[SYSTEMDS-3836] Fix distributive binary ops rewrite for broadcasting

The distributive binary operation rewrite, transforms the pattern
X-X*Y into (1-Y)*X but was so far not aware of broadcasting semantics.
If Y is a row or column vector but X a matrix, the rewrite yields
mismatching dimension exceptions during runtime. We now simply rewrite
the pattern to X*(1-Y) if Y is indeed a vector and X is not. Always
rewriting the pattern to the latter cause the mmchain rewrite to
no longer trigger (which is crucial for many end-to-end algorithms).

The tests have, however, also shown that for the multiLogReg test
we are not compiling mmchain (independent of this rewrite change)
something that needs fixing before the release.

Author: mboehm7

src/main/java/org/apache/sysds/hops/rewrite/RewriteAlgebraicSimplificationStatic.java

@@ -886,10 +886,14 @@ private static Hop simplifyDistributiveBinaryOperation( Hop parent, Hop hi, int
 						X = right;
 						Y = ( right == leftC1 ) ? leftC2 : leftC1;
 					}
-					if( X != null ){ //rewrite 'binary +/-' 
+					if( X != null && Y.dimsKnown() ){ //rewrite 'binary +/-' 
 						LiteralOp literal = new LiteralOp(1);
 						BinaryOp plus = HopRewriteUtils.createBinary(Y, literal, bop.getOp());
-						BinaryOp mult = HopRewriteUtils.createBinary(plus, X, OpOp2.MULT);
+						
+						BinaryOp mult = (plus.getDim1()==1 || plus.getDim2() == 1)
+								&& (X.getDim1()>1 && X.getDim2()>1) ?
+							HopRewriteUtils.createBinary(X, plus, OpOp2.MULT) :
+							HopRewriteUtils.createBinary(plus, X, OpOp2.MULT);
 						HopRewriteUtils.replaceChildReference(parent, hi, mult, pos);
 						HopRewriteUtils.cleanupUnreferenced(hi, left);
 						hi = mult;
@@ -908,10 +912,13 @@ private static Hop simplifyDistributiveBinaryOperation( Hop parent, Hop hi, int
 						X = left;
 						Y = ( left == rightC1 ) ? rightC2 : rightC1;
 					}
-					if( X != null ){ //rewrite '+/- binary'
+					if( X != null && Y.dimsKnown() ){ //rewrite '+/- binary'
 						LiteralOp literal = new LiteralOp(1);
 						BinaryOp plus = HopRewriteUtils.createBinary(literal, Y, bop.getOp());
-						BinaryOp mult = HopRewriteUtils.createBinary(plus, X, OpOp2.MULT);
+						BinaryOp mult = (plus.getDim1()==1 || plus.getDim2() == 1) 
+								&& (X.getDim1()>1 && X.getDim2()>1) ?
+							HopRewriteUtils.createBinary(X, plus, OpOp2.MULT) :
+							HopRewriteUtils.createBinary(plus, X, OpOp2.MULT);
 						HopRewriteUtils.replaceChildReference(parent, hi, mult, pos);
 						HopRewriteUtils.cleanupUnreferenced(hi, right);
 						hi = mult;

src/test/java/org/apache/sysds/test/functions/rewrite/RewriteSimplifyDistributiveBinaryOperationTest.java

@@ -86,6 +86,11 @@ public void testDistrBinaryOpMultAddNoRewrite() {
 	public void testDistrBinaryOpMultAddRewrite() {
 		testSimplifyDistributiveBinaryOperation(4, true);    //pattern: (Y*X+X) -> (Y+1)*X
 	}
+	
+	@Test
+	public void testDistrBinaryOpMultMinusVectorRewrite() {
+		testSimplifyDistributiveBinaryOperation(5, true);    //pattern: (X*Y-X) -> (Y+1)*X, Y vector
+	}
 
 	private void testSimplifyDistributiveBinaryOperation(int ID, boolean rewrites) {
 		boolean oldFlag1 = OptimizerUtils.ALLOW_ALGEBRAIC_SIMPLIFICATION;
@@ -104,7 +109,7 @@ private void testSimplifyDistributiveBinaryOperation(int ID, boolean rewrites) {
 
 			//create matrices
 			double[][] X = getRandomMatrix(rows, cols, -1, 1, 0.60d, 3);
-			double[][] Y = getRandomMatrix(rows, cols, -1, 1, 0.60d, 5);
+			double[][] Y = getRandomMatrix(rows, ID==5?1:cols, -1, 1, 0.60d, 5);
 			writeInputMatrixWithMTD("X", X, true);
 			writeInputMatrixWithMTD("Y", Y, true);
 

src/test/scripts/functions/rewrite/RewriteSimplifyDistributiveBinaryOperation.R

@@ -44,6 +44,8 @@ if( type == 1 ) {
     R = (X+Y*X)
 } else if( type == 4 ) {
     R = (Y*X+X)
+} else if( type == 5 ) {
+    R = (X*(Y%*%matrix(1,1,ncol(X)))-X) * 1
 }
 
 

src/test/scripts/functions/rewrite/RewriteSimplifyDistributiveBinaryOperation.dml

@@ -38,6 +38,9 @@ else if( type == 3 ) {
 else if( type == 4 ) {
     R = (Y*X+X) * 1
 }
+else if( type == 5 ) {
+    R = (X*Y-X) * 1
+}
 
 
 # Write the result matrix R