Реализация масштабированного векторного сложения многократной точности на GPU

К.С. Исупов, А.С. Куваев
2019 №1 (2018)  
Many modern problems solved on supercomputers require linear algebra operations to be performed with a precision exceeding the IEEE 754 formats. This article discusses the implementation of highprecision scaled vector addition (WAXPBY) on CUDAcompatible graphics processing units (GPUs). To represent multipleprecision floatingpoint numbers, a format based on a residue number system is used, which eliminates the need for carry propagation and allows to compute all significand digits
more » ... igits simultaneously. The parallel WAXPBY algorithm is divided into a number of steps, each of which is performed as a separate CUDA kernel, which allows for efficient use of available computing resources. Experiments have shown that the performance of the developed routine is higher compared to solutions for central processing units. The results can be useful in linear algebra applications that use GPUs to speed up calculations and are critical to rounding errors.
doi:10.25730/vsu.0536.19.017 fatcat:27d336lbinfahktrx26jev3rlu