <float>(widthOut * heightOut * channelsOut * batch);
_dy = new CudaDeviceVariable <float>(widthOut * heightOut * channelsOut * batch);
_z = new CudaDeviceVariable <float>(widthOut * heightOut * channelsOut * batch);
_ones = new CudaDeviceVariable <float>(batch);
_withBorderInput = new CudaDeviceVariable <float>((widthIn + filterWidth - 1) * (heightIn + filterHeight - 1) * channelsIn * batch);
_withBorderDx = new CudaDeviceVariable <float>((widthIn + filterWidth - 1) * (heightIn + filterHeight - 1) * channelsIn * batch);
_cudnn = cudnnCtx;
_blas = blasCtx;
_descActivation = new ActivationDescriptor();
_descActivation.SetActivationDescriptor(cudnnActivationMode.Relu, cudnnNanPropagation.NotPropagateNan, 0);
_descBias = new TensorDescriptor();
_descBias.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Float, 1, channelsOut, 1, 1);
_descDataInBorder = new TensorDescriptor();
_descDataIn = new TensorDescriptor();
_descDataIn.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Float, batch, channelsIn, heightIn + filterHeight - 1, widthIn + filterWidth - 1);
_descDataOut = new TensorDescriptor();
_descDataOut.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Float, batch, channelsOut, heightOut, widthOut);
_descFilter = new FilterDescriptor();
_descFilter.SetFilter4dDescriptor(cudnnDataType.Float, cudnnTensorFormat.NCHW, channelsOut, channelsIn, filterWidth, filterHeight);
_descConv = new ConvolutionDescriptor();
_descConvBorder = new ConvolutionDescriptor();
_descConv.SetConvolution2dDescriptor(0, 0, 1, 1, 1, 1, cudnnConvolutionMode.Convolution, cudnnDataType.Float);
int n = 0;
int c = 0;
int h = 0;
int w = 0;
_descConv.GetConvolution2dForwardOutputDim(_descDataIn, _descFilter, ref n, ref c, ref h, ref w);
_kernelAddBorder = new AddBorderKernel(moduleBorder, ctx);
_kernelAddBorder.BlockDimensions = new ManagedCuda.VectorTypes.dim3(widthIn + filterWidth - 1, (heightIn + filterHeight - 1) / 2 + 1, 1);
_kernelCropBorder = new CropBorderKernel(moduleBorder, ctx);
_kernelCropBorder.BlockDimensions = new ManagedCuda.VectorTypes.dim3(widthIn, heightIn / 2 + 1, 1);
if (_activation == Activation.PRelu || _activation == Activation.LeakyRelu)
{
_temp = new CudaDeviceVariable <float>(channelsOut * batch);
_aRelu = new CudaDeviceVariable <float>(channelsOut);
_dARelu = new CudaDeviceVariable <float>(channelsOut);
_KernelPReluForward = new PReluForwardKernel(modulePrelu, ctx);
_KernelPReluBackward = new PReluBackwardKernel(modulePrelu, ctx);
_KernelPReluBackward1 = new PReluBackward1Kernel(modulePrelu, ctx);
_KernelPReluBackward2 = new PReluBackward2Kernel(modulePrelu, ctx);
_KernelPReluForward.SetComputeSize((uint)widthOut * (uint)heightOut, (uint)channelsOut, (uint)batch);
_KernelPReluBackward.SetComputeSize((uint)channelsOut, 1, 1);
}
cudnnConvolutionFwdAlgoPerf[] algos =
_cudnn.FindConvolutionForwardAlgorithm(_descDataIn, _descFilter, _descConv, _descDataOut, 5);
cudnnConvolutionBwdDataAlgoPerf[] algos2 = _cudnn.FindConvolutionBackwardDataAlgorithm(_descFilter, _descDataOut, _descConv, _descDataIn, 5);
_algoFwd = _cudnn.GetConvolutionForwardAlgorithm(_descDataIn, _descFilter, _descConv,
_descDataOut, cudnnConvolutionFwdPreference.PreferFastest, 0);
SizeT sizeInBytes = 0, tmpsize = 0;
sizeInBytes = _cudnn.GetConvolutionForwardWorkspaceSize(_descDataIn, _descFilter,
_descConv, _descDataOut, _algoFwd);
_algoBwdFilter = _cudnn.GetConvolutionBackwardFilterAlgorithm(_descDataIn, _descDataOut, _descConv, _descFilter,
cudnnConvolutionBwdFilterPreference.PreferFastest, 0);
tmpsize = _cudnn.GetConvolutionBackwardFilterWorkspaceSize(_descDataIn, _descDataOut, _descConv, _descFilter, _algoBwdFilter);
sizeInBytes = Math.Max(sizeInBytes, tmpsize);
_algoBwdData = _cudnn.GetConvolutionBackwardDataAlgorithm(_descFilter, _descDataOut, _descConv, _descDataIn, cudnnConvolutionBwdDataPreference.PreferFastest, 0);
tmpsize = _cudnn.GetConvolutionBackwardDataWorkspaceSize(_descFilter, _descDataOut, _descConv, _descDataIn, _algoBwdData);
sizeInBytes = Math.Max(sizeInBytes, tmpsize);
if (sizeInBytes > 0)
{
_workspace = new CudaDeviceVariable <byte>(sizeInBytes);
}
else
{
_workspace = CudaDeviceVariable <byte> .Null;
}
}
