[systemds] branch master updated: [SYSTEMDS-3109] Performance sparse binary matrix-vector operations

[mboehm7]

This is an automated email from the ASF dual-hosted git repository.

mboehm7 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/systemds.git


The following commit(s) were added to refs/heads/master by this push:
     new 9e01614  [SYSTEMDS-3109] Performance sparse binary matrix-vector 
operations
9e01614 is described below

commit 9e016141264b3a83eedaf2b03c3e6463f242f272
Author: Matthias Boehm <mboe...@gmail.com>
AuthorDate: Sun Aug 29 21:14:36 2021 +0200

    [SYSTEMDS-3109] Performance sparse binary matrix-vector operations
    
    This patch makes several smaller improvements to sparse binary
    matrix-vector operations, especially for sparse-safe operations like
    multiply and division. This includes skipping zero values for division,
    and pre-allocating sparse output rows. On a scenario of 100K x 100 and
    50 operations, the dense X/rowVect required 11.3s while sparse (with
    sparsity 0.1) required 33.6s. With skipping it went down to 11.9s and
    with skipping and pre-allocation to 8.7s.
    
    Thanks to Sebastian Baunsgaard for catching this performance issue.
---
 .../java/org/apache/sysds/runtime/matrix/data/LibMatrixBincell.java  | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git 
a/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixBincell.java 
b/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixBincell.java
index 2fde00d..6d53ff7 100644
--- a/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixBincell.java
+++ b/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixBincell.java
@@ -605,7 +605,7 @@ public class LibMatrixBincell {
        
        private static void safeBinaryMVSparse(MatrixBlock m1, MatrixBlock m2, 
MatrixBlock ret, BinaryOperator op) {
                boolean isMultiply = (op.fn instanceof Multiply);
-               boolean skipEmpty = (isMultiply);
+               boolean skipEmpty = (isMultiply || isSparseSafeDivide(op, m2));
                
                int rlen = m1.rlen;
                int clen = m1.clen;
@@ -661,7 +661,8 @@ public class LibMatrixBincell {
                        for( int i=0; i<rlen; i++ ) {
                                if( skipEmpty && (a==null || a.isEmpty(i)) )
                                        continue; //skip empty rows
-                                       
+                               if( skipEmpty && ret.sparse )
+                                       ret.sparseBlock.allocate(i, a.size(i));
                                int lastIx = -1;
                                if( a!=null && !a.isEmpty(i) ) {
                                        int apos = a.pos(i);