public void Forward(SoftmaxLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
float max = float.MinValue;
for (int oc = 0; oc < argument.Channels; oc++)
{
max = Math.Max(max, src[oc]);
}
float sum = 0;
for (int oc = 0; oc < argument.Channels; oc++)
{
float value = (float)Math.Exp(src[oc] - max);
sum += value;
dest[oc] = value;
}
for (int oc = 0; oc < argument.Channels; oc++)
{
dest[oc] /= sum;
}
}
public void Forward(ResizeNearestNeighborLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
float heightScale = (float)argument.InputHeight / argument.OutputHeight;
float widthScale = (float)argument.InputWidth / argument.OutputWidth;
int destIdx = 0;
for (int oc = 0; oc < argument.Channels; oc++)
{
var channelSrc = src.Slice((int)(argument.InputWidth * argument.InputHeight * oc));
for (int oy = 0; oy < argument.OutputHeight; oy++)
{
var inY = (int)Math.Min(Math.Floor(oy * heightScale), argument.InputHeight - 1);
var yOrigin = channelSrc.Slice(inY * (int)argument.InputWidth);
for (int ox = 0; ox < argument.OutputWidth; ox++)
{
var inX = (int)Math.Min(Math.Floor(ox * widthScale), argument.InputWidth - 1);
dest[destIdx++] = yOrigin[inX];
}
}
}
}
public void Forward(LogisticLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
for (int oc = 0; oc < argument.Channels; oc++)
{
dest[oc] = 1f / (1f + (float)Math.Exp(-src[oc]));
}
}
public void Forward(RequantizeLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = context.GetMainRamAt((int)argument.MainMemoryOutputAddress);
for (int i = 0; i < argument.Count; i++)
{
dest[i] = argument.Table[src[i]];
}
}
public void Forward(K210RemovePaddingLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = context.GetMainRamAt((int)argument.MainMemoryOutputAddress);
for (int oc = 0; oc < argument.Channels; oc++)
{
dest[oc] = src[oc * 16];
}
}
public void Forward(DequantizeLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
var q = argument.QuantParam;
for (int i = 0; i < argument.Count; i++)
{
dest[i] = src[i] * q.Scale + q.Bias;
}
}
public void Forward(AddLayerArgument argument, ForwardContext context)
{
var srcA = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAAddress));
var srcB = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputBAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
for (int oc = 0; oc < argument.Count; oc++)
{
dest[oc] = srcA[oc] + srcB[oc];
}
}
public void Forward(ConcatenationLayerArgument argument, ForwardContext context)
{
var dest = context.GetMainRamAt((int)argument.MainMemoryOutputAddress);
foreach (var input in argument.InputsMainMemory)
{
var src = context.GetMainRamAt((int)input.Start, (int)input.Size);
src.CopyTo(dest);
dest = dest.Slice((int)input.Size);
}
}
public void Forward(QuantizeLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = context.GetMainRamAt((int)argument.MainMemoryOutputAddress);
var q = argument.QuantParam;
float scale = 1f / q.Scale;
for (int i = 0; i < argument.Count; i++)
{
int value = (int)Math.Round((src[i] - q.Bias) * scale);
dest[i] = (byte)FxExtensions.Clamp(value, 0, 0xFF);
}
}
public void Forward(FullyConnectedLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast<byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast<byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
for (int oc = 0; oc < argument.OutputChannels; oc++)
{
var weights = new ReadOnlySpan<float>(argument.Weights, (int)(oc * argument.InputChannels), (int)argument.InputChannels);
float sum = 0;
for (int ic = 0; ic < argument.InputChannels; ic++)
sum += src[ic] * weights[ic];
dest[oc] = sum + argument.Bias[oc];
}
}
public void Forward(K210UploadLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = context.GetKpuRamAt((int)argument.KPUMemoryOutputAddress);
K210Helper.KpuUpload(dest, src, (int)argument.Width, (int)argument.Height, (int)argument.Channels);
}
public void Forward(GlobalAveragePool2dLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
int i = 0;
for (int oc = 0; oc < argument.Channels; oc++)
{
float sum = 0;
for (int x = 0; x < argument.KernelSize; x++)
{
sum += src[i++];
}
dest[oc] = sum / argument.KernelSize;
}
}
public void Forward(ChannelwiseDequantizeLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
int idx = 0;
for (int oc = 0; oc < argument.Channels; oc++)
{
var q = argument.QuantParams[oc];
for (int i = 0; i < argument.ChannelSize; i++)
{
dest[idx] = src[idx] * q.Scale + q.Bias;
idx++;
}
}
}
public void Forward(TensorflowFlattenLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
int i = 0;
for (int oy = 0; oy < argument.Height; oy++)
{
for (int ox = 0; ox < argument.Width; ox++)
{
for (int oc = 0; oc < argument.Channels; oc++)
{
dest[i++] = src[(int)((oc * argument.Height + oy) * argument.Width + ox)];
}
}
}
}
public void Forward(AveragePool2dLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
int outIdx = 0;
for (int oc = 0; oc < argument.OutputChannels; oc++)
{
var channelSrc = src.Slice((int)(argument.InputWidth * argument.InputHeight * oc));
for (int oy = 0; oy < argument.OutputHeight; oy++)
{
for (int ox = 0; ox < argument.OutputWidth; ox++)
{
int inXOrigin = (int)(ox * argument.StrideWidth) - (int)argument.PaddingWidth;
int inYOrigin = (int)(oy * argument.StrideHeight) - (int)argument.PaddingHeight;
int kernelXStart = Math.Max(0, -inXOrigin);
int kernelXEnd = Math.Min((int)argument.KernelWidth, (int)argument.InputWidth - inXOrigin);
int kernelYStart = Math.Max(0, -inYOrigin);
int kernelYEnd = Math.Min((int)argument.KernelHeight, (int)argument.InputHeight - inYOrigin);
float value = 0;
float kernelCount = 0;
for (int ky = kernelYStart; ky < kernelYEnd; ky++)
{
for (int kx = kernelXStart; kx < kernelXEnd; kx++)
{
int inX = inXOrigin + kx;
int inY = inYOrigin + ky;
value += channelSrc[inY * (int)argument.InputWidth + inX];
kernelCount++;
}
}
dest[outIdx++] = value / kernelCount;
}
}
}
}
public void Forward(L2NormalizationLayerArgument argument, ForwardContext context)
{
var src = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryInputAddress));
var dest = MemoryMarshal.Cast <byte, float>(context.GetMainRamAt((int)argument.MainMemoryOutputAddress));
float sum = 0;
const float epsilon = 1e-10f;
for (int oc = 0; oc < argument.Channels; oc++)
{
sum += src[oc] * src[oc];
}
if (sum < epsilon)
{
sum = epsilon;
}
sum = 1f / (float)Math.Sqrt(sum);
for (int oc = 0; oc < argument.Channels; oc++)
{
dest[oc] = src[oc] * sum;
}
}
public void Forward(K210AddPaddingLayerArgument argument, ForwardContext context)
{
var src = context.GetMainRamAt((int)argument.MainMemoryInputAddress);
var dest = context.GetKpuRamAt((int)argument.KPUMemoryOutputAddress);
var height = 4;
(var groups, var rowLength, var rowPadding) = (4, 1, 16);
int srcIdx = 0;
for (int oc = 0; oc < argument.Channels; oc++)
{
var channel_origin = oc / groups * rowLength * height * 64 + oc % groups * rowPadding;
for (int y = 0; y < 1; y++)
{
var y_origin = channel_origin + y * rowLength * 64;
for (int x = 0; x < 1; x++)
{
dest[y_origin + x] = src[srcIdx++];
}
}
}
}