Computing system and transposition method therefor
A computing system and a transposition method therefor. One aspect of the present disclosure provides a computing system that variably determines an input/output unit of a memory depending on a processing unit or an output unit of a processor, and a transposition method for supporting same.
1 . A computing system comprising:
a memory module configured to input or output data from a memory according to a memory layout including a size of a lane corresponding to a size of decomposed data, the decomposed data being obtained from a tensor to be fed to or output from a processor and being addressed together by the memory; and
the processor,
wherein the memory layout is variably controlled based on a size of data simultaneously processed by the processor or based on a size of data simultaneously output by the processor.
2 . The computing system of claim 1 , wherein:
the tensor has multiple dimensions, and is decomposed along one or more of the multiple dimensions,
the size of the lane corresponds to a multiple of a basic processing unit, the basic processing unit representing a size of data located along to a specific one of the decomposed dimensions.
3 . The computing system of claim 1 , wherein the memory module determines the memory layout of the memory based on the size of the lane, and continuously stores consecutive elements of input data depending on the memory layout.
4 . The computing system of claim 1 , further comprising:
a transposition module configured to transpose input data based on consecutively located elements corresponding to the size of data simultaneously processed by the processor.
5 . The computing system of claim 4 , wherein the memory module determines the size of the lane to be R times a specific dimension in relation to a data, wherein a transposition module receives R pieces of input and output data, and the memory module determines the memory layout of the memory entries to perform memory input and output based on the size of the lane.
6 . The computing system of claim 5 , wherein the transposition module receives R element pairs from the memory module and performs R×R transposition, wherein the element pairs are alternately retrieved memory entries.
7 . The computing system of claim 6 , wherein
the memory module sequentially outputs data with alternate memory entries, and
the transposition module performs the Rx R transposition considering the element pairs, which correspond to the size of data simultaneously processed by the processor and are consecutively located among elements contained in the data in units of the memory entry, as a transposition unit.
8 . The computing system of claim 4 , wherein
the memory module determines the size of the lane to be R′ times a basic output unit representing the size of data simultaneously output by the processor, and determines a layout of the memory entries to perform memory input and output based on the size of the lane, wherein R′ is associated with pieces of data from processor to the memory module, and
the transposition module receives R′ element pairs from the processor and performs R′×R′ transposition.
9 . A method of performing transposition in a device, the method comprising:
determining a size of a lane of a memory module based at least in part on a specific dimension in relation to a data type of data of the memory module; and
sequentially receiving R element pairs from the memory module or a processor, and performing R×R transposition based at least in part on consecutively located elements corresponding to the data type of the data.