Open-Source: GEMM (General Matrix Multiply) acceleration as a computational backbone in Digital Signal Processing (DSP) pipelines
An example of a DSP pipeline is estimating the relative velocity of objects in a RADAR system. The computation involves Fast Fourier Transforms (FFTs) which in turn, can be expressed in terms of matrix multiplications (via the Cooley-Tukey algorithm). The demonstrator will feature:
- two or more RedMulE(Reduced-Precision Matrix Multiplication Engine) instances.
- a RISC-V core: ibex(fork)
- compiler2 support for splitting the original GEMM operation in two or more sub-GEMMs. The sub-GEMMs will be executed in parallel, using the multiple RedMulE instances available in the system.
- investigation of using OpenMP for allocating the sub-GEMMs to RedMulE instances (in collaboration with NXP-RO).