public override float Inference(CudaDeviceVariable <float> input)
{
_input = input;
NPPImage_32fC1 tempConv = new NPPImage_32fC1(_tempConvolution.DevicePointer, InWidth, InHeight, InWidth * sizeof(float));
for (int outLayer = 0; outLayer < OutChannels; outLayer++)
{
SizeT offsetOut = outLayer * OutWidth * OutHeight * sizeof(float);
CUdeviceptr ptrWithOffsetOut = _z.DevicePointer + offsetOut;
NPPImage_32fC1 imgOut = new NPPImage_32fC1(ptrWithOffsetOut, OutWidth, OutHeight, OutWidth * sizeof(float));
imgOut.Set(0);
for (int inLayer = 0; inLayer < InChannels; inLayer++)
{
SizeT offsetIn = inLayer * InWidth * InHeight * sizeof(float);
CUdeviceptr ptrWithOffsetIn = _input.DevicePointer + offsetIn;
NPPImage_32fC1 imgIn = new NPPImage_32fC1(ptrWithOffsetIn, InWidth, InHeight, InWidth * sizeof(float));
imgIn.SetRoi(_filterX / 2, _filterY / 2, InWidth - _filterX + 1, InHeight - _filterY + 1);
SizeT offsetFilter = (outLayer * InChannels * _filterX * _filterY + inLayer * _filterX * _filterY) * sizeof(float);
CudaDeviceVariable <float> filter = new CudaDeviceVariable <float>(_weights.DevicePointer + offsetFilter, false, _filterX * _filterY * sizeof(float));
imgIn.Filter(tempConv, filter, new NppiSize(_filterX, _filterY), new NppiPoint(_filterX / 2, _filterY / 2));
imgOut.Add(tempConv);
}
imgOut.Add(bHost[outLayer]);
}
switch (_activation)
{
case Activation.None:
_y.CopyToDevice(_z);
break;
case Activation.Relu:
//_aRelu is set to 0!
_KernelPReluForward.RunSafe(_z, _aRelu, _y, _outWidth * _outHeight, _outChannels, _batch);
break;
case Activation.PRelu:
_KernelPReluForward.RunSafe(_z, _aRelu, _y, _outWidth * _outHeight, _outChannels, _batch);
break;
case Activation.LeakyRelu:
_KernelPReluForward.RunSafe(_z, _aRelu, _y, _outWidth * _outHeight, _outChannels, _batch);
break;
default:
break;
}
return(_nextLayer.Inference(_y));
}
public void Track(NPPImage_32fC1 imgTrack, NPPImage_32fC1 imgRef, NPPImage_32fC2 preShift, int i, float2 baseShiftRef, float baseRotationRef, float2 baseShifttoTrack, float baseRotationtoTrack, float threshold)
{
if (imgTrack.WidthRoi != imgRef.WidthRoi || imgTrack.HeightRoi != imgRef.HeightRoi ||
imgTrack.WidthRoi != currentWidth || imgTrack.HeightRoi != currentHeight)
{
throw new ArgumentOutOfRangeException();
}
int level = imgTrack.Width / imgTrack.WidthRoi;
convertToTilesBorder.RunSafe(imgRef, imgRefSortedTiles, currentTileSize, currentMaxShift, CurrentBlockCountX, CurrentBlockCountY, baseShiftRef, baseRotationRef); //template
forward[i].Exec(imgRefSortedTiles.DevicePointer, imgRefCplx.DevicePointer);
convertToTiles.RunSafe(imgTrack, imgToTrackSortedTiles, preShift, currentTileSize, currentMaxShift, CurrentBlockCountX, CurrentBlockCountY, baseShifttoTrack, baseRotationtoTrack); //image in paper
//DumpFloat(imgToTrackSortedTiles, currentTileSize + 2* currentMaxShift, currentTileSize + 2 * currentMaxShift, CurrentBlockCountX * CurrentBlockCountY, tileIdx, "tilesTrack_" + level + "_" + debugCallCounter + ".bin");
//DumpFloat(imgRefSortedTiles, currentTileSize + 2 * currentMaxShift, currentTileSize + 2 * currentMaxShift, CurrentBlockCountX * CurrentBlockCountY, tileIdx, "tilesRef_" + level + "_" + debugCallCounter + ".bin");
forward[i].Exec(imgToTrackSortedTiles.DevicePointer, imgToTrackCplx.DevicePointer);
conjKernel.RunSafe(imgRefCplx, imgToTrackCplx);
backward[i].Exec(imgToTrackCplx.DevicePointer, imgCrossCorrelation.DevicePointer);
imgCrossCorrelation.DivC(CurrentBlockSize * CurrentBlockSize);
squaredSumKernel.RunSafe(imgRefSortedTiles, squaredSumsOfTiles, currentMaxShift, currentTileSize, CurrentBlockCountX * CurrentBlockCountY);
//DumpFloat(squaredSumsOfTiles, 1, 1, CurrentBlockCountX * CurrentBlockCountY, tileIdx, "squaredSums_" + level + "_" + debugCallCounter + ".bin");
boxFilterXKernel.RunSafe(imgToTrackSortedTiles, imgRefSortedTiles, currentMaxShift, currentTileSize, CurrentBlockCountX * CurrentBlockCountY);
boxFilterYKernel.RunSafe(imgRefSortedTiles, imgToTrackSortedTiles, currentMaxShift, currentTileSize, CurrentBlockCountX * CurrentBlockCountY);
//DumpFloat(imgToTrackSortedTiles, currentTileSize + 2 * currentMaxShift, currentTileSize + 2 * currentMaxShift, CurrentBlockCountX * CurrentBlockCountY, tileIdx, "boxFilter_" + level + "_" + debugCallCounter + ".bin");
normalizedCCKernel.RunSafe(imgCrossCorrelation, squaredSumsOfTiles, imgToTrackSortedTiles, shiftImages, currentMaxShift, currentTileSize, CurrentBlockCountX * CurrentBlockCountY);
//DumpFloat(shiftImages, (2 * currentMaxShift + 1), (2 * currentMaxShift + 1), CurrentBlockCountX * CurrentBlockCountY, tileIdx, "tilesShift_" + level + "_" + debugCallCounter + ".bin");
patchShift.SetRoi(0, 0, CurrentBlockCountX, CurrentBlockCountY);
findMinimumKernel.RunSafe(shiftImages, patchShift, currentMaxShift, CurrentBlockCountX, CurrentBlockCountY, threshold);
NPPImage_32fC1 preShiftFloat = new NPPImage_32fC1(preShift.DevicePointer, 2 * CurrentBlockCountX, CurrentBlockCountY, preShift.Pitch);
NPPImage_32fC1 patchShiftFloat = new NPPImage_32fC1(patchShift.DevicePointer, 2 * CurrentBlockCountX, CurrentBlockCountY, patchShift.Pitch);
preShiftFloat.Add(patchShiftFloat);
debugCallCounter++;
}
