An Approach for Matrix Multiplication of 32-Bit Fixed Point Numbers by Means of 16-Bit SIMD Instructions on DSP

Matrix multiplication is an important operation for many engineering applications. Sometimes new features that include matrix should be added to existing and even out-of-date embedded platforms. In this paper, unusual problem considered: how implement of 32-bit signed integers fixed-point numbers on DSP having SIMD instructions 16-bit only. For examined tasks, size may vary from several tens two hundred. The proposed mathematical approach dense rectangular comprises decomposition matrices numbers, four multiplications unsigned via outer product, correction outcome fixed point numbers. Several tricks performance optimization are analyzed. addition, ways block-wise parallel implementations described. An implementation the method by means vector faster than using scalar demonstrates close a theoretically achievable limit. described technique can generalized n-bit handling with n/2-bit integers. conclusion, recommendations practitioners who work various presented.