BNNSDataLayoutConvolutionWeightsOIHW_Pack32
A constant that represents a 4D array of packed convolution weights with 32-output channel packing and 128-byte array address alignment.
Declaration
var BNNSDataLayoutConvolutionWeightsOIHW_Pack32: BNNSDataLayout { get }Discussion
The Value (kx, ky, InChannel, OutChannel) is at index:
OutChannelPositionInGroup + kw * 32 + ky * kernel_width * 32 + InChannel * kernel_height * kernel_width * 32 + OutChannelGroup * input_channels * kernel_height * kernel_width * 32
Where:
kernel_widthis the kernel width.kernel_heightis the kernel height.input_channelsis the number of input channels.output_channelsis the number of output channels.OutChannelGroup = OutChannel / 32.OutChannelPositionInGroup = OutChannel % 32.kwissize[0]andkxis between0tokw-1.khissize[1]andkyis between0tokh-1.