The MLP kernels provided by this release include a one CU (Compute Unit) and a 4-CU Fcn designs.
As shown in the Figure 1, 2 L2 level designs are provided by this release, both are targeting the Alveo U250
platform. The single Fcn CU design instantiates a Fcn kernel that is connected to a DDR channel.
The 4-CU Fcn design instantiates the Fcn Kernel, shown on the left side of Figure 1, 4 times to form
4 CUs. Each CU is connected to a dedicated DDR channel.
Figure 50 Figure 1. FCN support at L2 level
The Fcn Kernel shown in Figure 1 includes a fully connected (FC) engine implementation that
realizes the dense matrix-matrix multiplication and sigmoid function for each adjacent fully connected layers.
The chaining of the operations is implemented by the instruction decoder. That is, using instruction
to encode the relationship between the input and output matrices. Each instruction causes the FC engine to
be triggered once. Each run of the CU can involve several instruction; hence, triggering FC
engine several times to complete the inference process of the fully connected neural network.
You can find the usage of these two designs from directories L2/tests/mlp/fcn_1CU and L2/tests/mlp/fcn_4CU.