[SYSTEMDS-2854] Followup to bugfix in SUM_SQ CUDA codegen

in commit 197a14b

Tested with org.apache.sysds.test.functions.codegen.CellwiseTmplTest.testCodegenCellwiseRewrite9()

Author: corepointer

src/main/cuda/headers/agg_ops.cuh

@@ -92,8 +92,6 @@ struct IdentityOp {
 template<typename T>
 struct SumOp {
 	__device__  __forceinline__ T operator()(T a, T b) const {
-//		 if(blockIdx.x==0 && threadIdx.x ==0)
-//		 	printf("a=%f + b=%f => %f\n", a, b, a+b);
 		return a + b;
 	}
 
@@ -124,16 +122,6 @@ struct MinusOp {
 	}
 };
 
-/**
- * Functor op for sum of squares operation (returns a + b * b)
- */
-template<typename T>
-struct SumSqOp {
-	__device__  __forceinline__ T operator()(T a, T b) const {
-		return a + b * b;
-	}
-};
-
 /**
  * Functor op for min operation
  */

src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeCell.java

@@ -162,8 +162,9 @@ public String codegen(boolean sparse, GeneratorAPI _api) {
 
 		tmp = tmp.replace("%BODY_dense%", tmpDense);
 
-		//return last TMP
-		tmp = tmp.replaceAll("%OUT%", _output.getVarname());
+		//Return last TMP. Square it for CUDA+SUM_SQ
+		tmp = (api.isJava() || _aggOp != AggOp.SUM_SQ) ? tmp.replaceAll("%OUT%", _output.getVarname()) :
+			tmp.replaceAll("%OUT%", _output.getVarname() + " * " + _output.getVarname());
 
 		//replace meta data information
 		tmp = tmp.replaceAll("%TYPE%", getCellType().name());
@@ -181,11 +182,8 @@ public String codegen(boolean sparse, GeneratorAPI _api) {
 			if(_aggOp != null)
 			switch(_aggOp) {
 				case SUM:
-					agg_op = "SumOp";
-					initial_value = "(T)0.0";
-					break;
 				case SUM_SQ:
-					agg_op = "SumSqOp";
+					agg_op = "SumOp";
 					initial_value = "(T)0.0";
 					break;
 				case MIN:

src/test/scripts/functions/codegen/SystemDS-config-codegen6.xml

@@ -27,4 +27,6 @@
 
    <!-- The number of theads for the spark instance artificially selected-->
    <sysds.local.spark.number.threads>16</sysds.local.spark.number.threads>
+
+   <sysds.codegen.api>auto</sysds.codegen.api>
 </root>