Information processing apparatus, control method, and program

An information processing apparatus comprising:

• one or more processors;
• and
• one or more memories storing program instructions executable by the one or more processors, wherein the one or more processors are configured to execute the program instructions to control:
• an acquiring unit to acquire input matrix data information that represents target matrix data in a dense representation format or a sparse representation format, the target matrix data being represented with all data elements of the target matrix data when the target matrix data is represented in the dense representation formats, the target matrix data being represented with non-zero data elements of the target matrix data when the target matrix data is represented in a sparse representation format;
• a sparsity calculation unit to calculate sparsity of the target matrix data;
• a selection unit to select one of a plurality of representation formats based on the calculated sparsity, the plurality of representation formats including the dense representation format and at least two types of sparse representation formats;
• an output unit to provide output matrix data information in which the target matrix data is represented in the selected representation format;
• and
• a conversion unit that to acquire three-or-more-dimensional array data, retrieve a plurality of matrix data from the three-or-more-dimensional array data, and generate a plurality pieces of the input matrix data information each of which includes respective ones of the retrieved matrix data,
• wherein the one or more processors are further configured to execute the program instructions to control the acquiring unit to acquire the plurality pieces of the input matrix data information generated by the conversion unit.

A computer-implemented control method comprising:

• acquiring input matrix data information that represents target matrix data in a dense representation format or a sparse representation format, the target matrix data being represented with all data elements of the target matrix data when the target matrix data is represented in the dense representation formats, the target matrix data being represented with non-zero data elements of the target matrix data when the target matrix data is represented in a sparse representation format;
• calculating sparsity of the target matrix data;
• selecting one of a plurality of representation formats based on the calculated sparsity, the plurality of representation formats including the dense representation format and at least two types of sparse representation formats;
• outputting output matrix data information in which the target matrix data is represented by in the selected representation format;
• determining whether or not the calculated sparsity is greater than low-sparsity threshold;
• selecting the dense representation format when it is determined that the calculated sparsity is smaller than the low-sparsity threshold;
• determining whether or not the calculated sparsity is smaller than high-sparsity threshold when it is determined that the calculated sparsity is not smaller than the low-sparsity threshold, the high-sparsity threshold being greater than the low-sparsity threshold;
• selecting a first sparse representation format when it is determined that the calculated sparsity is smaller than the high-sparsity threshold;
• and
• selecting a second sparse representation format when it is determined that the calculated sparsity is not smaller than the high-sparsity threshold,
• wherein the high-sparsity threshold is defined by comparing the number of bits used to represent the target matrix data in the first sparse representation format with the number of bits used to represent the target matrix data in the second sparse representation format, and
• wherein the low-sparsity threshold is defined by comparing the number of bits used to represent the target matrix data in the dense representation format with the number of bits used to represent the target matrix data in the first sparse representation format.

A non-transitory computer-readable medium storing a program that causes a computer to execute:

• acquiring input matrix data information that represents target matrix data in a dense representation format or a sparse representation format, the target matrix data being represented with all data elements of the target matrix data when the target matrix data is represented in the dense representation formats, the target matrix data being represented with non-zero data elements of the target matrix data when the target matrix data is represented in a sparse representation format;
• calculating sparsity of the target matrix data;
• selecting one of a plurality of representation formats based on the calculated sparsity, the plurality of representation formats including the dense representation format and at least two types of sparse representation formats;
• outputting output matrix data information in which the target matrix data is represented by in the selected representation format;
• determining whether or not the calculated sparsity is greater than low-sparsity threshold;
• selecting the dense representation format when it is determined that the calculated sparsity is smaller than the low-sparsity threshold;
• determining whether or not the calculated sparsity is smaller than high-sparsity threshold when it is determined that the calculated sparsity is not smaller than the low-sparsity threshold, the high-sparsity threshold being greater than the low-sparsity threshold;
• selecting a first sparse representation format when it is determined that the calculated sparsity is smaller than the high-sparsity threshold;
• and
• selecting a second sparse representation format when it is determined that the calculated sparsity is not smaller than the high-sparsity threshold,
• wherein the high-sparsity threshold is defined by comparing the number of bits used to represent the target matrix data in the first sparse representation format with the number of bits used to represent the target matrix data in the second sparse representation format, and
• wherein the low-sparsity threshold is defined by comparing the number of bits used to represent the target matrix data in the dense representation format with the number of bits used to represent the target matrix data in the first sparse representation format.

A computer-implemented control method comprising:

• acquiring input matrix data information that represents target matrix data in a dense representation format or a sparse representation format, the target matrix data being represented with all data elements of the target matrix data when the target matrix data is represented in the dense representation formats, the target matrix data being represented with non-zero data elements of the target matrix data when the target matrix data is represented in a sparse representation format;
• calculating sparsity of the target matrix data;
• selecting one of a plurality of representation formats based on the calculated sparsity, the plurality of representation formats including the dense representation format and at least two types of sparse representation formats;
• outputting output matrix data information in which the target matrix data is represented by in the selected representation format;
• acquiring three-or-more-dimensional array data, retrieving a plurality of matrix data from the three-or-more-dimensional array data, and generating a plurality pieces of the input matrix data information each of which includes respective ones of the retrieved matrix data,
• wherein, the acquiring of the input matrix data information acquired the input matrix data information generated from the three-or-more-dimensional array data.

A non-transitory computer-readable medium storing a program that causes a computer to execute:

• acquiring input matrix data information that represents target matrix data in a dense representation format or a sparse representation format, the target matrix data being represented with all data elements of the target matrix data when the target matrix data is represented in the dense representation formats, the target matrix data being represented with non-zero data elements of the target matrix data when the target matrix data is represented in a sparse representation format;
• calculating sparsity of the target matrix data;
• selecting one of a plurality of representation formats based on the calculated sparsity, the plurality of representation formats including the dense representation format and at least two types of sparse representation formats;
• outputting output matrix data information in which the target matrix data is represented by in the selected representation format;
• acquiring three-or-more-dimensional array data, retrieving a plurality of matrix data from the three-or-more-dimensional array data, and generating a plurality pieces of the input matrix data information each of which includes respective ones of the retrieved matrix data,
• wherein, the acquiring of the input matrix data information acquired the input matrix data information generated from the three-or-more-dimensional array data.