Electronic device and method with compressed storage format conversion
An electronic device includes a host processor configured to: convert a sparse matrix compressed and expressed in a first compressed format into a second compressed storage format, based on a feature of the sparse matrix; preprocess a vector based on the second compressed storage format; and transmit the sparse matrix converted into the second compressed storage format and the preprocessed vector to a computing device; and the computing device configured to multiply the sparse matrix converted into the second compressed storage format by the preprocessed vector.
1. An electronic device comprising:
a host processor configured to:
convert a sparse matrix compressed and expressed in a first compressed format into a second compressed storage format, based on a feature of the sparse matrix, wherein the second compressed storage format is determined by inputting the feature extracted from the sparse matrix to a machine learning model;
preprocess a vector based on the second compressed storage format;
transmit the sparse matrix converted into the second compressed storage format and the preprocessed vector to a computing device; and
the computing device configured to multiply the sparse matrix converted into the second compressed storage format by the preprocessed vector.
2. The electronic device of claim 1 , wherein, for the converting of the sparse matrix into the second compressed storage format, the host processor is further configured to:
extract the feature from the sparse matrix;
output candidate compressed storage formats by inputting the feature to the machine learning model; and
determine the second compressed storage format among the candidate compressed storage formats.
3. The electronic device of claim 2 , wherein the feature comprises:
a sparsity rate referring to a rate of 0 values comprised by the sparse matrix; and
a sparsity pattern referring to a pattern of the 0 values comprised by the sparse matrix.
4. The electronic device of claim 2 , wherein, for the converting of the sparse matrix into the second compressed storage format, the host processor is further configured to convert the first compressed storage format of the sparse matrix into the second compressed storage format in response to the second compressed storage format being different from the first compressed storage format of the sparse matrix.
5. The electronic device of claim 2 , wherein, for the determining of the second compressed storage format, the host processor is further configured to determine one of the candidate compressed storage formats to be the second compressed storage format, based on a number of cycles that is a number of operations performed by the computing device in multiplying the sparse matrix converted into a candidate compressed storage format by the vector.
6. The electronic device of claim 1 , wherein, for the preprocessing of the vector, the host processor is further configured to extract and align elements of the vector to be used for the multiplication based on the second compressed storage format.
7. The electronic device of claim 1 , wherein, for the converting of the sparse matrix into the second compressed storage format, the host processor is further configured to:
divide the sparse matrix into a plurality of blocks; and
convert the sparse matrix compressed and expressed in the first compressed storage format into the second compressed storage format by converting a compressed storage format of the plurality of blocks.
8. The electronic device of claim 7 , wherein, for the converting the sparse matrix into the second compressed storage format, the host processor is further configured to:
extract the feature of the sparse matrix;
output a candidate compressed storage format for each of the plurality of blocks by inputting the feature to the machine learning model; and
determine the second compressed storage format among a plurality of combinations of the candidate compressed storage formats of the plurality of blocks.
9. The electronic device of claim 8 , wherein the feature comprises a sparsity rate referring to a rate of 0 values comprised by the sparse matrix and a sparsity pattern referring to a pattern of the 0 values comprised by the sparse matrix.
10. The electronic device of claim 8 , wherein the host processor is further configured to:
determine the number of cycles that is a number of operations performed by the computing device in multiplying a block in a candidate compressed storage format comprised in the plurality of combinations by the vector; and
determine one of the plurality of combinations to be the second compressed storage format based on the number of the cycles.
11. The electronic device of claim 7 , wherein, for the preprocessing of the vector, the host processor is further configured to extract and align elements of the vector to be used for the multiplication for each of the plurality of blocks, based on a converted compressed storage format of the plurality of blocks comprised in the sparse matrix converted into the second compressed storage format.
12. A processor-implemented method comprising:
converting a sparse matrix compressed and expressed in a first compressed storage format into a second compressed storage format, based on a feature of the sparse matrix, wherein the second compressed storage format is determined by inputting the feature extracted from the sparse matrix to a machine learning model;
preprocessing a vector based on the second compressed storage format; and
transmitting the sparse matrix converted into the second compressed storage format and the preprocessed vector to a computing device,
wherein the computing device is configured to multiply the sparse matrix converted into the second compressed storage format by the preprocessed vector.
13. The method of claim 12 , wherein the converting of the sparse matrix into the second compressed storage format comprises:
extracting the feature from the sparse matrix;
outputting candidate compressed storage formats by inputting the feature to the machine learning model; and
determining the second compressed storage format among the candidate compressed storage formats.
14. The method of claim 13 , wherein the feature comprises:
a sparsity rate referring to a rate of 0 values comprised by the sparse matrix; and
a sparsity pattern referring to a pattern of the 0 values comprised by the sparse matrix.
15. The method of claim 13 , wherein the converting of the sparse matrix into the second compressed storage format further comprises converting the first compressed storage format of the sparse matrix into the second compressed storage format in response to the second compressed storage format being different from the first compressed storage format of the sparse matrix.
16. The method of claim 13 , wherein the determining of the second compressed storage format comprises determining one of the candidate compressed storage formats to be the second compressed storage format, based on a number of cycles that is a number of operations performed by the computing device in multiplying the sparse matrix converted into a candidate compressed storage format with the vector.
17. The method of claim 12 , wherein the preprocessing the vector comprises extracting and arranging elements of the vector to be used for the multiplication based on the second compressed storage format.
18. The method of claim 12 , wherein the converting of the sparse matrix into the second compressed storage format further comprises:
dividing the sparse matrix into a plurality of blocks; and
converting the sparse matrix compressed and expressed in the first compressed storage format into the second compressed storage format by converting a compressed storage format of the plurality of blocks.
19. The method of claim 18 , wherein the converting the sparse matrix into the second compressed storage format further comprises:
extracting the feature of the sparse matrix;
outputting a candidate compressed storage format for each of the plurality of blocks by inputting the feature to the machine learning model; and
determining the second compressed storage format among a plurality of combinations of the candidate compressed storage formats of the plurality of blocks,
wherein the feature comprises a sparsity rate referring to a rate of 0 values comprised by the sparse matrix and a sparsity pattern referring to a pattern of the 0 values comprised by the sparse matrix.
20. The method of claim 19 , wherein the determining the second compressed storage format further comprises:
determining the number of cycles that is a number of operations performed by the computing device in multiplying a block in a candidate compressed storage format comprised in the plurality of combinations by the vector; and
determining one of the plurality of combinations to be the second compressed storage format based on the number of the cycles.