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 K210Conv2dLayerArgument Convert(K210Conv2d layer, ConvertContext context)
{
var config = new K210ConvLayerConfig {
BNConfigs = new K210LayerBNConfig[layer.OutputChannels], ActConfigs = new K210LayerActConfig[16]
};
(var sw, var bw) = QuantizeWeights(layer.Conv2dType == K210Conv2dType.Conv2d, layer.Weights, config, context.WeightsBits);
(var sx, var bx) = QuantizeInput(context.Quantization.Distributions[layer.Input.Connection.From].Global, config);
config.ArgAdd = (long)Math.Round(bw * bx * layer.KernelWidth * layer.KernelHeight);
var scale = new double[layer.OutputChannels];
for (int i = 0; i < scale.Length; i++)
{
scale[i] = sw[i] * sx;
}
QuantizeBiasAndOutput(layer, layer.Bias, context.Quantization.Distributions[layer.Output], context.Quantization.AdditionalDistributions[layer.OutputBeforeActivation], scale, config);
config.InputChannels = layer.InputChannels;
config.OutputChannels = layer.OutputChannels;
config.InputWidth = layer.Input.Dimensions[3];
config.InputHeight = layer.Input.Dimensions[2];
(config.InputGroups, config.InputRowLength) = K210Helper.GetRowLayout(config.InputWidth);
config.OutputWidth = layer.Output.Dimensions[3];
config.OutputHeight = layer.Output.Dimensions[2];
(config.OutputGroups, config.OutputRowLength) = K210Helper.GetRowLayout(config.OutputWidth);
config.KernelType = layer.KernelWidth == 3 ? 1 : 0;
config.IsDepthwise = layer.Conv2dType == K210Conv2dType.DepthwiseConv2d;
config.PoolType = (int)layer.PoolType;
config.PadValue = (int)Math.Round(-bx);
if (layer.Conv2dType == K210Conv2dType.Conv2d)
{
var kernelSize = (int)layer.Weights.Length * context.WeightsBits / 8;
var oneChannelSize = layer.KernelWidth * layer.KernelHeight * layer.InputChannels * context.WeightsBits / 8;
var sizeLimit = context.WeightsBits == 8 ? 30 : 60;
var oneLoadChannels = Math.Min(layer.OutputChannels, (int)Math.Floor(sizeLimit * 1024.0 / oneChannelSize));
config.OneLoadKernelsSize = oneChannelSize * oneLoadChannels;
config.LoadTimes = (int)Math.Ceiling(layer.OutputChannels / (double)oneLoadChannels);
config.OutputChannelsOnTime = oneLoadChannels;
}
else
{
config.OneLoadKernelsSize = (int)layer.Weights.Length * context.WeightsBits / 8;
config.LoadTimes = 1;
config.OutputChannelsOnTime = layer.OutputChannels;
}
var inputOneLineChannels = Math.Min(layer.InputChannels, config.InputGroups);
config.InputSize = config.InputRowLength * config.InputHeight * config.InputChannels / inputOneLineChannels;
var outputOneLineChannels = Math.Min(layer.OutputChannels, config.OutputGroups);
config.OutputSize = config.OutputRowLength * config.OutputHeight * config.OutputChannels / outputOneLineChannels;
return(new K210Conv2dLayerArgument
{
Config = config,
ParamAddress = new K210Conv2dParamAddress()
});
}
