Network

Tiny-dpcpp-nn is an implementation of multi-layer perceptrions (MLPs) for Intel GPUs. In particular, this implementation focuses on MLPs with constant, small power of 2 network widths (16, 32, 64, 128) and aims to maximize the performance of inference and training by fusing the requisite matrix operations in the separate layers into a single operation.

Multi-Layer Perceptrons

Fused MLPs

As indicate above, Fused MLPs fuse the matrix operations of several layers into a single operation. In general, this decreases the number of global memory access and thus increases the arithmetic intensity and performance. Especially Inference benefits from this since the load of the input of each layer and the store of the output of each layer can be reduced to a single load of the input to the network and a single store of the output of the network.

ESIMD Kernels

The fused forward pass and the backward pass are implemented in Intel’s ESIMD SYCL extension. They can be found in include/network/kernel_esimd.h in the functions forward_impl_general and backward_impl, respectively. The forward_impl_general is templated and re-used for the forward pass required as part of the training of the network as well as the Inference.