debug setupSparseThreadLocalMem

Author: Twolvero

src/main/java/org/apache/sysds/runtime/codegen/LibSpoofPrimitives.java

@@ -2195,7 +2195,7 @@ public static SparseRowVector vectMultWrite(int len, double bval, double[] a, in
 		return vectMultWrite(len, a, bval, aix, ai, alen);
 	}
 
-	//version with branching
+	//old version with branching (not used)
 	public static SparseRowVector vectMultWriteB(int len, double[] a, double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
 		SparseRowVector c = allocSparseVector(Math.min(alen, blen));
 		if( a == null || b == null ) return c;
@@ -2279,7 +2279,7 @@ public static SparseRowVector vectDivWrite(int len, double bval, double[] a, int
 		return c;
 	}
 
-	//version with branching
+	//old version with branching (not used)
 	public static SparseRowVector vectDivWriteB(int len, double[] a, double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
 		SparseRowVector c = allocSparseVector(alen);
 		int aItr = ai;
@@ -3318,31 +3318,6 @@ public static SparseRowVector vectSignWrite(int len, double[] a, int[] aix, int
 		return c;
 	}
 
-	//todo MatrixMult, pow2 and mult2 drafts
-//	public static SparseRowVector vectMatrixMult(int len, double[] a, double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
-//		//note: assumption b is already transposed for efficient dot products
-//		int m2clen = b.length / len;
-//		SparseRowVector c = allocSparseVector(m2clen);
-//		for(int i = 0; i < m2clen; i++) {
-//			c.set(bix[i], LibMatrixMult.dotProduct(a, aix, ai, alen, b, bix, bi, blen));
-//		}
-//		return c;
-//	}
-//
-//	public static SparseRowVector vectPow2Write(int len, double[] a, int[] aix, int ai, int alen) {
-//		SparseRowVector c = allocSparseVector(len);
-//		for(int j = 0; j < ai+alen; j++)
-//			c.set(aix[j], a[j] * a[j]);
-//		return c;
-//	}
-//
-//	public static SparseRowVector vectMult2Write(int len, double[] a, int[] aix, int ai, int alen) {
-//		SparseRowVector c = allocSparseVector(len);
-//		for(int j = 0; j < ai+alen; j++)
-//			c.set(aix[j], a[j] + a[j]);
-//		return c;
-//	}
-
 	//complex builtin functions that are not directly generated
 	//(included here in order to reduce the number of imports)
 

src/main/java/org/apache/sysds/runtime/codegen/SpoofRowwise.java

@@ -189,7 +189,7 @@ public MatrixBlock execute(ArrayList<MatrixBlock> inputs, ArrayList<ScalarObject
 		//setup thread-local memory if necessary
 		if( allocTmp &&_reqVectMem > 0 )
 			if(inputs.get(0).isInSparseFormat() && DMLScript.SPARSE_INTERMEDIATE) {
-				LibSpoofPrimitives.setupSparseThreadLocalMemory(_reqVectMem, n/2, n2);
+				LibSpoofPrimitives.setupSparseThreadLocalMemory(_reqVectMem, n, n2);
 				LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, n, n2);
 			} else {
 				LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, n, n2);
@@ -442,7 +442,12 @@ public DenseBlock call() {
 
 			//allocate vector intermediates and partial output
 			if( _reqVectMem > 0 )
-				LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				if(_a.isInSparseFormat() && DMLScript.SPARSE_INTERMEDIATE) {
+					LibSpoofPrimitives.setupSparseThreadLocalMemory(_reqVectMem, _clen, _clen2);
+					LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				} else {
+					LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				}
 			DenseBlock c = DenseBlockFactory.createDenseBlock(1, _outLen);
 
 			if( !_a.isInSparseFormat() )
@@ -451,7 +456,12 @@ public DenseBlock call() {
 				executeSparse(_a.getSparseBlock(), _b, _scalars, c, _clen, _rl, _ru, 0);
 
 			if( _reqVectMem > 0 )
-				LibSpoofPrimitives.cleanupThreadLocalMemory();
+				if(_a.isInSparseFormat() && DMLScript.SPARSE_INTERMEDIATE) {
+					LibSpoofPrimitives.cleanupSparseThreadLocalMemory();
+					LibSpoofPrimitives.cleanupThreadLocalMemory();
+				} else {
+					LibSpoofPrimitives.cleanupThreadLocalMemory();
+				}
 			return c;
 		}
 	}
@@ -485,15 +495,25 @@ protected ParExecTask( MatrixBlock a, SideInput[] b, MatrixBlock c, double[] sca
 		public Long call() {
 			//allocate vector intermediates
 			if( _reqVectMem > 0 )
-				LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				if(_a.isInSparseFormat() && DMLScript.SPARSE_INTERMEDIATE) {
+					LibSpoofPrimitives.setupSparseThreadLocalMemory(_reqVectMem, _clen, _clen2);
+					LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				} else {
+					LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, _clen, _clen2);
+				}
 
 			if( !_a.isInSparseFormat() )
 				executeDense(_a.getDenseBlock(), _b, _scalars, _c.getDenseBlock(), _clen, _rl, _ru, 0);
 			else
 				executeSparse(_a.getSparseBlock(), _b, _scalars, _c.getDenseBlock(), _clen, _rl, _ru, 0);
-			
+
 			if( _reqVectMem > 0 )
-				LibSpoofPrimitives.cleanupThreadLocalMemory();
+				if(_a.isInSparseFormat() && DMLScript.SPARSE_INTERMEDIATE) {
+					LibSpoofPrimitives.cleanupSparseThreadLocalMemory();
+					LibSpoofPrimitives.cleanupThreadLocalMemory();
+				} else {
+					LibSpoofPrimitives.cleanupThreadLocalMemory();
+				}
 
 			//maintain nnz for row partition
 			return _c.recomputeNonZeros(_rl, _ru-1, 0, _c.getNumColumns()-1);