public override void FeedForward()
{
convolutionalLayer1.FeedForward();
/*
*
* float[] conv1outputAll = new float[convolutionalLayer1.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize];
* OpenCLSpace.ClError = Cl.EnqueueReadBuffer(OpenCLSpace.Queue,
* convolutionalLayer1.OutputNeurons.ActivationsGPU, // source
* Bool.True,
* (IntPtr)0,
* (IntPtr)(convolutionalLayer1.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize * sizeof(float)),
* conv1outputAll, // destination
* 0,
* null,
* out OpenCLSpace.ClEvent);
* OpenCLSpace.CheckErr(OpenCLSpace.ClError, "NeuralNetwork.ForwardPass Cl.clEnqueueReadBuffer layerInput");
*
* Console.WriteLine("\nConvLayer1 output activations:");
* for (int m = 0; m < inputNeurons.MiniBatchSize; m++)
* {
* float[] layerOutput = new float[convolutionalLayer1.OutputNeurons.NumberOfUnits];
* Array.Copy(conv1outputAll, m * convolutionalLayer1.OutputNeurons.NumberOfUnits, layerOutput, 0, convolutionalLayer1.OutputNeurons.NumberOfUnits);
*
* Console.WriteLine("\n --- Mini-batch item {0} -----", m);
* for (int j = 0; j < layerOutput.Length; j++)
* Console.Write("{0} ", layerOutput[j]);
* Console.WriteLine();
* Console.ReadKey();
* }
*/
if (nonlinearityType == "ReLU")
{
nonlinearityReLU.FeedForward();
}
else if (nonlinearityType == "ELU")
{
nonlinearityELU.FeedForward();
}
/*
* float[] nonlinOutputAll = new float[nonlinearity.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize];
* OpenCLSpace.ClError = Cl.EnqueueReadBuffer(OpenCLSpace.Queue,
* nonlinearity.OutputNeurons.ActivationsGPU, // source
* Bool.True,
* (IntPtr)0,
* (IntPtr)(convolutionalLayer1.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize * sizeof(float)),
* nonlinOutputAll, // destination
* 0,
* null,
* out OpenCLSpace.ClEvent);
* OpenCLSpace.CheckErr(OpenCLSpace.ClError, "NeuralNetwork.ForwardPass Cl.clEnqueueReadBuffer layerInput");
*
* Console.WriteLine("\nNonlinearity output activations:");
* for (int m = 0; m < inputNeurons.MiniBatchSize; m++)
* {
* float[] layerOutput = new float[nonlinearity.OutputNeurons.NumberOfUnits];
* Array.Copy(nonlinOutputAll, m * nonlinearity.OutputNeurons.NumberOfUnits, layerOutput, 0, nonlinearity.OutputNeurons.NumberOfUnits);
*
* Console.WriteLine("\n --- Mini-batch item {0} -----", m);
* for (int j = 0; j < layerOutput.Length; j++)
* Console.Write("{0} ", layerOutput[j]);
* Console.WriteLine();
* Console.ReadKey();
* }
*/
convolutionalLayer2.FeedForward();
/*
* float[] conv2outputAll = new float[convolutionalLayer2.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize];
* OpenCLSpace.ClError = Cl.EnqueueReadBuffer(OpenCLSpace.Queue,
* convolutionalLayer2.OutputNeurons.ActivationsGPU, // source
* Bool.True,
* (IntPtr)0,
* (IntPtr)(convolutionalLayer2.OutputNeurons.NumberOfUnits * inputNeurons.MiniBatchSize * sizeof(float)),
* conv2outputAll, // destination
* 0,
* null,
* out OpenCLSpace.ClEvent);
* OpenCLSpace.CheckErr(OpenCLSpace.ClError, "NeuralNetwork.ForwardPass Cl.clEnqueueReadBuffer layerInput");
*
* Console.WriteLine("\nConvLayer2 output activations:");
* for (int m = 0; m < inputNeurons.MiniBatchSize; m++)
* {
* float[] layerOutput = new float[convolutionalLayer2.OutputNeurons.NumberOfUnits];
* Array.Copy(conv2outputAll, m * convolutionalLayer2.OutputNeurons.NumberOfUnits, layerOutput, 0, convolutionalLayer2.OutputNeurons.NumberOfUnits);
*
* Console.WriteLine("\n --- Mini-batch item {0} -----", m);
* for (int j = 0; j < layerOutput.Length; j++)
* Console.Write("{0} ", layerOutput[j]);
* Console.WriteLine();
* Console.ReadKey();
* }
*/
// Additionally, cumulate inputs onto outputs
OpenCLSpace.ClError = Cl.SetKernelArg(OpenCLSpace.SkipForward, 0, outputNeurons.ActivationsGPU);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.SkipForward, 1, inputNeurons.ActivationsGPU);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.SkipForward, 2, (IntPtr)sizeof(int), nInputUnits);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.SkipForward, 3, (IntPtr)sizeof(int), inputNeurons.MiniBatchSize);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.SetKernelArg");
// Run kernel
OpenCLSpace.ClError = Cl.EnqueueNDRangeKernel(OpenCLSpace.Queue,
OpenCLSpace.SkipForward,
1,
null,
globalWorkSizePtr,
localWorkSizePtr,
0,
null,
out OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.EnqueueNDRangeKernel");
OpenCLSpace.ClError = Cl.ReleaseEvent(OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.ReleaseEvent");
OpenCLSpace.ClError = Cl.Finish(OpenCLSpace.Queue);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.Finish");
}
