[ https://issues.apache.org/jira/browse/SYSTEMML-824?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Niketan Pansare updated SYSTEMML-824: ------------------------------------- Description: The cellwise (matrix-matrix as well as matrix-scalar) operations take significant amount of time while training Lenet. Here are few ways to improve the performance of cell-wise operations: 1. Inject in-place updates [1] (saving on zero-ing out the matrix). 2. Fused cell-wise operations (as an example, recently added axpy operations: https://github.com/apache/incubator-systemml/commit/b584aecf6b3a1eb96ff83b78cc3ad7c7c6d15baa). 3. Parallelize cellwise operations (initial investigation need to be conducted before proceeding in this direction especially in sparse case: https://github.com/apache/incubator-systemml/blob/master/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixBincell.java#L274). [~nakul02] [~mwdus...@us.ibm.com] [~prithvi_r_s] [~mboehm7] [~reinwald] Reference: [1] http://www.diku.dk/hjemmesider/ansatte/torbenm/ICD/Register.pdf was: The cellwise (matrix-matrix as well as matrix-scalar) operations take significant amount of time while training Lenet. Here are few ways to improve the performance of cell-wise operations: 1. Inject in-place updates [1] (saving on zero-ing out the matrix). 2. Fused cell-wise operations (as an example, recently added axpy operations). 3. Parallelize cellwise operations (initial investigation need to be conducted before proceeding in this direction especially in sparse case: https://github.com/apache/incubator-systemml/blob/master/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixBincell.java#L274). [~nakul02] [~mwdus...@us.ibm.com] [~prithvi_r_s] [~mboehm7] [~reinwald] Reference: [1] http://www.diku.dk/hjemmesider/ansatte/torbenm/ICD/Register.pdf > Improve the performance of binary cell-wise operations > ------------------------------------------------------ > > Key: SYSTEMML-824 > URL: https://issues.apache.org/jira/browse/SYSTEMML-824 > Project: SystemML > Issue Type: Task > Reporter: Niketan Pansare > > The cellwise (matrix-matrix as well as matrix-scalar) operations take > significant amount of time while training Lenet. Here are few ways to improve > the performance of cell-wise operations: > 1. Inject in-place updates [1] (saving on zero-ing out the matrix). > 2. Fused cell-wise operations (as an example, recently added axpy operations: > https://github.com/apache/incubator-systemml/commit/b584aecf6b3a1eb96ff83b78cc3ad7c7c6d15baa). > > 3. Parallelize cellwise operations (initial investigation need to be > conducted before proceeding in this direction especially in sparse case: > https://github.com/apache/incubator-systemml/blob/master/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixBincell.java#L274). > [~nakul02] [~mwdus...@us.ibm.com] [~prithvi_r_s] [~mboehm7] [~reinwald] > Reference: > [1] http://www.diku.dk/hjemmesider/ansatte/torbenm/ICD/Register.pdf -- This message was sent by Atlassian JIRA (v6.3.4#6332)