To search, Click below search items. |
| All Published Papers Search Service |
|
|
|
Title |
Architecture of a Graphics Floating-Point Processing Unit with Multi-Word Load and Selective Result Store Mechanisms |
|
Author |
Jiro Miyake, Shigeo Kuninobu, Takaaki Baba |
| Citation |
Vol. 9 No. 6 pp. 200-207 |
|
Abstract |
We have proposed a graphics floating-point processing unit (G-FPU) with 48% reduction of hardware for a conventional processing unit that has both functions of a SIMD-type execution unit dedicated for multiply-accumulate operations and a general-purpose execution unit. The hardware reduction is obtained by realizing a dual-structured general-purpose execution unit that can handle both repeated operations of multiply-accumulate for geometry transformations and irregular operations such as ray-tracing in graphics processing with 9% increase in the hardware for a SIMD-type execution unit. To utilize multiple execution units that can operate in parallel, the high performance of data transfer is indispensable. Therefore, we have proposed a multi-word load mechanism and a selective result store mechanism to load and store data in parallel with executions. These mechanisms reduce the number of load/store instructions and achieve the high performance of data transfer required for parallel operations. Moreover, they remove a buffer memory of 7.9 K gates that temporarily stores data for executions. The effective data transfer reduces the processing cycles for intersection calculation by 26% and geometry transformation by 39%, compared with the case that conventional load/store instructions are used. |
|
Keywords |
Floating-point processor, Parallel operation, Data transfer, Graphics processing |
|
URL |
