軟體實體層之圖形處理器實作

Abstract

在無線通訊中5G的規範下，現行的C-RAN/Fronthaul架構已無法支撐Remote Radio Head(RRH)與Baseband Unit(BBU)間的傳輸容量，為了克服此問題，BBU與RRH中所執行的功能需要重新定義來解決問題。 在參考文獻[ 1]當中提供了重新定義的方法。在重新定義的BBU與RRH中，必須先將各個功能分割出來。並在各種連線情形中去分配BBU與RRH的功能分配，為了達成功能的快速分配，soft-PHY的概念就被提了出來，而在PC實作soft-PHY下，CPU為系統中唯一的處理單元，整個系統的處理速度和吞吐量受限於CPU的能力，而CPU本身並不擅長大量資料處理，而在現有的系統下，為了讓CPU可以處理soft-PHY的功能，特地降低運算精準度來增加平行度，為了改善現有soft-PHY中CPU處理的缺點，本論文採用GPU來輔助CPU處理資料運算。本論文研究貢獻是試做出soft-PHY使用CPU與GPU偕同運算，並使商用UE可以成功連線，當中試作出PUSCH中傅立葉轉換、通道估測以及通道補償，而在本論文中將CPU與GPU偕同運算過程時，資料在main memory與GPU memory間的次數縮為3次、運算精準度提升為32 bit浮點運算、GPU只使用7~8%。透過試做結果可以推測出使用GPU與CPU協同運算可以支援13個UE同時連線。

According to 5G requirements, the architectures of modem C-RAN/Fronthaul could not be supported the throughputs between the Remote Radio Head (RRH) and Baseband Unit (BBU). For overcoming this issue, we could adopt the approaches [1] of slicing network between the RRH and BBU. For flexible fitting and agile prototyping of RRH and BBU features, the concepts of soft-PHY have been proposed. The implementation of soft-PHY on PC has already been completed in many platforms, but the CPU computing is not suitable for the communication signal processing, especially the single-instruction-multiple-data (SIMD) type computing. Therefore, in this work, we utilize the GPU computing for assisting SIMD architecture efficiency. The collaboration of CPU and GPU for soft-PHY computing, we reduce the times of data copy to 3 between main memory and GPU memory. The data precision is improved from 16-bit fixed point to 32-bit floating point, and finally the GPU utilization decreases the in 7% - 8% for supporting simultaneously connecting with 13 UEs.