| 摘要 |
PURPOSE:To reduce overhead and to reduce a data transfer amount by classifying arrangement access inside a loop into groups based on preceding order relation and generating batch access corresponding to each group outside the loop. CONSTITUTION:When there is access A1, A2... of arrangement data at a global memory 17 concerning a program 10 of an optimizing target, this access is classified into groups for making shadow common, and a transfer range and the shape of the shadow are decided. Then, an instruction for simultaneously receiving the required arrangement data from the global memory 17 all together for each of shadow is added before the loop, and an instruction for simultaneously transmitting the arrangement data as the processing result to the global memory 17 all together is added after the loop. Inside the loop, the shadow at a local memory 18 is accessed. Thus, it is not necessary for each loop to transfer the data of an element unit with the global memory 17, the number of times of data transfer is reduced due to batch transfer, and the transfer amount can be reduced.
|
