public static void ConvolutionBackwardFilter(DNNConvolutionBwdFilterAlgo algo, Cpu.ConvolutionDesc2d cd, CudaStorage workspace, Tensor x, Tensor dy, Tensor dw)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var convDesc = GetConvDescriptor(cd, x.ElementType);
using (var workspacePtr = new CudaDeviceVariable <byte>(workspace.DevicePtrAtElement(0), false, workspace.ByteLength))
using (var xPtr = GetDeviceVar(x))
using (var dyPtr = GetDeviceVar(dy))
using (var dwPtr = GetDeviceVar(dw))
using (var xDesc = GetDescriptor(x))
using (var dyDesc = GetDescriptor(dy))
using (var dwDesc = GetFilterDescriptor(dw))
{
dnn.Value.ConvolutionBackwardFilter(1,
xDesc, xPtr,
dyDesc, dyPtr,
convDesc,
(cudnnConvolutionBwdFilterAlgo)algo,
workspacePtr,
0,
dwDesc, dwPtr);
}
}
}
public static void PoolingBackward(DNNPoolingDesc desc, Tensor x, Tensor y, Tensor dx, Tensor dy)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var poolingDesc = new PoolingDescriptor();
poolingDesc.SetPoolingNdDescriptor((cudnnPoolingMode)desc.Mode, cudnnNanPropagation.PropagateNan, desc.WindowDims.Length,
desc.WindowDims, desc.Padding, desc.Strides);
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var dxPtr = GetDeviceVar(dx))
using (var dyPtr = GetDeviceVar(dy))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
using (var dxDesc = GetDescriptor(dx))
using (var dyDesc = GetDescriptor(dy))
{
// Note: ManagedCUDA argument names may be slightly misleading (src refers to 'y' here, and dest to 'x')
dnn.Value.PoolingBackward(poolingDesc, 1,
yDesc, yPtr,
dyDesc, dyPtr,
xDesc, xPtr,
0,
dxDesc, dxPtr);
}
}
}
public static void ActivationBackward(Tensor x, Tensor y, Tensor dx, Tensor dy, DNNActivation activationType, double clippedReluCeiling)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var activationDesc = new ActivationDescriptor();
activationDesc.SetActivationDescriptor((cudnnActivationMode)activationType,
cudnnNanPropagation.PropagateNan,
clippedReluCeiling);
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var dxPtr = GetDeviceVar(dx))
using (var dyPtr = GetDeviceVar(dy))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
using (var dxDesc = GetDescriptor(dx))
using (var dyDesc = GetDescriptor(dy))
{
dnn.Value.ActivationBackward(activationDesc, 1,
xDesc, xPtr,
dxDesc, dxPtr,
yDesc, yPtr,
0,
dyDesc, dyPtr);
}
}
}
/// <summary>
/// Convs the forward.
/// </summary>
/// <param name="algo">The algo.</param>
/// <param name="cd">The cd.</param>
/// <param name="workspace">The workspace.</param>
/// <param name="x">The x.</param>
/// <param name="w">The w.</param>
/// <param name="y">The y.</param>
public static void ConvForward(DNNConvolutionFwdAlgo algo, Cpu.ConvolutionDesc2d cd, CudaStorage workspace, NDArray x, NDArray w, NDArray y)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var convDesc = GetConvDescriptor(cd, x.ElementType);
using (var workspacePtr = new CudaDeviceVariable <byte>(workspace.DevicePtrAtElement(0), false, workspace.ByteLength))
using (var xPtr = GetDeviceVar(x))
using (var wPtr = GetDeviceVar(w))
using (var yPtr = GetDeviceVar(y))
using (var xDesc = GetDescriptor(x))
using (var wDesc = GetFilterDescriptor(w))
using (var yDesc = GetDescriptor(y))
{
dnn.Value.ConvolutionForward(1,
xDesc, xPtr,
wDesc, wPtr,
convDesc,
(cudnnConvolutionFwdAlgo)algo,
workspacePtr,
0,
yDesc, yPtr);
}
}
}
public Tensor AddmmBatch(Tensor result, float beta, Tensor src, float alpha, Tensor m1, Tensor m2)
{
TSCudaContext context = CudaHelpers.TSContextForTensor(src);
if (src.ElementType != m1.ElementType || src.ElementType != m2.ElementType || (result != null && result.ElementType != src.ElementType))
{
throw new InvalidOperationException("All tensors must have the same element type");
}
if (result != null && !(result.Storage is CudaStorage))
{
throw new ArgumentException("result must be a CUDA tensor", "result");
}
if (!(m1.Storage is CudaStorage))
{
throw new ArgumentException("m1 must be a CUDA tensor", "m1");
}
if (!(m2.Storage is CudaStorage))
{
throw new ArgumentException("m2 must be a CUDA tensor", "m2");
}
if (src.DimensionCount != 3)
{
throw new ArgumentException("src must be a matrix", "src");
}
if (m1.DimensionCount != 3)
{
throw new ArgumentException("m1 must be a matrix", "m1");
}
if (m2.DimensionCount != 3)
{
throw new ArgumentException("m2 must be a matrix", "m2");
}
if (src.Sizes[1] != m1.Sizes[1] || src.Sizes[2] != m2.Sizes[2] || m1.Sizes[2] != m2.Sizes[1])
{
throw new InvalidOperationException($"Size mismatch, srcSize0 = {src.Sizes[0]}, m1Size0 = {m1.Sizes[0]}, srcSize1 = {src.Sizes[1]}, m2Size1 = {m2.Sizes[1]}, m1Size1 = '{m1.Sizes[1]}', m2Size0 = '{m2.Sizes[0]}'");
}
Tensor writeTarget = TensorResultBuilder.GetWriteTarget(result, src, true, src.Sizes);
if (writeTarget != src)
{
Ops.Copy(writeTarget, src);
}
CudaMatrixMulMM.GemmBatch(context, alpha, m1, m2, beta, writeTarget);
return(writeTarget);
}
public NDArray Addmm(NDArray result, float beta, NDArray src, float alpha, NDArray m1, NDArray m2)
{
var context = CudaHelpers.TSContextForTensor(src);
if (src.ElementType != m1.ElementType || src.ElementType != m2.ElementType || (result != null && result.ElementType != src.ElementType))
{
throw new InvalidOperationException("All tensors must have the same element type");
}
if (result != null && !(result.Storage is CudaStorage))
{
throw new ArgumentException("result must be a CUDA tensor", "result");
}
if (!(m1.Storage is CudaStorage))
{
throw new ArgumentException("m1 must be a CUDA tensor", "m1");
}
if (!(m2.Storage is CudaStorage))
{
throw new ArgumentException("m2 must be a CUDA tensor", "m2");
}
if (src.DimensionCount != 2)
{
throw new ArgumentException("src must be a matrix", "src");
}
if (m1.DimensionCount != 2)
{
throw new ArgumentException("m1 must be a matrix", "m1");
}
if (m2.DimensionCount != 2)
{
throw new ArgumentException("m2 must be a matrix", "m2");
}
if (src.Shape[0] != m1.Shape[0] || src.Shape[1] != m2.Shape[1] || m1.Shape[1] != m2.Shape[0])
{
throw new InvalidOperationException("Size mismatch");
}
var writeTarget = TensorResultBuilder.GetWriteTarget(result, src, true, src.Shape);
if (writeTarget != src)
{
Ops.Copy(writeTarget, src);
}
CudaMatrixMulMM.Gemm(context, alpha, m1, m2, beta, writeTarget);
return(writeTarget);
}
public static void ConvolutionBackwardBias(Cpu.ConvolutionDesc2d cd, Tensor dy, Tensor db)
{
using (var dnn = CudaHelpers.TSContextForTensor(dy).DNNForTensor(dy))
{
using (var dyPtr = GetDeviceVar(dy))
using (var dbPtr = GetDeviceVar(db))
using (var dyDesc = GetDescriptor(dy))
using (var dbDesc = GetDescriptor(db))
{
dnn.Value.ConvolutionBackwardBias(1,
dyDesc, dyPtr,
0,
dbDesc, dbPtr);
}
}
}
public static void AddTensor(Tensor src, Tensor result)
{
using (var dnn = CudaHelpers.TSContextForTensor(src).DNNForTensor(src))
{
using (var srcPtr = GetDeviceVar(src))
using (var resultPtr = GetDeviceVar(result))
using (var srcDesc = GetDescriptor(src))
using (var resultDesc = GetDescriptor(result))
{
dnn.Value.AddTensor(1,
srcDesc, srcPtr,
1,
resultDesc, resultPtr);
}
}
}
public static void SoftmaxForward(DNNSoftmaxAlgorithm algorithm, DNNSoftmaxMode mode, Tensor x, Tensor y)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
{
dnn.Value.SoftmaxForward((cudnnSoftmaxAlgorithm)algorithm, (cudnnSoftmaxMode)mode, 1,
xDesc, xPtr,
0,
yDesc, yPtr);
}
}
}
public static void SoftmaxBackward(DNNSoftmaxAlgorithm algorithm, DNNSoftmaxMode mode, Tensor y, Tensor dx, Tensor dy)
{
using (var dnn = CudaHelpers.TSContextForTensor(y).DNNForTensor(y))
{
using (var yPtr = GetDeviceVar(y))
using (var dxPtr = GetDeviceVar(dx))
using (var dyPtr = GetDeviceVar(dy))
using (var yDesc = GetDescriptor(y))
using (var dxDesc = GetDescriptor(dx))
using (var dyDesc = GetDescriptor(dy))
{
dnn.Value.SoftmaxBackward((cudnnSoftmaxAlgorithm)algorithm, (cudnnSoftmaxMode)mode, 1,
yDesc, yPtr,
dyDesc, dyPtr,
0,
dxDesc, dxPtr);
}
}
}
public static void PoolingForward(DNNPoolingDesc desc, Tensor x, Tensor y)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var poolingDesc = new PoolingDescriptor();
poolingDesc.SetPoolingNdDescriptor((cudnnPoolingMode)desc.Mode, cudnnNanPropagation.PropagateNan, desc.WindowDims.Length,
desc.WindowDims, desc.Padding, desc.Strides);
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
{
dnn.Value.PoolingForward(poolingDesc, 1,
xDesc, xPtr,
0,
yDesc, yPtr);
}
}
}
public Tensor Dot(Tensor result, Tensor lhs, Tensor rhs)
{
var context = CudaHelpers.TSContextForTensor(lhs);
if (lhs.DimensionCount == 1 && rhs.DimensionCount == 1)
{
return(CudaMatrixMulDot.Dot(context, result, lhs, rhs));
}
else if (lhs.DimensionCount == 2 && rhs.DimensionCount == 1)
{
return(CudaMatrixMulMV.Mul_M_V(context, result, lhs, rhs));
}
else if (lhs.DimensionCount == 2 && rhs.DimensionCount == 2)
{
return(CudaMatrixMulMM.Mul_M_M(context, result, lhs, rhs));
}
else
{
throw new NotSupportedException(string.Format("Multiplication of {0}D with {1}D tensor is not supported"));
}
}
public NDArray Dot(NDArray result, NDArray lhs, NDArray rhs)
{
var context = CudaHelpers.TSContextForTensor(lhs);
if (lhs.DimensionCount == 1 && rhs.DimensionCount == 1)
{
return(CudaMatrixMulDot.Dot(context, result, lhs, rhs));
}
else if (lhs.DimensionCount == 2 && (rhs.DimensionCount == 1 || rhs.PossibleVector))
{
return(CudaMatrixMulMV.Mul_M_V(context, result, lhs, rhs.Ravel()).Reshape(lhs.Shape[0], 1));
}
else if (lhs.DimensionCount == 2 && rhs.DimensionCount == 2)
{
return(CudaMatrixMulMM.Mul_M_M(context, result, lhs, rhs));
}
else
{
throw new NotSupportedException(string.Format("Multiplication of {0}D with {1}D tensor is not supported"));
}
}
/// <summary>
/// Convolutions the backward data.
/// </summary>
/// <param name="algo">The algo.</param>
/// <param name="cd">The cd.</param>
/// <param name="workspace">The workspace.</param>
/// <param name="w">The w.</param>
/// <param name="dy">The dy.</param>
/// <param name="dx">The dx.</param>
public static void ConvolutionBackwardData(DNNConvolutionBwdDataAlgo algo, Cpu.ConvolutionDesc2d cd, CudaStorage workspace, NDArray w, NDArray dy, NDArray 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,
(cudnnConvolutionBwdDataAlgo)algo,
workspacePtr, 0f,
dxDesc, dxPtr);
}
}
}
public static Tensor Addmm(Tensor result, float beta, Tensor src, float alpha, Tensor m1, Tensor m2)
{
try
{
TSCudaContext context = CudaHelpers.TSContextForTensor(src);
if (src.ElementType != m1.ElementType || src.ElementType != m2.ElementType || (result != null && result.ElementType != src.ElementType))
{
throw new InvalidOperationException("All tensors must have the same element type");
}
if (result != null && !(result.Storage is CudaStorage))
{
throw new ArgumentException("result must be a CUDA tensor", nameof(result));
}
if (!(m1.Storage is CudaStorage))
{
throw new ArgumentException("m1 must be a CUDA tensor", nameof(m1));
}
if (!(m2.Storage is CudaStorage))
{
throw new ArgumentException("m2 must be a CUDA tensor", nameof(m2));
}
if (src.DimensionCount != 2)
{
throw new ArgumentException("src must be a matrix", nameof(src));
}
if (m1.DimensionCount != 2)
{
throw new ArgumentException("m1 must be a matrix", nameof(m1));
}
if (m2.DimensionCount != 2)
{
throw new ArgumentException("m2 must be a matrix", nameof(m2));
}
if (src.Sizes[0] != m1.Sizes[0] || src.Sizes[1] != m2.Sizes[1] || m1.Sizes[1] != m2.Sizes[0])
{
throw new InvalidOperationException($"Size mismatch, srcSize0 = {src.Sizes[0]}, m1Size0 = {m1.Sizes[0]}, srcSize1 = {src.Sizes[1]}, m2Size1 = {m2.Sizes[1]}, m1Size1 = '{m1.Sizes[1]}', m2Size0 = '{m2.Sizes[0]}'");
}
Tensor writeTarget = TensorResultBuilder.GetWriteTarget(result, src, false, src.Sizes);
if (writeTarget != src)
{
Ops.Copy(writeTarget, src);
}
CudaMatrixMulMM.Gemm(context, alpha, m1, m2, beta, writeTarget);
return(writeTarget);
}
catch (Exception err)
{
Logger.WriteLine($"Exception in Addmm: '{err.Message}'");
Logger.WriteLine($"Call stack: '{err.StackTrace}'");
throw;
}
}
