public override void Conv2DInputGradient(Tensor gradient, Tensor rotKernels, int stride, int paddingX, int paddingY, Tensor inputGradients)
{
GpuShape[] shapes = new[] { new GpuShape(gradient.Shape),
new GpuShape(rotKernels.Shape),
new GpuShape(inputGradients.Shape),
new GpuShape(rotKernels.Width, rotKernels.Height, 1, rotKernels.BatchSize) };
float[] devGradient = Gpu.CopyToDevice(gradient.Values);
float[] devRotKernels = Gpu.CopyToDevice(rotKernels.Values);
GpuShape[] devShapes = Gpu.CopyToDevice(shapes);
int threadsRequiredPerResultElem = rotKernels.BatchSize * rotKernels.Height * rotKernels.Width;
float[,] resultPartials = new float[inputGradients.Length, GetBlocksNum(threadsRequiredPerResultElem)];
float[,] devResultPartials = Gpu.Allocate(resultPartials);
// simulate
//GpuConv2DInputGradient(GetSimulatedThread(blockSize, new dim3(bx, by, bz), new dim3(tx, ty, tz)), gradient.Values, rotKernels.Values, resultPartials, shapes, paddingX, paddingY, stride);
Gpu.Launch(new dim3(inputGradients.Length, GetBlocksNum(threadsRequiredPerResultElem)), THREADS_PER_BLOCK).GpuConv2DInputGradient(devGradient, devRotKernels, devResultPartials, devShapes, paddingX, paddingY, stride);
Gpu.Synchronize();
Gpu.CopyFromDevice(devResultPartials, resultPartials);
Gpu.FreeAll();
for (int k = 0; k < resultPartials.GetLength(0); ++k)
{
for (int partialId = 0; partialId < resultPartials.GetLength(1); ++partialId)
{
inputGradients.Values[k] += resultPartials[k, partialId];
}
}
}
//public override void Add(Tensor t1, Tensor t2, Tensor result)
//{
// int threadsRequired = result.Length;
// float[] devT1 = Gpu.CopyToDevice(t1.Values);
// float[] devT2 = Gpu.CopyToDevice(t2.Values);
// float[] devResult = Gpu.Allocate(result.Values);
// Gpu.Launch(GetBlocksNum(threadsRequired), THREADS_PER_BLOCK).GpuAdd(devT1, devT2, devResult);
// Gpu.Synchronize();
// Gpu.CopyFromDevice(devResult, result.Values);
// Gpu.FreeAll();
//}
//public override void Sub(Tensor t1, Tensor t2, Tensor result)
//{
// int threadsRequired = result.Length;
// float[] devT1 = Gpu.CopyToDevice(t1.Values);
// float[] devT2 = Gpu.CopyToDevice(t2.Values);
// float[] devResult = Gpu.Allocate(result.Values);
// Gpu.Launch(GetBlocksNum(threadsRequired), THREADS_PER_BLOCK).GpuSub(devT1, devT2, devResult);
// Gpu.Synchronize();
// Gpu.CopyFromDevice(devResult, result.Values);
// Gpu.FreeAll();
//}
//public override void Mul(Tensor t1, Tensor t2, Tensor result)
//{
// int threadsRequired = result.BatchSize * t1.Depth * t1.Height * t2.Width;
// GpuShape[] shapes = new [] { new GpuShape(t1.Shape), new GpuShape(t2.Shape), new GpuShape(result.Shape) };
// float[] devT1 = Gpu.CopyToDevice(t1.Values);
// float[] devT2 = Gpu.CopyToDevice(t2.Values);
// float[] devResult = Gpu.Allocate(result.Values);
// GpuShape[] devShapes = Gpu.CopyToDevice(shapes);
// Gpu.Launch(GetBlocksNum(threadsRequired), THREADS_PER_BLOCK).GpuMul(devT1, devT2, devResult, devShapes);
// Gpu.Synchronize();
// Gpu.CopyFromDevice(devResult, result.Values);
// Gpu.FreeAll();
//}
public override void Conv2D(Tensor t, Tensor kernels, int stride, int paddingX, int paddingY, Tensor result)
{
int threadsRequired = t.BatchSize * kernels.BatchSize * result.Width * result.Height;
GpuShape[] shapes = new[] { new GpuShape(t.Shape), new GpuShape(kernels.Shape), new GpuShape(result.Shape) };
float[] devT = Gpu.CopyToDevice(t.Values);
float[] devKernels = Gpu.CopyToDevice(kernels.Values);
float[] devResult = Gpu.Allocate(result.Values);
GpuShape[] devShapes = Gpu.CopyToDevice(shapes);
Gpu.Launch(GetBlocksNum(threadsRequired), THREADS_PER_BLOCK).GpuConv2D(devT, devKernels, devResult, devShapes, paddingX, paddingY, stride);
Gpu.Synchronize();
Gpu.CopyFromDevice(devResult, result.Values);
Gpu.FreeAll();
}