Fig. 1: Basic principles of the PTFP.
From: High-order tensor flow processing using integrated photonic circuits
a Principles of the conventional GeMM and the PTFP. Before matrix multiplication, the GeMM reshapes and duplicates the input tensor ([Cin, L, L]) to the transformed data ([Cin × Dkernel, L × L]). Input data is duplicated Dkernel times. After the matrix multiplier, outputs are yielded and can be reconstructed to convolved feature maps. In the PTFP approach (marked with ‘Flow’), the input tensor is reshaped and enters the PTFP in serial. Each input channel is temporally modulated onto an individual wavelength. A line in the PTFP schematic represents a convolutional operation between an input channel and an output channel. Inside each line, signals are delayed, weighted, and summed so that a temporal convolution (a.k.a. FIR filter) is completed. An output channel is yielded by combining convolved signals from different input channels. Other output channels can be realized by spatially duplicating the same structure. b Conceptual schematic of the PTFP chip. EOM electro-optic modulator, MUX wavelength multiplexer, ODL optical delay line. The directional couplers and delay lines perform data duplication and shifting in the optical domain. Multiple wavelengths are split and delayed in parallel. Crossing waveguides are virtually broken for the succinctness of the graph.
