/// <summary>
/// Called every time a new frame shows up from the camera
/// </summary>
/// <param name="sender"></param>
/// <param name="e"></param>
private void provider_ImageTransaction(object sender, ImageReceivedEventArgs e)
{
if (stopping || e == null || e.PixelData == null)
{
return;
}
if (Monitor.TryEnter(this))
{
imgGPU.Scatter(e.PixelData);
if (calibrating)
{
// compute a running mean of each pixel
worker.Launch(RunningMean_Kernel, lp, e.Width, imgGPU.Ptr, meanImg.Ptr, numCalibrationSamples);
numCalibrationSamples++;
// compute the mean over the full mean image, using a reduction operation
addReduce.Reduce(meanImg.Ptr, scalarOutput.Ptr, numPixels);
var mean = scalarOutput.GatherScalar();
mean /= numPixels;
// update the correction factor for each pixel
worker.Launch(UpdateCorrectionFactor_Kernel, lp, e.Width, meanImg.Ptr, correctionFactor.Ptr, mean);
}
if (correctBrightness)
{
worker.Launch(ApplyCorrectionFactor_Kernel, lp, e.Width, correctionFactor.Ptr, imgGPU.Ptr);
imgGPU.Gather(e.PixelData);
}
img.Bytes = e.PixelData; // note: have to do this last, because if you set it earlier but then modify the bytes it won't update the image
// trigger a new frame event
OnFrameAvailable(new VideoFrame {
Image = img, ImageGPU = imgGPU, Timestamp = e.TimeStamp
});
Monitor.Exit(this);
}
}
protected override bool Execute(Assignment assignment, ILValue <T> output)
{
var input = assignment.GetInput(Input).ToRValue();
var reduce = Reduction;
var reductionIndices = ReductionIndices;
var shape = Shape;
// unit stride case and reduction is full reduction (to scalar)
if (input.Layout.IsFullyUnitStride && shape.Rank == 0)
{
var length = input.Layout.Shape.Length;
Util.EnsureTrue(length > 0L);
var read = input.BufferReader.GetFlatReader1();
var write = output.Buffer.FlatWriter1;
if (assignment.Context.Type == ContextType.Gpu)
{
var stream = assignment.Context.ToGpuContext().Stream;
var numItems = (int)length;
Func <int, T> sourceOp = i => read(i);
Action <int, T> outputOp = (i, value) => write(i, value);
// first pass, get the size of temp memory
var tempStorageSize = 0;
DeviceReduce.Reduce(stream, new deviceptr <byte>(), ref tempStorageSize, numItems, sourceOp, outputOp, reduce);
// now allocate temp memory
using (var tempMemoryRcpt = assignment.Context.Device.ToGpuDevice().MemoryRepository.Acquire <byte>(tempStorageSize))
{
DeviceReduce.Reduce(stream, tempMemoryRcpt.Memory.Ptr, ref tempStorageSize, numItems, sourceOp, outputOp, reduce);
}
return(true);
}
if (assignment.Context.Type == ContextType.Cpu)
{
var acc = read(0L);
for (var i = 1L; i < length; ++i)
{
acc = reduce(acc, read(i));
}
write(0L, acc);
return(true);
}
}
// currently we only support matrix partial reduce, need TODO to fix this with more generic cases
if (input.Layout.IsFullyUnitStride && input.Layout.Rank == 2 && reductionIndices.Length == 1)
{
var rows = input.Layout.Shape[0];
var cols = input.Layout.Shape[1];
var read = input.BufferReader.GetReader2();
var write = output.Buffer.FlatWriter1;
if (reductionIndices[0] == 1)
{
if (assignment.Context.Type == ContextType.Gpu)
{
var stream = assignment.Context.ToGpuContext().Stream;
var numSegments = (int)rows;
var numItems = (int)cols;
Func <int, int, T> sourceOp = (i, j) => read(i, j);
Action <int, int, T> outputOp = (i, _, value) => write(i, value);
// first pass, get the size of temp memory
var tempStorageSize = 0;
DeviceReduce.Reduce(stream, new deviceptr <byte>(), ref tempStorageSize, numSegments, numItems, sourceOp, outputOp, reduce);
// now allocate temp memory
// TODO: move to assigmnet, manage the temp memory
using (var tempMemoryRcpt = assignment.Context.Device.ToGpuDevice().MemoryRepository.Acquire <byte>(tempStorageSize))
{
DeviceReduce.Reduce(stream, tempMemoryRcpt.Memory.Ptr, ref tempStorageSize, numSegments, numItems, sourceOp, outputOp, reduce);
}
return(true);
}
if (assignment.Context.Type == ContextType.Cpu)
{
for (var row = 0L; row < rows; ++row)
{
var acc = read(row, 0L);
for (var col = 1L; col < cols; ++col)
{
acc = reduce(acc, read(row, col));
}
write(row, acc);
}
return(true);
}
}
if (reductionIndices[0] == 0)
{
if (assignment.Context.Type == ContextType.Gpu)
{
var stream = assignment.Context.ToGpuContext().Stream;
// This is a quick fix, it is not good performance
stream.LongFor(0L, cols, col =>
{
var acc = read(0L, col);
for (var row = 1L; row < rows; ++row)
{
acc = reduce(acc, read(row, col));
}
write(col, acc);
});
return(true);
}
if (assignment.Context.Type == ContextType.Cpu)
{
for (var col = 0L; col < cols; ++col)
{
var acc = read(0L, col);
for (var row = 1L; row < rows; ++row)
{
acc = reduce(acc, read(row, col));
}
write(col, acc);
}
return(true);
}
}
}
return(false);
}