[systemml] branch master updated: [SYSTEMML-540] Reduce the memory pressure of CP lstm_backward instruction

[ni...@apache.org]

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

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


The following commit(s) were added to refs/heads/master by this push:
     new c7b9745  [SYSTEMML-540] Reduce the memory pressure of CP lstm_backward instruction
c7b9745 is described below

commit c7b9745800e0c71f0c6c76b8284c78e33a5cdb01
Author: Niketan Pansare <np...@us.ibm.com>
AuthorDate: Tue Mar 5 15:48:09 2019 -0800

    [SYSTEMML-540] Reduce the memory pressure of CP lstm_backward instruction
    
    - When lstm_backward in invoked, this commit avoids memory allocation and left indexing of output and carry activations of the corresponding forward invocation.
---
 .../runtime/instructions/cp/DnnCPInstruction.java    |  6 +++---
 .../sysml/runtime/matrix/data/LibMatrixDNN.java      | 20 +++++++++-----------
 2 files changed, 12 insertions(+), 14 deletions(-)

diff --git a/src/main/java/org/apache/sysml/runtime/instructions/cp/DnnCPInstruction.java b/src/main/java/org/apache/sysml/runtime/instructions/cp/DnnCPInstruction.java
index 50a11de..35ac5b6 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/cp/DnnCPInstruction.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/cp/DnnCPInstruction.java
@@ -388,8 +388,6 @@ public class DnnCPInstruction extends UnaryCPInstruction {
 					+ "but found [" + dc.getNumRows() + "," + dc.getNumColumns() + "]");
 		}
 		
-		MatrixBlock out = new MatrixBlock(N, return_seq ? (T*M) : M, false);
-		MatrixBlock c = new MatrixBlock(N, M, false);
 		MatrixBlock cache_out = new MatrixBlock(T, N*M, false);
 		MatrixBlock cache_c = new MatrixBlock(T, N*M, false);
 		MatrixBlock cache_ifog = new MatrixBlock(T, N*4*M, false);
@@ -401,7 +399,9 @@ public class DnnCPInstruction extends UnaryCPInstruction {
 		cache_ifog.allocateDenseBlock();
 		LibMatrixDNN.lstm(X, W, b, out0, c0, 
 				return_seq, N, T, D, M,
-				out,  c, cache_out, cache_c, cache_ifog,
+				// Avoid out and c computation in lstm forward call
+				null, null, 
+				cache_out, cache_c, cache_ifog,
 				_numThreads);
 		
 		MatrixBlock dX = new MatrixBlock(N, T*D, false);
diff --git a/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java b/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
index 0005932..3ec9fb3 100644
--- a/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
+++ b/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
@@ -20,7 +20,6 @@ package org.apache.sysml.runtime.matrix.data;
 
 import java.util.ArrayList;
 import java.util.Arrays;
-import java.util.HashSet;
 import java.util.List;
 import java.util.concurrent.Callable;
 import java.util.concurrent.ExecutorService;
@@ -41,7 +40,6 @@ import org.apache.sysml.runtime.functionobjects.Multiply;
 import org.apache.sysml.runtime.functionobjects.Plus;
 import org.apache.sysml.runtime.functionobjects.PlusMultiply;
 import org.apache.sysml.runtime.functionobjects.Power;
-import org.apache.sysml.runtime.functionobjects.Power2;
 import org.apache.sysml.runtime.functionobjects.SwapIndex;
 import org.apache.sysml.runtime.functionobjects.Builtin.BuiltinCode;
 import org.apache.sysml.runtime.instructions.cp.KahanObject;
@@ -57,8 +55,6 @@ import org.apache.sysml.runtime.matrix.operators.UnaryOperator;
 import org.apache.sysml.runtime.util.CommonThreadPool;
 import org.apache.sysml.runtime.util.DnnUtils;
 
-import com.sun.org.apache.xpath.internal.operations.Minus;
-
 /*
  * This class allows users to invoke deep learning related operations 
  * (such as conv2d, conv2d_backward_data, conv2d_backward_filter, maxpooling, maxpooling_backward, bias_add)
@@ -514,7 +510,7 @@ public class LibMatrixDNN {
 	
 	public static void lstm(MatrixBlock X, MatrixBlock W, MatrixBlock b, MatrixBlock out0, MatrixBlock c0, 
 			boolean return_seq, int N, int T, int D, int M,
-			MatrixBlock out, MatrixBlock c, // output 
+			MatrixBlock out, MatrixBlock c, // output: if null, the output and c are not passed back
 			MatrixBlock cache_out, MatrixBlock cache_c, MatrixBlock cache_ifog, // if null, the cache values are not computed
 			int numThreads) {
 		MatrixBlock out_prev = out0;
@@ -624,7 +620,7 @@ public class LibMatrixDNN {
 				updateIfogCache(cache_ifog, ifo, g, t, N, M);
 			}
 			
-			if(return_seq) {
+			if(return_seq && out != null) {
 				out = out.leftIndexingOperations(out_t, 0, N-1, (t-1)*M, t*M-1, out, UpdateType.INPLACE);
 			}
 			out_prev = out_t;
@@ -635,12 +631,14 @@ public class LibMatrixDNN {
 				reshapeAsRowMatrixAndLeftIndex(cache_c, c_t, t-1, N*M);
 			}
 		}
-		if(out_t != null && !return_seq)
+		if(out_t != null && !return_seq && out != null)
 			out.copy(out_t);
-		if(c_t != null)
-			c.copy(c_t);
-		else
-			c.copy(c0);
+		if(c != null) {
+			if(c_t != null)
+				c.copy(c_t);
+			else
+				c.copy(c0);
+		}
 		if(cache_out != null) {
 			cache_out.recomputeNonZeros();
 			cache_c.recomputeNonZeros();