A Novel Architecture of Special Arithmetic Function Unit for Area-Efficient Programmable Vertex Shader

A novel architecture of high precision, floating-point special Arithmetic function unit (SFU) for elementary transcendental functions is presented in this paper to provide area efficiency as well as high performance for programmable vertex shader. From the architecture point of view, the evaluation of quadratic approximation for special functions is performed by sharing the SIMD vector unit in shader architecture to minimize processing latency and to reduce area cost in SFU. An optimized minimax approach is proposed as well to obtain the finite-length and normalized quadratic coefficients for high precision. The experiment result shows that the proposed SFU can significantly reduce area cost and by adopting the proposed SFU, a vertex shader with Transport triggered architecture (TTA) can achieve 15.0% improvement on average in performance/area ratio for various shading benchmarks.