/// <summary>
/// Col2s the im.
/// </summary>
/// <param name="col">The col.</param>
/// <param name="im">The im.</param>
/// <param name="channels">The channels.</param>
/// <param name="height">The height.</param>
/// <param name="width">The width.</param>
/// <param name="patch_h">The patch h.</param>
/// <param name="patch_w">The patch w.</param>
/// <param name="pad_h">The pad h.</param>
/// <param name="pad_w">The pad w.</param>
/// <param name="stride_h">The stride h.</param>
/// <param name="stride_w">The stride w.</param>
/// <param name="dilation_h">The dilation h.</param>
/// <param name="dilation_w">The dilation w.</param>
public void Col2Im(Tensor col, Tensor im, int channels, int height, int width,
int patch_h, int patch_w, int pad_h,
int pad_w, int stride_h, int stride_w,
int dilation_h, int dilation_w)
{
var context = CudaHelpers.TSContextForTensor(im);
var cudaContext = context.CudaContextForTensor(im);
int height_col = (height + 2 * pad_h - (dilation_h * (patch_h - 1) + 1))
/ stride_h + 1;
int width_col = (width + 2 * pad_w - (dilation_w * (patch_w - 1) + 1))
/ stride_w + 1;
int num_kernels = channels * height * width;
var data_im = CudaHelpers.GetBufferStart(im);
var data_col = CudaHelpers.GetBufferStart(col);
// From Torch source:
// To avoid involving atomic operations, we will launch one kernel per
// bottom dimension, and then in the kernel add up the top dimensions.
Invoke(context, cudaContext, "col2im_kernel", new dim3(NNThreads.NumBlocks(num_kernels)), new dim3(NNThreads.NumThreads), 0, CUstream.NullStream,
num_kernels, data_col, height, width, channels, patch_h, patch_w, pad_h, pad_w, stride_h, stride_w,
dilation_h, dilation_w,
height_col, width_col, data_im);
}
public void SpatialMaxPoolingBackward(Tensor input, Tensor gradOutput, Tensor gradInput, Tensor indices, ConvolutionDesc2d cd, bool ceilMode)
{
var context = CudaHelpers.TSContextForTensor(gradOutput);
var cudaContext = context.CudaContextForTensor(gradOutput);
var dimw = 3;
var dimh = 2;
var dimc = 1;
var nbatch = input.Sizes[0];
var nslices = input.Sizes[dimc];
var iheight = input.Sizes[dimh];
var iwidth = input.Sizes[dimw];
var owidth = gradOutput.Sizes[dimw];
var oheight = gradOutput.Sizes[dimh];
using var gradOutputContig = Ops.AsContiguous(gradOutput);
var gradOutputPtr = CudaHelpers.GetBufferStart(gradOutputContig);
var indicesPtr = CudaHelpers.GetBufferStart(indices);
var gradInputPtr = CudaHelpers.GetBufferStart(gradInput);
var count = (int)input.ElementCount();
this.Invoke(context, cudaContext, "MaxPoolBackward", new dim3(NNThreads.NumBlocks(count)), new dim3(NNThreads.NumThreads), 0, CUstream.NullStream,
count, gradOutputPtr, indicesPtr, nbatch, nslices, iheight, iwidth, oheight, owidth,
cd.kH, cd.kW, cd.dH, cd.dW, cd.padH, cd.padW, gradInputPtr);
}
/// <summary>
/// Im2s the col.
/// </summary>
/// <param name="im">The im.</param>
/// <param name="col">The col.</param>
/// <param name="channels">The channels.</param>
/// <param name="height">The height.</param>
/// <param name="width">The width.</param>
/// <param name="ksize_h">The ksize h.</param>
/// <param name="ksize_w">The ksize w.</param>
/// <param name="pad_h">The pad h.</param>
/// <param name="pad_w">The pad w.</param>
/// <param name="stride_h">The stride h.</param>
/// <param name="stride_w">The stride w.</param>
/// <param name="dilation_h">The dilation h.</param>
/// <param name="dilation_w">The dilation w.</param>
public void Im2Col(Tensor im, Tensor col, int channels,
int height, int width,
int ksize_h, int ksize_w, int pad_h,
int pad_w, int stride_h, int stride_w,
int dilation_h, int dilation_w)
{
var context = CudaHelpers.TSContextForTensor(im);
var cudaContext = context.CudaContextForTensor(im);
// From Torch source:
// We are going to launch channels * height_col * width_col kernels, each
// kernel responsible for copying a single-channel grid.
int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1))
/ stride_h + 1;
int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1))
/ stride_w + 1;
int num_kernels = channels * height_col * width_col;
var data_im = CudaHelpers.GetBufferStart(im);
var data_col = CudaHelpers.GetBufferStart(col);
Invoke(context, cudaContext, "im2col_kernel", new dim3(NNThreads.NumBlocks(num_kernels)), new dim3(NNThreads.NumThreads), 0, CUstream.NullStream,
num_kernels, data_im, height, width, ksize_h, ksize_w,
pad_h, pad_w, stride_h, stride_w,
dilation_h, dilation_w,
height_col, width_col, data_col);
}
public void SpatialMaxPoolingForward(Tensor input, Tensor output, Tensor indices, ConvolutionDesc2d cd, bool ceilMode)
{
var context = CudaHelpers.TSContextForTensor(input);
var cudaContext = context.CudaContextForTensor(input);
var iwidth = input.Sizes[3];
var iheight = input.Sizes[2];
var nInputPlane = input.Sizes[1];
var batchSize = input.Sizes[0];
long owidth;
long oheight;
if (ceilMode)
{
// ReSharper disable once ArrangeRedundantParentheses
oheight = (long)(Math.Ceiling((float)(iheight - cd.kH + 2 * cd.padH) / cd.dH)) + 1;
// ReSharper disable once ArrangeRedundantParentheses
owidth = (long)(Math.Ceiling((float)(iwidth - cd.kW + 2 * cd.padW) / cd.dW)) + 1;
}
else
{
// ReSharper disable once ArrangeRedundantParentheses
oheight = (long)(Math.Floor((float)(iheight - cd.kH + 2 * cd.padH) / cd.dH)) + 1;
// ReSharper disable once ArrangeRedundantParentheses
owidth = (long)(Math.Floor((float)(iwidth - cd.kW + 2 * cd.padW) / cd.dW)) + 1;
}
if (cd.padW != 0 || cd.padH != 0)
{
// ensure that the last pooling starts inside the image
if ((oheight - 1) * cd.dH >= iheight + cd.padH)
{
--oheight;
}
if ((owidth - 1) * cd.dW >= iwidth + cd.padW)
{
--owidth;
}
}
using var inputContig = Ops.AsContiguous(input);
var inputPtr = CudaHelpers.GetBufferStart(inputContig);
var outputPtr = CudaHelpers.GetBufferStart(output);
var indicesPtr = CudaHelpers.GetBufferStart(indices);
var count = (int)output.ElementCount();
this.Invoke(context, cudaContext, "MaxPoolForward", new dim3(NNThreads.NumBlocks(count)), new dim3(NNThreads.NumThreads), 0, CUstream.NullStream,
count, inputPtr, batchSize, nInputPlane, iheight, iwidth, oheight, owidth,
cd.kH, cd.kW, cd.dH, cd.dW, cd.padH, cd.padW, outputPtr, indicesPtr);
}