This is an automated email from the ASF dual-hosted git repository.
mboehm7 pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/systemds.git
The following commit(s) were added to refs/heads/main by this push:
new b1c5d64d78 [SYSTEMDS-3860] Extended sparsity exploitation in codegen
row templates
b1c5d64d78 is described below
commit b1c5d64d7884f3c63284bbfea6003d13faec4a33
Author: Frxms <[email protected]>
AuthorDate: Wed Aug 20 14:00:30 2025 +0200
[SYSTEMDS-3860] Extended sparsity exploitation in codegen row templates
Finalized runtime kernels, code generation, and optimization
Closes #2297.
Closes #2277.
Closes #2276.
---
src/main/java/org/apache/sysds/api/DMLOptions.java | 15 +-
src/main/java/org/apache/sysds/api/DMLScript.java | 4 +
.../apache/sysds/hops/codegen/SpoofCompiler.java | 2 +-
.../org/apache/sysds/hops/codegen/cplan/CNode.java | 20 +-
.../sysds/hops/codegen/cplan/CNodeBinary.java | 197 +++-
.../apache/sysds/hops/codegen/cplan/CNodeNary.java | 5 +-
.../apache/sysds/hops/codegen/cplan/CNodeRow.java | 4 +-
.../sysds/hops/codegen/cplan/CNodeTernary.java | 7 +-
.../sysds/hops/codegen/cplan/CNodeUnary.java | 65 +-
.../sysds/hops/codegen/cplan/CodeTemplate.java | 18 +-
.../sysds/hops/codegen/cplan/java/Binary.java | 62 +-
.../sysds/hops/codegen/cplan/java/Ternary.java | 2 +-
.../sysds/hops/codegen/cplan/java/Unary.java | 25 +-
.../sysds/hops/codegen/template/TemplateRow.java | 26 +-
.../sysds/runtime/codegen/LibSpoofPrimitives.java | 1226 ++++++++++++++++++++
.../apache/sysds/runtime/codegen/SpoofRowwise.java | 45 +-
.../codegen/CPlanVectorPrimitivesTest.java | 461 ++++++++
.../component/codegen/SparseVectorAllocTest.java | 133 +++
.../test/functions/codegen/RowAggTmplTest.java | 12 +-
.../scripts/functions/codegen/rowAggPattern49.R | 54 +
.../scripts/functions/codegen/rowAggPattern49.dml | 52 +
.../scripts/functions/codegen/rowAggPattern50.R | 43 +
.../scripts/functions/codegen/rowAggPattern50.dml | 40 +
23 files changed, 2436 insertions(+), 82 deletions(-)
diff --git a/src/main/java/org/apache/sysds/api/DMLOptions.java
b/src/main/java/org/apache/sysds/api/DMLOptions.java
index 97d5f54a4a..917aecc4ab 100644
--- a/src/main/java/org/apache/sysds/api/DMLOptions.java
+++ b/src/main/java/org/apache/sysds/api/DMLOptions.java
@@ -86,6 +86,7 @@ public class DMLOptions {
public boolean federatedCompilation = false; //
Compile federated instructions based on input federation state and privacy
constraints.
public boolean noFedRuntimeConversion = false; // If
activated, no runtime conversion of CP instructions to FED instructions will be
performed.
public int seed = -1; // The
general seed for the execution, if -1 random (system time).
+ public boolean sparseIntermediate = false;
// whether SparseRowIntermediates should be used for rowwise operations
public final static DMLOptions defaultOptions = new DMLOptions(null);
@@ -119,7 +120,8 @@ public class DMLOptions {
", w=" + fedWorker +
", federatedCompilation=" + federatedCompilation +
", noFedRuntimeConversion=" + noFedRuntimeConversion +
- ", seed=" + seed +
+ ", seed=" + seed +
+ ", sparseIntermediate=" + sparseIntermediate +
'}';
}
@@ -353,6 +355,11 @@ public class DMLOptions {
dmlOptions.seed =
Integer.parseInt(line.getOptionValue("seed"));
}
+ //TODO move to systemds-config instead of command-line arg
+ if(line.hasOption("sparseIntermediate")){
+ dmlOptions.sparseIntermediate = true;
+ }
+
return dmlOptions;
}
@@ -436,7 +443,10 @@ public class DMLOptions {
Option commandlineSeed = OptionBuilder
.withDescription("A general seed for the execution
through the commandline")
.hasArg().create("seed");
-
+ Option sparseRowIntermediates = OptionBuilder
+ .withDescription("If activated, sparseRowVector
intermediates will be used to calculate rowwise operations.")
+ .create("sparseIntermediate");
+
options.addOption(configOpt);
options.addOption(cleanOpt);
options.addOption(statsOpt);
@@ -457,6 +467,7 @@ public class DMLOptions {
options.addOption(federatedCompilation);
options.addOption(noFedRuntimeConversion);
options.addOption(commandlineSeed);
+ options.addOption(sparseRowIntermediates);
// Either a clean(-clean), a file(-f), a script(-s) or
help(-help) needs to be specified
OptionGroup fileOrScriptOpt = new OptionGroup()
diff --git a/src/main/java/org/apache/sysds/api/DMLScript.java
b/src/main/java/org/apache/sysds/api/DMLScript.java
index 2bc8d3b816..65805b5c2e 100644
--- a/src/main/java/org/apache/sysds/api/DMLScript.java
+++ b/src/main/java/org/apache/sysds/api/DMLScript.java
@@ -155,6 +155,9 @@ public class DMLScript
// Global seed
public static int SEED = -1;
+ // Sparse row flag
+ public static boolean SPARSE_INTERMEDIATE = false;
+
public static String MONITORING_ADDRESS = null;
// flag that indicates whether or not to suppress any prints to stdout
@@ -278,6 +281,7 @@ public class DMLScript
LINEAGE_ESTIMATE = dmlOptions.lineage_estimate;
LINEAGE_DEBUGGER = dmlOptions.lineage_debugger;
SEED = dmlOptions.seed;
+ SPARSE_INTERMEDIATE =
dmlOptions.sparseIntermediate;
String fnameOptConfig = dmlOptions.configFile;
diff --git a/src/main/java/org/apache/sysds/hops/codegen/SpoofCompiler.java
b/src/main/java/org/apache/sysds/hops/codegen/SpoofCompiler.java
index 34329ca64d..307205dbc8 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/SpoofCompiler.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/SpoofCompiler.java
@@ -470,7 +470,7 @@ public class SpoofCompiler {
* @param recompile true if invoked during dynamic recompilation
* @return dag root nodes of modified dag
*/
- public static ArrayList<Hop> optimize(ArrayList<Hop> roots, boolean
recompile)
+ public static ArrayList<Hop> optimize(ArrayList<Hop> roots, boolean
recompile)
{
if( roots == null || roots.isEmpty() )
return roots;
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNode.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNode.java
index 36cc8f4979..36ebd238ac 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNode.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNode.java
@@ -19,6 +19,7 @@
package org.apache.sysds.hops.codegen.cplan;
+import org.apache.sysds.api.DMLScript;
import org.apache.sysds.common.Types.DataType;
import org.apache.sysds.hops.codegen.SpoofCompiler.GeneratorAPI;
import org.apache.sysds.hops.codegen.template.TemplateUtils;
@@ -77,6 +78,14 @@ public abstract class CNode
_genVar = "TMP"+_seqVar.getNextID();
return _genVar;
}
+
+ public String createVarname(boolean sparse) {
+ if(!sparse) {
+ return createVarname();
+ } else {
+ return _genVar = "S" + createVarname();
+ }
+ }
public String getVarname() {
return _genVar;
@@ -98,6 +107,8 @@ public abstract class CNode
return "len";
if(getVarname().startsWith("b"))
return getVarname() + ".clen";
+ else if(getVarname().startsWith("STMP"))
+ return "len";
else if(_dataType == DataType.MATRIX)
return getVarname() + ".length";
}
@@ -222,8 +233,13 @@ public abstract class CNode
protected String replaceUnaryPlaceholders(String tmp, String varj,
boolean vectIn, GeneratorAPI api) {
//replace sparse and dense inputs
- tmp = tmp.replace("%IN1v%", varj+"vals");
- tmp = tmp.replace("%IN1i%", varj+"ix");
+ if(DMLScript.SPARSE_INTERMEDIATE) {
+ tmp = tmp.replace("%IN1v%", varj.startsWith("STMP") ?
varj+".values()" : varj+"vals");
+ tmp = tmp.replace("%IN1i%", varj.startsWith("STMP") ?
varj+".indexes()" :varj+"ix");
+ } else {
+ tmp = tmp.replace("%IN1v%", varj+"vals");
+ tmp = tmp.replace("%IN1i%", varj+"ix");
+ }
tmp = tmp.replace("%IN1%",
(vectIn && TemplateUtils.isMatrix(_inputs.get(0))) ?
((api == GeneratorAPI.JAVA) ? varj +
".values(rix)" : varj + ".vals(0)" ) :
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeBinary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeBinary.java
index b29d586c38..6031d8492a 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeBinary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeBinary.java
@@ -22,6 +22,7 @@ package org.apache.sysds.hops.codegen.cplan;
import java.util.Arrays;
import org.apache.commons.lang3.StringUtils;
+import org.apache.sysds.api.DMLScript;
import org.apache.sysds.common.Opcodes;
import org.apache.sysds.hops.codegen.template.TemplateUtils;
import org.apache.sysds.common.Types.DataType;
@@ -126,7 +127,8 @@ public class CNodeBinary extends CNode {
}
private final BinType _type;
-
+ private boolean sparseTemplate;
+
public CNodeBinary( CNode in1, CNode in2, BinType type ) {
//canonicalize commutative matrix-scalar operations
//to increase reuse potential
@@ -143,6 +145,23 @@ public class CNodeBinary extends CNode {
setOutputDims();
}
+ public CNodeBinary( CNode in1, CNode in2, BinType type, double
sparsityEst, double scalarVal ) {
+ //canonicalize commutative matrix-scalar operations
+ //to increase reuse potential
+ if( type.isCommutative() && in1 instanceof CNodeData
+ && in1.getDataType()==DataType.SCALAR ) {
+ CNode tmp = in1;
+ in1 = in2;
+ in2 = tmp;
+ }
+
+ _inputs.add(in1);
+ _inputs.add(in2);
+ _type = type;
+ setOutputDims();
+ sparseTemplate = getTemplateType(sparsityEst, scalarVal);
+ }
+
public BinType getType() {
return _type;
}
@@ -157,60 +176,63 @@ public class CNodeBinary extends CNode {
//generate children
sb.append(_inputs.get(0).codegen(sparse, api));
sb.append(_inputs.get(1).codegen(sparse, api));
-
+
//generate binary operation (use sparse template, if data input)
- boolean lsparseLhs = sparse && _inputs.get(0) instanceof
CNodeData
- && _inputs.get(0).getVarname().startsWith("a");
- boolean lsparseRhs = sparse && _inputs.get(1) instanceof
CNodeData
- && _inputs.get(1).getVarname().startsWith("a");
+ boolean lsparseLhs = sparse ? _inputs.get(0) instanceof
CNodeData
+ && _inputs.get(0).getVarname().startsWith("a") ||
+ _inputs.get(0).getVarname().startsWith("STMP") : false;
+ boolean lsparseRhs = sparse ? _inputs.get(1) instanceof
CNodeData
+ && _inputs.get(1).getVarname().startsWith("a") ||
+ _inputs.get(1).getVarname().startsWith("STMP") : false;
boolean scalarInput = _inputs.get(0).getDataType().isScalar();
boolean scalarVector = (_inputs.get(0).getDataType().isScalar()
&& _inputs.get(1).getDataType().isMatrix());
boolean vectorVector = _inputs.get(0).getDataType().isMatrix()
&& _inputs.get(1).getDataType().isMatrix();
- String var = createVarname();
+ String var = createVarname(sparse && sparseTemplate &&
getOutputType(scalarVector, lsparseLhs, lsparseRhs));
String tmp = getLanguageTemplateClass(this, api)
- .getTemplate(_type, lsparseLhs, lsparseRhs,
scalarVector, scalarInput, vectorVector);
+ .getTemplate(_type, lsparseLhs, lsparseRhs,
scalarVector, scalarInput, vectorVector, sparseTemplate);
tmp = tmp.replace("%TMP%", var);
-
+
//replace input references and start indexes
for( int j=0; j<2; j++ ) {
String varj = _inputs.get(j).getVarname(api);
-
//replace sparse and dense inputs
- tmp = tmp.replace("%IN"+(j+1)+"v%", varj+"vals");
- tmp = tmp.replace("%IN"+(j+1)+"i%", varj+"ix");
+ tmp = tmp.replace("%IN"+(j+1)+"v%",
varj.startsWith("STMP") ? varj+".values()" : varj+"vals");
+ tmp = tmp.replace("%IN"+(j+1)+"i%",
varj.startsWith("STMP") ? varj+".indexes()" : varj+"ix");
tmp = tmp.replace("%IN"+(j+1)+"%",
- varj.startsWith("a") ? (api ==
GeneratorAPI.JAVA ? varj :
- (_inputs.get(j).getDataType()
== DataType.MATRIX ? varj + ".vals(0)" : varj)) :
- varj.startsWith("b") ? (api ==
GeneratorAPI.JAVA ? varj + ".values(rix)" :
- (_type ==
BinType.VECT_MATRIXMULT ? varj : varj + ".vals(0)")) :
-
_inputs.get(j).getDataType() == DataType.MATRIX ? (api == GeneratorAPI.JAVA ?
varj : varj + ".vals(0)") : varj);
-
+ varj.startsWith("a") ? (api ==
GeneratorAPI.JAVA ? varj :
+ (_inputs.get(j).getDataType() ==
DataType.MATRIX ? varj + ".vals(0)" : varj)) :
+ varj.startsWith("b") ? (api ==
GeneratorAPI.JAVA ? varj + ".values(rix)" :
+ (_type ==
BinType.VECT_MATRIXMULT ? varj : varj + ".vals(0)")) :
+ _inputs.get(j).getDataType() ==
DataType.MATRIX ? (api == GeneratorAPI.JAVA ? varj : varj + ".vals(0)") : varj);
+
+ tmp = tmp.replace("%SLEN"+(j+1)+"%",
varj.startsWith("STMP") ? varj+".size()" : varj.startsWith("a") ? "alen" :
"blen");
+
//replace start position of main input
- tmp = tmp.replace("%POS"+(j+1)+"%", (_inputs.get(j)
instanceof CNodeData
- &&
_inputs.get(j).getDataType().isMatrix()) ? (!varj.startsWith("b")) ? varj+"i" :
-
((TemplateUtils.isMatrix(_inputs.get(j)) || (_type.isElementwise()
- &&
TemplateUtils.isColVector(_inputs.get(j)))) && _type!=BinType.VECT_MATRIXMULT) ?
+ tmp = tmp.replace("%POS"+(j+1)+"%", (_inputs.get(j)
instanceof CNodeData
+ && _inputs.get(j).getDataType().isMatrix()) ?
(!varj.startsWith("b")) ? varj+"i" :
+ ((TemplateUtils.isMatrix(_inputs.get(j)) ||
(_type.isElementwise()
+ &&
TemplateUtils.isColVector(_inputs.get(j)))) && _type!=BinType.VECT_MATRIXMULT) ?
varj + ".pos(rix)" : "0" : "0");
}
//replace length information (e.g., after matrix mult)
- if( _type == BinType.VECT_OUTERMULT_ADD || (_type ==
BinType.VECT_CBIND && vectorVector) ) {
+ if( _type == BinType.VECT_OUTERMULT_ADD || (_type ==
BinType.VECT_CBIND && vectorVector)) {
for( int j=0; j<2; j++ )
tmp = tmp.replace("%LEN"+(j+1)+"%",
_inputs.get(j).getVectorLength(api));
}
- else { //general case
+ else { //general case
CNode mInput = getIntermediateInputVector();
if( mInput != null )
tmp = tmp.replace("%LEN%",
mInput.getVectorLength(api));
}
-
+
sb.append(tmp);
-
+
//mark as generated
_generated = true;
-
+
return sb.toString();
}
@@ -219,7 +241,126 @@ public class CNodeBinary extends CNode {
if( getInput().get(i).getDataType().isMatrix() )
return getInput().get(i);
return null;
- }
+ }
+
+ private boolean getTemplateType(double sparsityEst, double scalarVal) {
+ if(!DMLScript.SPARSE_INTERMEDIATE)
+ return false;
+ else {
+ switch(_type) {
+ case VECT_MULT:
+ case VECT_DIV:
+ case VECT_LESS:
+ case VECT_MINUS:
+ case VECT_PLUS:
+ case VECT_XOR:
+ case VECT_BITWAND:
+ case VECT_BIASADD:
+ case VECT_BIASMULT:
+ case VECT_MIN:
+ case VECT_MAX:
+ case VECT_NOTEQUAL:
+ case VECT_GREATER:
+ case VECT_EQUAL:
+ case VECT_LESSEQUAL:
+ case VECT_GREATEREQUAL: return sparsityEst <
0.1;
+ case VECT_MULT_SCALAR:
+ case VECT_DIV_SCALAR:
+ case VECT_XOR_SCALAR:
+ case VECT_BITWAND_SCALAR: return sparsityEst <
0.3;
+ case VECT_GREATER_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal >= 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal < 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_GREATEREQUAL_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal > 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal <= 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_MIN_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal >= 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal >= 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_LESS_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal <= 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal > 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_LESSEQUAL_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal < 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal >= 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_MAX_SCALAR: {
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal <= 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal <= 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_POW_SCALAR:
+ case VECT_EQUAL_SCALAR:{
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal != 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal != 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ case VECT_NOTEQUAL_SCALAR:{
+ if(scalarVal != Double.NaN) {
+ return
_inputs.get(1).getDataType().isScalar() ? scalarVal == 0 && sparsityEst < 0.2
+ :
_inputs.get(0).getDataType().isScalar() && scalarVal == 0 && sparsityEst < 0.2;
+ } else
+ return false;
+ }
+ default: return sparsityEst < 0.3;
+ }
+ }
+ }
+
+ public boolean getOutputType(boolean scalarVector, boolean lsparseLhs,
boolean lsparseRhs) {
+ switch(_type) {
+ case VECT_POW_SCALAR: return !scalarVector &&
lsparseLhs;
+ case VECT_MULT_SCALAR:
+ case VECT_DIV_SCALAR:
+ case VECT_XOR_SCALAR:
+ case VECT_MIN_SCALAR:
+ case VECT_MAX_SCALAR:
+ case VECT_EQUAL_SCALAR:
+ case VECT_NOTEQUAL_SCALAR:
+ case VECT_LESS_SCALAR:
+ case VECT_LESSEQUAL_SCALAR:
+ case VECT_GREATER_SCALAR:
+ case VECT_GREATEREQUAL_SCALAR:
+ case VECT_BITWAND_SCALAR: return lsparseLhs ||
lsparseRhs;
+ case VECT_MULT:
+ case VECT_DIV:
+ case VECT_MINUS:
+ case VECT_PLUS:
+ case VECT_XOR:
+ case VECT_BITWAND:
+ case VECT_BIASADD:
+ case VECT_BIASMULT:
+ case VECT_MIN:
+ case VECT_MAX:
+ case VECT_NOTEQUAL:
+ case VECT_LESS:
+ case VECT_GREATER: return lsparseLhs && lsparseRhs;
+ default: return false;
+ }
+ }
@Override
public String toString() {
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeNary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeNary.java
index dcf18ec656..35c351546d 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeNary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeNary.java
@@ -60,7 +60,10 @@ public class CNodeNary extends CNode
sb.append( sparseInput ?
"
LibSpoofPrimitives.vectWrite("+varj+"vals, %TMP%, "
+varj+"ix, "+pos+", "+off+", "+input._cols+");\n" :
- "
LibSpoofPrimitives.vectWrite("+(varj.startsWith("b")?varj+".values(rix)":varj)
+
varj.startsWith("STMP") ?
+ "
LibSpoofPrimitives.vectWrite("+varj+".values(), %TMP%, "
+
+varj+".indexes(), "+pos+", "+off+", "+varj+".size());\n" :
+ "
LibSpoofPrimitives.vectWrite("+(varj.startsWith("b")?varj+".values(rix)":varj)
+",
%TMP%, "+pos+", "+off+", "+input._cols+");\n");
off += input._cols;
}
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeRow.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeRow.java
index 77dec97cbe..c0d06b4bcb 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeRow.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeRow.java
@@ -37,6 +37,7 @@ public class CNodeRow extends CNodeTpl
+ "import
org.apache.sysds.runtime.codegen.SpoofOperator.SideInput;\n"
+ "import org.apache.sysds.runtime.codegen.SpoofRowwise;\n"
+ "import
org.apache.sysds.runtime.codegen.SpoofRowwise.RowType;\n"
+ + "import org.apache.sysds.runtime.data.SparseRowVector;\n"
+ "import org.apache.commons.math3.util.FastMath;\n"
+ "\n"
+ "public final class %TMP% extends SpoofRowwise { \n"
@@ -162,7 +163,8 @@ private static final String TEMPLATE_ROWAGG_OUT_CUDA =
"\t\tif(threadIdx.x == 0
case NO_AGG_B1:
case NO_AGG_CONST:
if(api == GeneratorAPI.JAVA)
- return
TEMPLATE_NOAGG_OUT.replace("%IN%", varName).replace("%LEN%",
_output.getVarname()+".length");
+ return
TEMPLATE_NOAGG_OUT.replace("%IN%",
varName.startsWith("STMP")?varName+".values(),
"+varName+".indexes()":varName).replace("%LEN%",
+ varName.startsWith("STMP") ?
varName+".size()" : _output.getVarname()+".length");
else
return
TEMPLATE_NOAGG_CONST_OUT_CUDA.replace("%IN%", varName +
".vals(0)").replaceAll("%LEN%", _output.getVarname()+".length");
case FULL_AGG:
diff --git
a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeTernary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeTernary.java
index 5e81109283..c6ff9802b1 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeTernary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeTernary.java
@@ -82,10 +82,13 @@ public class CNodeTernary extends CNode
String varj = _inputs.get(j-1).getVarname();
//replace sparse and dense inputs
tmp = tmp.replace("%IN"+j+"v%",
- varj+(varj.startsWith("a")?"vals":"") );
+ varj+(varj.startsWith("a")?"vals" :
varj.startsWith("STMP") ? ".values()" :"") );
tmp = tmp.replace("%IN"+j+"i%",
- varj+(varj.startsWith("a")?"ix":"") );
+ varj+(varj.startsWith("a")?"ix":
varj.startsWith("STMP") ? ".indexes()" :"") );
tmp = tmp.replace("%IN"+j+"%", varj );
+ tmp = tmp.replace("%POS%", varj.startsWith("a") ?
varj+"i" : varj.startsWith("STMP") ? "0" : "");
+ tmp = tmp.replace("%LEN%",
+ varj.startsWith("a") ? "alen" :
varj.startsWith("STMP") ? varj+".size()" : "");
}
sb.append(tmp);
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeUnary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeUnary.java
index fe67995b6b..ffbe0087f1 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeUnary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CNodeUnary.java
@@ -23,6 +23,7 @@ import java.util.Arrays;
import org.apache.commons.lang3.ArrayUtils;
import org.apache.commons.lang3.StringUtils;
+import org.apache.sysds.api.DMLScript;
import org.apache.sysds.common.Opcodes;
import org.apache.sysds.common.Types.DataType;
import org.apache.sysds.runtime.util.UtilFunctions;
@@ -85,13 +86,21 @@ public class CNodeUnary extends CNode
}
private UnaryType _type;
-
+ private boolean sparseTemplate;
+
public CNodeUnary( CNode in1, UnaryType type ) {
_inputs.add(in1);
_type = type;
setOutputDims();
}
-
+
+ public CNodeUnary( CNode in1, UnaryType type, double sparsity ) {
+ _inputs.add(in1);
+ _type = type;
+ setOutputDims();
+ sparseTemplate = getTemplateType(sparsity);
+ }
+
public UnaryType getType() {
return _type;
}
@@ -111,11 +120,13 @@ public class CNodeUnary extends CNode
sb.append(_inputs.get(0).codegen(sparse, api));
//generate unary operation
- boolean lsparse = sparse && (_inputs.get(0) instanceof CNodeData
- && _inputs.get(0).getVarname().startsWith("a")
- && !_inputs.get(0).isLiteral());
- String var = createVarname();
- String tmp = getLanguageTemplateClass(this,
api).getTemplate(_type, lsparse);
+ boolean lsparse = sparse &&
+ ((_inputs.get(0) instanceof CNodeData
+ && _inputs.get(0).getVarname().startsWith("a")
+ && !_inputs.get(0).isLiteral())
+ ||
_inputs.get(0).getVarname().startsWith("STMP"));
+ String var = createVarname(sparseTemplate && lsparse &&
getOutputType());
+ String tmp = getLanguageTemplateClass(this,
api).getTemplate(_type, lsparse, sparseTemplate);
tmp = tmp.replaceAll("%TMP%", var);
//replace sparse and dense inputs
@@ -130,6 +141,46 @@ public class CNodeUnary extends CNode
return sb.toString();
}
+
+ public boolean getTemplateType(double sparsity) {
+ if(!DMLScript.SPARSE_INTERMEDIATE)
+ return false;
+ else {
+ switch(_type) {
+ case VECT_SQRT:
+ case VECT_ABS:
+ case VECT_ROUND:
+ case VECT_CEIL:
+ case VECT_FLOOR:
+ case VECT_SIN:
+ case VECT_TAN:
+ case VECT_ASIN:
+ case VECT_ATAN:
+ case VECT_SINH:
+ case VECT_TANH:
+ case VECT_SIGN: return sparsity <= 0.3;
+ default: return false;
+ }
+ }
+ }
+
+ public boolean getOutputType() {
+ switch(_type) {
+ case VECT_SQRT:
+ case VECT_ABS:
+ case VECT_ROUND:
+ case VECT_CEIL:
+ case VECT_FLOOR:
+ case VECT_SIN:
+ case VECT_TAN:
+ case VECT_ASIN:
+ case VECT_ATAN:
+ case VECT_SINH:
+ case VECT_TANH:
+ case VECT_SIGN: return true;
+ default: return false;
+ }
+ }
@Override
public String toString() {
diff --git
a/src/main/java/org/apache/sysds/hops/codegen/cplan/CodeTemplate.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/CodeTemplate.java
index e29594f525..1ca8b4a3c1 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/CodeTemplate.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/CodeTemplate.java
@@ -31,9 +31,17 @@ public abstract class CodeTemplate {
public String getTemplate() {
throw new RuntimeException("Calling wrong getTemplate method on
" + getClass().getCanonicalName());
}
-
+
+ /**
+ * @param sparseTemplate added to turn SparseRowVector intermediates on
and off
+ */
public String getTemplate(CNodeBinary.BinType type, boolean sparseLhs,
boolean sparseRhs, boolean scalarVector,
- boolean scalarInput, boolean vectorVector) {
+ boolean scalarInput, boolean vectorVector, boolean
sparseTemplate) {
+ throw new RuntimeException("Calling wrong getTemplate method on
" + getClass().getCanonicalName());
+ }
+
+ public String getTemplate(CNodeBinary.BinType type, boolean sparseLhs,
boolean sparseRhs,
+ boolean scalarVector, boolean scalarInput, boolean
vectorVector) {
throw new RuntimeException("Calling wrong getTemplate method on
" + getClass().getCanonicalName());
}
@@ -44,6 +52,10 @@ public abstract class CodeTemplate {
public String getTemplate(CNodeUnary.UnaryType type, boolean sparse) {
throw new RuntimeException("Calling wrong getTemplate method on
" + getClass().getCanonicalName());
}
+
+ public String getTemplate(CNodeUnary.UnaryType type, boolean sparse,
boolean sparseTemplate) {
+ throw new RuntimeException("Calling wrong getTemplate method on
" + getClass().getCanonicalName());
+ }
public static String getTemplate(String templateFileName) {
try {
@@ -68,5 +80,5 @@ public abstract class CodeTemplate {
return null;
}
}
-
+
}
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Binary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Binary.java
index 40496249e5..34c4ee9088 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Binary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Binary.java
@@ -25,7 +25,7 @@ import org.apache.sysds.hops.codegen.cplan.CodeTemplate;
public class Binary extends CodeTemplate {
public String getTemplate(BinType type, boolean sparseLhs, boolean
sparseRhs,
- boolean scalarVector, boolean scalarInput, boolean vectorVector)
+ boolean scalarVector, boolean scalarInput, boolean
vectorVector, boolean sparseTemplate)
{
switch (type) {
case ROWMAXS_VECTMULT:
@@ -68,13 +68,22 @@ public class Binary extends CodeTemplate {
}
//vector-scalar operations
+ case VECT_POW_SCALAR: {
+ String vectName = type.getVectorPrimitiveName();
+ if( scalarVector )
+ return sparseRhs ? " double[] %TMP%
= LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %IN2i%, %POS2%, alen,
%LEN%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS2%, %LEN%);\n";
+ else if(sparseTemplate) {
+ return sparseLhs ? " SparseRowVector
%TMP% = LibSpoofPrimitives.vect"+vectName+"Write(%LEN%, %IN1v%, %IN2%, %IN1i%,
%POS1%, %SLEN1%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %LEN%);\n";
+ } else {
+ return sparseLhs ? " double[] %TMP%
= LibSpoofPrimitives.vect"+vectName+"Write(%IN1v%, %IN2%, %IN1i%, %POS1%, alen,
%LEN%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %LEN%);\n";
+ }
+ }
case VECT_MULT_SCALAR:
case VECT_DIV_SCALAR:
- case VECT_MINUS_SCALAR:
- case VECT_PLUS_SCALAR:
- case VECT_POW_SCALAR:
case VECT_XOR_SCALAR:
- case VECT_BITWAND_SCALAR:
case VECT_MIN_SCALAR:
case VECT_MAX_SCALAR:
case VECT_EQUAL_SCALAR:
@@ -82,7 +91,22 @@ public class Binary extends CodeTemplate {
case VECT_LESS_SCALAR:
case VECT_LESSEQUAL_SCALAR:
case VECT_GREATER_SCALAR:
- case VECT_GREATEREQUAL_SCALAR: {
+ case VECT_GREATEREQUAL_SCALAR:
+ case VECT_BITWAND_SCALAR: {
+ String vectName = type.getVectorPrimitiveName();
+ if(scalarVector) {
+ if(sparseRhs)
+ return sparseTemplate ? "
SparseRowVector %TMP% = LibSpoofPrimitives.vect"+vectName+"Write(%LEN%, %IN1%,
%IN2v%, %IN2i%, %POS2%, %SLEN1%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %IN2i%, %POS2%, alen,
%LEN%);\n";
+ } else {
+ if(sparseLhs)
+ return sparseTemplate ? "
SparseRowVector %TMP% = LibSpoofPrimitives.vect"+vectName+"Write(%LEN%, %IN1v%,
%IN2%, %IN1i%, %POS1%, %SLEN1%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1v%, %IN2%, %IN1i%, %POS1%, alen,
%LEN%);\n";
+ }
+ return " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %LEN%);\n";
+ }
+ case VECT_MINUS_SCALAR:
+ case VECT_PLUS_SCALAR: {
String vectName = type.getVectorPrimitiveName();
if( scalarVector )
return sparseRhs ? " double[] %TMP%
= LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %IN2i%, %POS2%, alen,
%LEN%);\n" :
@@ -115,20 +139,34 @@ public class Binary extends CodeTemplate {
case VECT_BIASMULT:
case VECT_MIN:
case VECT_MAX:
- case VECT_EQUAL:
case VECT_NOTEQUAL:
case VECT_LESS:
+ case VECT_GREATER:{
+ String vectName = type.getVectorPrimitiveName();
+ if(sparseTemplate && sparseLhs && sparseRhs) {
+ return " SparseRowVector %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%LEN%, %IN1v%, %IN2v%, %IN1i%, %IN2i%,
%POS1%, %POS2%, %SLEN1%, %SLEN2%);\n";
+ } else {
+ return sparseLhs ?
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1v%, %IN2%, %IN1i%, %POS1%, %POS2%,
alen, %LEN%);\n" :
+ sparseRhs ?
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %POS1%, %IN2i%, %POS2%,
alen, %LEN%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %POS2%,
%LEN%);\n";
+ }
+ }
+ case VECT_EQUAL:
case VECT_LESSEQUAL:
- case VECT_GREATER:
case VECT_GREATEREQUAL: {
String vectName = type.getVectorPrimitiveName();
- return sparseLhs ?
+ if(sparseTemplate && sparseLhs && sparseRhs) {
+ return " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%LEN%, %IN1v%, %IN2v%, %IN1i%, %IN2i%,
%POS1%, %POS2%, %SLEN1%, %SLEN2%);\n";
+ } else {
+ return sparseLhs ?
" double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1v%, %IN2%, %IN1i%, %POS1%, %POS2%,
alen, %LEN%);\n" :
sparseRhs ?
- " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %POS1%, %IN2i%, %POS2%,
alen, %LEN%);\n" :
- " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %POS2%,
%LEN%);\n";
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2v%, %POS1%, %IN2i%, %POS2%,
alen, %LEN%);\n" :
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %IN2%, %POS1%, %POS2%,
%LEN%);\n";
+ }
}
-
//scalar-scalar operations
case MULT:
return " double %TMP% = %IN1% * %IN2%;\n";
diff --git
a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Ternary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Ternary.java
index a86d51cca8..64d282bbb8 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Ternary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Ternary.java
@@ -51,7 +51,7 @@ public class Ternary extends CodeTemplate {
case LOOKUP_RC1:
return sparse ?
- " double %TMP% = getValue(%IN1v%,
%IN1i%, ai, alen, %IN3%-1);\n" :
+ " double %TMP% = getValue(%IN1v%,
%IN1i%, %POS%, %LEN%, %IN3%-1);\n" :
" double %TMP% = getValue(%IN1%,
%IN2%, rix, %IN3%-1);\n";
case LOOKUP_RVECT1:
diff --git a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Unary.java
b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Unary.java
index d8a1085df5..ef256223b9 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Unary.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/cplan/java/Unary.java
@@ -25,7 +25,7 @@ import org.apache.sysds.hops.codegen.cplan.CodeTemplate;
public class Unary extends CodeTemplate {
@Override
- public String getTemplate(UnaryType type, boolean sparse) {
+ public String getTemplate(UnaryType type, boolean sparse, boolean
sparseTemplate) {
switch( type ) {
case ROW_SUMS:
case ROW_SUMSQS:
@@ -38,25 +38,32 @@ public class Unary extends CodeTemplate {
return sparse ? " double %TMP% =
LibSpoofPrimitives.vect"+vectName+"(%IN1v%, %IN1i%, %POS1%, alen, len);\n":
" double %TMP% =
LibSpoofPrimitives.vect"+vectName+"(%IN1%, %POS1%, %LEN%);\n";
}
- case VECT_EXP:
- case VECT_POW2:
- case VECT_MULT2:
+
case VECT_SQRT:
- case VECT_LOG:
case VECT_ABS:
case VECT_ROUND:
case VECT_CEIL:
case VECT_FLOOR:
- case VECT_SIGN:
case VECT_SIN:
- case VECT_COS:
case VECT_TAN:
case VECT_ASIN:
- case VECT_ACOS:
case VECT_ATAN:
case VECT_SINH:
- case VECT_COSH:
case VECT_TANH:
+ case VECT_SIGN:{
+ String vectName = type.getVectorPrimitiveName();
+ return sparse ? sparseTemplate ?
+ " SparseRowVector %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(len, %IN1v%, %IN1i%, %POS1%, alen);\n"
:
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1v%, %IN1i%, %POS1%, alen, len);\n"
:
+ " double[] %TMP% =
LibSpoofPrimitives.vect"+vectName+"Write(%IN1%, %POS1%, %LEN%);\n";
+ }
+ case VECT_EXP:
+ case VECT_POW2:
+ case VECT_MULT2:
+ case VECT_LOG:
+ case VECT_COS:
+ case VECT_ACOS:
+ case VECT_COSH:
case VECT_CUMSUM:
case VECT_CUMMIN:
case VECT_CUMMAX:
diff --git
a/src/main/java/org/apache/sysds/hops/codegen/template/TemplateRow.java
b/src/main/java/org/apache/sysds/hops/codegen/template/TemplateRow.java
index 955bf778b8..8efb0245e9 100644
--- a/src/main/java/org/apache/sysds/hops/codegen/template/TemplateRow.java
+++ b/src/main/java/org/apache/sysds/hops/codegen/template/TemplateRow.java
@@ -37,6 +37,7 @@ import org.apache.sysds.hops.NaryOp;
import org.apache.sysds.hops.ParameterizedBuiltinOp;
import org.apache.sysds.hops.TernaryOp;
import org.apache.sysds.hops.UnaryOp;
+import org.apache.sysds.hops.OptimizerUtils;
import org.apache.sysds.hops.codegen.cplan.CNode;
import org.apache.sysds.hops.codegen.cplan.CNodeBinary;
import org.apache.sysds.hops.codegen.cplan.CNodeData;
@@ -391,7 +392,7 @@ public class TemplateRow extends TemplateBase
{
if( HopRewriteUtils.isUnary(hop,
SUPPORTED_VECT_UNARY) ) {
String opname =
"VECT_"+((UnaryOp)hop).getOp().name();
- out = new CNodeUnary(cdata1,
UnaryType.valueOf(opname));
+ out = new CNodeUnary(cdata1,
UnaryType.valueOf(opname), hop.getInput(0).getSparsity());
if( cdata1 instanceof CNodeData &&
!inHops2.containsKey("X") )
inHops2.put("X",
hop.getInput().get(0));
}
@@ -403,7 +404,7 @@ public class TemplateRow extends TemplateBase
{
cdata1 =
TemplateUtils.wrapLookupIfNecessary(cdata1, hop.getInput().get(0));
String primitiveOpName =
((UnaryOp)hop).getOp().name();
- out = new CNodeUnary(cdata1,
UnaryType.valueOf(primitiveOpName));
+ out = new CNodeUnary(cdata1,
UnaryType.valueOf(primitiveOpName), hop.getInput(0).getSparsity());
}
}
else if(HopRewriteUtils.isBinary(hop, OpOp2.CBIND)) {
@@ -440,7 +441,14 @@ public class TemplateRow extends TemplateBase
cdata1 = new CNodeUnary(cdata1,
UnaryType.LOOKUP_R);
if( TemplateUtils.isColVector(cdata2) )
cdata2 = new CNodeUnary(cdata2,
UnaryType.LOOKUP_R);
- out = getVectorBinary(cdata1, cdata2,
((BinaryOp)hop).getOp().name());
+ String opName =
((BinaryOp)hop).getOp().name();
+ Hop hopIn1 = hop.getInput(0);
+ Hop hopIn2 = hop.getInput(1);
+ double sparsityEst =
OptimizerUtils.getBinaryOpSparsity(
+ hopIn1.getSparsity(),
hopIn2.getSparsity(), OpOp2.valueOf(opName), false);
+ double literalVal = hopIn1 instanceof
LiteralOp ? ((LiteralOp) hopIn1).getDoubleValue()
+ : hopIn2 instanceof LiteralOp ?
((LiteralOp) hopIn2).getDoubleValue() : Double.NaN;
+ out = getVectorBinary(cdata1, cdata2,
opName, sparsityEst, literalVal);
if( cdata1 instanceof CNodeData &&
!inHops2.containsKey("X")
&&
!(cdata1.getDataType()==DataType.SCALAR) ) {
inHops2.put("X",
hop.getInput().get(0));
@@ -569,7 +577,17 @@ public class TemplateRow extends TemplateBase
return new CNodeBinary(cdata1, cdata2,
BinType.valueOf("VECT_"+name+"_SCALAR"));
}
}
-
+
+ private static CNodeBinary getVectorBinary(CNode cdata1, CNode cdata2,
String name, double sparsity, double literalVal) {
+ if( TemplateUtils.isMatrix(cdata1) &&
(TemplateUtils.isMatrix(cdata2)
+ || TemplateUtils.isRowVector(cdata2)) ) {
+ return new CNodeBinary(cdata1, cdata2,
BinType.valueOf("VECT_"+name), sparsity, literalVal);
+ }
+ else {
+ return new CNodeBinary(cdata1, cdata2,
BinType.valueOf("VECT_"+name+"_SCALAR"), sparsity, literalVal);
+ }
+ }
+
/**
* Comparator to order input hops of the row aggregate template. We try
* to order matrices-vectors-scalars via sorting by number of cells but
diff --git
a/src/main/java/org/apache/sysds/runtime/codegen/LibSpoofPrimitives.java
b/src/main/java/org/apache/sysds/runtime/codegen/LibSpoofPrimitives.java
index 6c0dc395c3..bc6ba19895 100644
--- a/src/main/java/org/apache/sysds/runtime/codegen/LibSpoofPrimitives.java
+++ b/src/main/java/org/apache/sysds/runtime/codegen/LibSpoofPrimitives.java
@@ -23,6 +23,7 @@ import java.util.Arrays;
import org.apache.commons.math3.util.FastMath;
import org.apache.sysds.runtime.data.DenseBlockFP64;
+import org.apache.sysds.runtime.data.SparseRowVector;
import org.apache.sysds.runtime.functionobjects.BitwAnd;
import org.apache.sysds.runtime.functionobjects.IntegerDivide;
import org.apache.sysds.runtime.functionobjects.Modulus;
@@ -51,6 +52,10 @@ public class LibSpoofPrimitives
@Override protected VectorBuffer initialValue() { return new
VectorBuffer(0,0,0); }
};
+ private static ThreadLocal<SparseVectorBuffer> sparseMemPool = new
ThreadLocal<>() {
+ @Override protected SparseVectorBuffer initialValue() { return
new SparseVectorBuffer(0,0,0); }
+ };
+
public static double rowMaxsVectMult(double[] a, double[] b, int ai,
int bi, int len) {
double val = Double.NEGATIVE_INFINITY;
int j=0;
@@ -2164,6 +2169,1155 @@ public class LibSpoofPrimitives
return (vectSum(avalsSqr, 0, len)-len*meanVal)/(len-1);
}
+ /**
+ * Vector primitives with SparseRowVector intermediates
+ * Changes:
+ * - Changed method signature to avoid method duplicate conflicts
+ * e.g. (double[], double, int[], int, int, int) --> (int,
double[], double, int[], int, int)
+ * - Added blen for vector - vector calculations to be able to use
both vectors as SparseRowVectors
+ * - Implemented a new SparseVectorBuffer class that creates a ring
buffer for SparseRowVectors in different sizes
+ */
+
+ public static SparseRowVector vectMultWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ if( a == null ) return c;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j]*bval;
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectMultWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectMultWrite(len, a, bval, aix, ai, alen);
+ }
+
+ //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;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ if(aIdx == bIdx) {
+ indexes[index] = aIdx;
+ values[index] = a[aItr] * b[bItr];
+ aItr++;
+ bItr++;
+ index++;
+ } else if(aIdx < bIdx)
+ aItr++;
+ else
+ bItr++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ //version without branching
+ public static SparseRowVector vectMultWrite(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));
+ int index = 0;
+ int aItr = ai;
+ int bItr = bi;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ indexes[index] = aIdx;
+ values[index] = a[aItr] * b[bItr];
+ index += aIdx == bIdx ? 1 : 0;
+ aItr += aIdx <= bIdx ? 1 : 0;
+ bItr += aIdx >= bIdx ? 1 : 0;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static void vectWrite(double[] a, int[] aix, double[] c, int ci,
int len) {
+ if( a == null ) return;
+ for(int j = 0; j < len; j++)
+ c[ci+aix[j]] = a[j];
+ }
+
+ public static void vectWrite(double[] a, double[] c, int[] aix, int ai,
int ci, int alen) {
+ if( a == null ) return;
+ for(int j = 0; j < alen; j++)
+ c[ci+aix[ai+j]] = a[ai+j];
+ }
+
+ public static SparseRowVector vectDivWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for( int j = 0; j < alen; j++ ) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] / bval;
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectDivWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = bval / a[ai+j];
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ //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;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ if(aIdx == bIdx) {
+ indexes[index] = aIdx;
+ values[index] = a[aItr] / b[bItr];
+ aItr++;
+ bItr++;
+ index++;
+ } else if(aIdx < bIdx) {
+ indexes[index] = aIdx;
+ values[index] = (a[aItr]>0) ?
Double.POSITIVE_INFINITY : Double.NEGATIVE_INFINITY;
+ aItr++;
+ index++;
+ } else {
+ bItr++;
+ }
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ //version without branching
+ public static SparseRowVector vectDivWrite(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));
+ int index = 0;
+ int aItr = ai;
+ int bItr = bi;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ indexes[index] = aIdx;
+ values[index] = a[aItr] / b[bItr];
+ index += aIdx == bIdx ? 1 : 0;
+ aItr += aIdx <= bIdx ? 1 : 0;
+ bItr += aIdx >= bIdx ? 1 : 0;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectMinusWriteB(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ if(aIdx == bIdx) {
+ indexes[index] = aIdx;
+ values[index] = a[aItr] - b[bItr];
+ aItr++;
+ bItr++;
+ index++;
+ } else if(aIdx < bIdx) {
+ indexes[index] = aIdx;
+ values[index] = a[aItr];
+ aItr++;
+ index++;
+ } else {
+ indexes[index] = bIdx;
+ values[index] = -b[bItr];
+ bItr++;
+ index++;
+ }
+ }
+ for (; aItr < ai+alen; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr];
+ index++;
+ }
+ for (; bItr < bi+blen; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = -b[bItr];
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectMinusWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = av - bv;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++, index++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr];
+ }
+ for (; bItr < bEnd; bItr++, index++) {
+ indexes[index] = bix[bItr];
+ values[index] = -b[bItr];
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectPlusWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = av + bv;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr];
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = b[bItr];
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectXorWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ if(bval != 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = !(a[aItr] != 0) ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = (a[ai+j] != 0) ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectXorWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectXorWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectXorWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = ((av != 0) != (bv != 0)) ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = (a[aItr] != 0) ? 1 : 0;
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = (b[bItr] != 0) ? 1 : 0;
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectPowWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ if(bval == 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = Math.pow(a[aItr], bval)
- 1;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.pow(a[ai+j], bval);
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectMinWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ if(bval < 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = Math.min(a[aItr], bval);
+ aItr++;
+ } else {
+ values[index] = bval;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = bval;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.min(a[ai+j], bval);
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectMinWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectMinWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectMinWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = Math.min(av, bv);
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = Math.min(a[aItr], 0);
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = Math.min(b[bItr], 0);
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectMaxWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ if(bval > 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = Math.max(a[aItr], bval);
+ aItr++;
+ } else {
+ values[index] = bval;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = bval;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.max(a[ai+j], bval);
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectMaxWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectMaxWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectMaxWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = Math.max(av, bv);
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = Math.max(a[aItr], 0);
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = Math.max(b[bItr], 0);
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectEqualWrite(int len, double[] a,
double bval, int[] aix, int ai, int alen) {
+ if(bval == 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] == bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] == bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectEqualWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectEqualWrite(len, a, bval, aix, ai, alen);
+ }
+
+ //doesn't return SparseRowVector, but still uses two sparse vectors as
inputs
+ public static double[] vectEqualWrite(int len, double[] a, double[] b,
int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ double[] c = allocVector(len, true, 1);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ while(aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ int index = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ c[index] = av == bv ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ }
+ for (; aItr < aEnd; aItr++) c[aix[aItr]] = 0;
+ for (; bItr < bEnd; bItr++) c[bix[bItr]] = 0;
+ return c;
+ }
+
+ public static SparseRowVector vectNotequalWrite(int len, double[] a,
double bval, int[] aix, int ai, int alen) {
+ if(bval != 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] != bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] != bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectNotequalWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectNotequalWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectNotequalWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = (av != bv) ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr] != 0 ? 1 : 0;
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = b[bItr] != 0 ? 1 : 0;
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectLessWrite(int len, double[] a, double
bval, int[] aix, int ai, int alen) {
+ if(bval > 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] < bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] < bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectLessWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectGreaterequalWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectLessWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = av < bv ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < aEnd; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr] < 0? 1 : 0;
+ index++;
+ }
+ for (; bItr < bEnd; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = 0 < b[bItr] ? 1 : 0;
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectLessequalWrite(int len, double[] a,
double bval, int[] aix, int ai, int alen) {
+ if(bval >= 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] <= bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] <= bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectLessequalWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectGreaterWrite(len, a, bval, aix, ai, alen);
+ }
+
+ //doesn't return SparseRowVector, but still uses two sparse vectors as
inputs
+ public static double[] vectLessequalWrite(int len, double[] a, double[]
b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ double[] c = allocVector(len, true, 1);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ while(aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ int index = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ c[index] = av <= bv ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ }
+ for(; aItr < ai+alen; aItr++) c[aix[aItr]] = (a[aItr] <= 0) ? 1
: 0;
+ for(; bItr < bi+blen; bItr++) c[bix[bItr]] = (0 <= b[bItr]) ?
1 : 0;
+ return c;
+ }
+
+ public static SparseRowVector vectGreaterWrite(int len, double[] a,
double bval, int[] aix, int ai, int alen) {
+ if(bval < 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] > bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] > bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectGreaterWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectLessequalWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectGreaterWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ SparseRowVector c = allocSparseVector(alen+blen);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ int index = 0;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while (aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ indexes[index] = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ values[index] = av >bv ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ index++;
+ }
+ for (; aItr < ai+alen; aItr++) {
+ indexes[index] = aix[aItr];
+ values[index] = a[aItr] > 0 ? 1 : 0;
+ index++;
+ }
+ for (; bItr < bi+blen; bItr++) {
+ indexes[index] = bix[bItr];
+ values[index] = 0 > b[bItr] ? 1 : 0;
+ index++;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectGreaterequalWrite(int len, double[]
a, double bval, int[] aix, int ai, int alen) {
+ if(bval <= 0) {
+ SparseRowVector c = allocSparseVector(len);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int index = 0;
+ int aItr = 0;
+ while(aItr < ai+alen && index < len) {
+ indexes[index] = index;
+ if(aix[aItr] == index) {
+ values[index] = a[aItr] >= bval ? 1 : 0;
+ aItr++;
+ } else {
+ values[index] = 1;
+ }
+ index++;
+ }
+ for(; index < len; index++) {
+ indexes[index] = index;
+ values[index] = 1;
+ }
+ c.setSize(len);
+ return c;
+ } else {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = a[ai+j] >= bval ? 1 : 0;
+ }
+ c.setSize(alen);
+ return c;
+ }
+ }
+
+ public static SparseRowVector vectGreaterequalWrite(int len, double
bval, double[] a, int[] aix, int ai, int alen) {
+ return vectLessWrite(len, a, bval, aix, ai, alen);
+ }
+
+ //doesn't return SparseRowVector, but still uses two sparse vectors as
inputs
+ public static double[] vectGreaterequalWrite(int len, double[] a,
double[] b, int[] aix, int[] bix, int ai, int bi, int alen, int blen) {
+ double[] c = allocVector(len, true, 1);
+ int aEnd = ai+alen;
+ int bEnd = bi+blen;
+ int aItr = ai;
+ int bItr = bi;
+ while(aItr < aEnd && bItr < bEnd) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ int useA = (aIdx <= bIdx) ? 1 : 0;
+ int useB = (aIdx >= bIdx) ? 1 : 0;
+ int index = (useA == 1) ? aIdx : bIdx;
+ double av = (useA == 1) ? a[aItr] : 0.0;
+ double bv = (useB == 1) ? b[bItr] : 0.0;
+
+ c[index] = av >= bv ? 1 : 0;
+ aItr += useA;
+ bItr += useB;
+ }
+ for(; aItr < ai+alen; aItr++) c[aix[aItr]] = (a[aItr] >= 0) ? 1
: 0;
+ for(; bItr < bi+blen; bItr++) c[bix[bItr]] = (0 >= b[bItr]) ?
1 : 0;
+ return c;
+ }
+
+ public static SparseRowVector vectBitwandWrite(int len, double[] a,
double bval, int[] aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ int bval1 = (int) bval;
+ for( int j = 0; j < alen; j++ ) {
+ indexes[j] = aix[ai+j];
+ values[j] = bwAnd(a[ai+j], bval1);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectBitwandWrite(int len, double bval,
double[] a, int[] aix, int ai, int alen) {
+ return vectBitwandWrite(len, a, bval, aix, ai, alen);
+ }
+
+ public static SparseRowVector vectBitwandWrite(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));
+ int index = 0;
+ int aItr = ai;
+ int bItr = bi;
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ while(aItr < ai+alen && bItr < bi+blen) {
+ int aIdx = aix[aItr];
+ int bIdx = bix[bItr];
+ indexes[index] = aIdx;
+ values[index] = bwAnd(a[aItr], b[bItr]);
+ index += aIdx == bIdx ? 1 : 0;
+ aItr += aIdx <= bIdx ? 1 : 0;
+ bItr += aIdx >= bIdx ? 1 : 0;
+ }
+ c.setSize(index);
+ return c;
+ }
+
+ public static SparseRowVector vectSqrtWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.sqrt(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectAbsWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.abs(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectRoundWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.round(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectCeilWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.ceil(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectFloorWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.floor(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectSinWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.sin(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectTanWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.tan(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectAsinWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.asin(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectAtanWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.atan(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectSinhWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.sinh(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectTanhWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.tanh(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
+ public static SparseRowVector vectSignWrite(int len, double[] a, int[]
aix, int ai, int alen) {
+ SparseRowVector c = allocSparseVector(alen);
+ int[] indexes = c.indexes();
+ double[] values = c.values();
+ for(int j = 0; j < alen; j++) {
+ indexes[j] = aix[ai+j];
+ values[j] = Math.signum(a[ai+j]);
+ }
+ c.setSize(alen);
+ return c;
+ }
+
//complex builtin functions that are not directly generated
//(included here in order to reduce the number of imports)
@@ -2194,10 +3348,19 @@ public class LibSpoofPrimitives
if( numVectors > 0 )
memPool.set(new VectorBuffer(numVectors, len, len2));
}
+
+ public static void setupSparseThreadLocalMemory(int numVectors, int
len, int len2) {
+ if( numVectors > 0 )
+ sparseMemPool.set(new SparseVectorBuffer(numVectors,
len, len2));
+ }
public static void cleanupThreadLocalMemory() {
memPool.remove();
}
+
+ public static void cleanupSparseThreadLocalMemory() {
+ sparseMemPool.remove();
+ }
public static double[] allocVector(int len, boolean reset) {
return allocVector(len, reset, 0);
@@ -2217,6 +3380,21 @@ public class LibSpoofPrimitives
Arrays.fill(vect, resetVal);
return vect;
}
+
+ public static SparseRowVector allocSparseVector(int len) {
+ SparseVectorBuffer buff = sparseMemPool.get();
+
+ //find next matching vector in ring buffer or
+ //allocate new vector if no vector was returned
+ SparseRowVector vect = buff.next(len);
+ if(vect == null)
+ vect = new SparseRowVector(len);
+ //reset vector for normal outputs
+ else if(vect.size() != 0)
+ vect.reset(len, len);
+
+ return vect;
+ }
/**
* Simple ring buffer of allocated vectors, where
@@ -2265,4 +3443,52 @@ public class LibSpoofPrimitives
&& _data.length == lnum);
}
}
+
+ /**
+ * Simple ring buffer of allocated SparseRowVectors, where
+ * vectors of different sizes are interspersed.
+ */
+ private static class SparseVectorBuffer {
+ private static final int MAX_SIZE = 512*1024; //4MB
+ private final SparseRowVector[] _data;
+ private int _pos;
+ private int _len1;
+ private int _len2;
+
+ public SparseVectorBuffer(int num, int len1, int len2) {
+ //best effort size restriction since large intermediates
+ //not necessarily used (num refers to the total number)
+ len1 = Math.min(len1, MAX_SIZE);
+ len2 = Math.min(len2, MAX_SIZE);
+ //pre-allocate ring buffer
+ int lnum = (len2>0 && len1!=len2) ? 2*num : num;
+ _data = new SparseRowVector[lnum];
+ for( int i=0; i<num; i++ ) {
+ if( lnum > num ) {
+ _data[2*i] = new SparseRowVector(len1);
+ _data[2*i+1] = new
SparseRowVector(len2);
+ }
+ else {
+ _data[i] = new SparseRowVector(len1);
+ }
+ }
+ _pos = -1;
+ _len1 = len1;
+ _len2 = len2;
+ }
+ public SparseRowVector next(int len) {
+ if( _len1<len && _len2<len )
+ return null;
+ do {
+ _pos = (_pos+1>=_data.length) ? 0 : _pos+1;
+ } while( _data[_pos].values().length<len );
+ return _data[_pos];
+ }
+ @SuppressWarnings("unused")
+ public boolean isReusable(int num, int len1, int len2) {
+ int lnum = (len2>0 && len1!=len2) ? 2*num : num;
+ return (_len1 == len1 && _len2 == len2
+ && _data.length == lnum);
+ }
+ }
}
diff --git a/src/main/java/org/apache/sysds/runtime/codegen/SpoofRowwise.java
b/src/main/java/org/apache/sysds/runtime/codegen/SpoofRowwise.java
index eab223bc97..e48d44f33f 100644
--- a/src/main/java/org/apache/sysds/runtime/codegen/SpoofRowwise.java
+++ b/src/main/java/org/apache/sysds/runtime/codegen/SpoofRowwise.java
@@ -27,6 +27,7 @@ import java.util.concurrent.ExecutorService;
import java.util.concurrent.Future;
import java.util.stream.IntStream;
+import org.apache.sysds.api.DMLScript;
import org.apache.sysds.runtime.DMLRuntimeException;
import org.apache.sysds.runtime.compress.CompressedMatrixBlock;
import org.apache.sysds.runtime.controlprogram.caching.MatrixObject;
@@ -187,7 +188,12 @@ public abstract class SpoofRowwise extends SpoofOperator
//setup thread-local memory if necessary
if( allocTmp &&_reqVectMem > 0 )
- LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem,
n, n2);
+ if(inputs.get(0).isInSparseFormat() &&
DMLScript.SPARSE_INTERMEDIATE) {
+
LibSpoofPrimitives.setupSparseThreadLocalMemory(_reqVectMem, n, n2);
+
LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, n, n2);
+ } else {
+
LibSpoofPrimitives.setupThreadLocalMemory(_reqVectMem, n, n2);
+ }
//core sequential execute
MatrixBlock a = inputs.get(0);
@@ -201,7 +207,12 @@ public abstract class SpoofRowwise extends SpoofOperator
//post-processing
if( allocTmp &&_reqVectMem > 0 )
- LibSpoofPrimitives.cleanupThreadLocalMemory();
+ if(inputs.get(0).isInSparseFormat() &&
DMLScript.SPARSE_INTERMEDIATE) {
+
LibSpoofPrimitives.cleanupSparseThreadLocalMemory();
+ LibSpoofPrimitives.cleanupThreadLocalMemory();
+ } else {
+ LibSpoofPrimitives.cleanupThreadLocalMemory();
+ }
if( flipOut ) {
fixTransposeDimensions(out);
out = LibMatrixReorg.transpose(out, new MatrixBlock(
@@ -431,7 +442,12 @@ public abstract class SpoofRowwise extends SpoofOperator
//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() )
@@ -440,7 +456,12 @@ public abstract class SpoofRowwise extends SpoofOperator
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;
}
}
@@ -474,15 +495,25 @@ public abstract class SpoofRowwise extends SpoofOperator
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);
diff --git
a/src/test/java/org/apache/sysds/test/component/codegen/CPlanVectorPrimitivesTest.java
b/src/test/java/org/apache/sysds/test/component/codegen/CPlanVectorPrimitivesTest.java
index 7d8b4c9096..bae22b1c38 100644
---
a/src/test/java/org/apache/sysds/test/component/codegen/CPlanVectorPrimitivesTest.java
+++
b/src/test/java/org/apache/sysds/test/component/codegen/CPlanVectorPrimitivesTest.java
@@ -21,6 +21,7 @@ package org.apache.sysds.test.component.codegen;
import java.lang.reflect.Method;
+import org.apache.sysds.runtime.data.SparseRowVector;
import org.junit.Test;
import org.apache.sysds.common.Types.OpOp2;
import org.apache.sysds.hops.codegen.cplan.CNodeBinary.BinType;
@@ -716,6 +717,287 @@ public class CPlanVectorPrimitivesTest extends
AutomatedTestBase
testVectorBinaryPrimitive(BinType.VECT_BITWAND,
InputType.VECTOR_SPARSE, InputType.VECTOR_DENSE);
}
+ //********************testing with sparse
intermediates********************//
+ //vector - scalar
+
+ @Test
+ public void testVectorScalarMultSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MULT_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarDivSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_DIV_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarMinSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MIN_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarMaxSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MAX_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarPowSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_POW_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_EQUAL_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarNotEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_NOTEQUAL_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarLessSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESS_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarLessEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESSEQUAL_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarGreaterSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_GREATER_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarGreaterEqualSparseToSparse() {
+
testVectorBinarySparsePrimitive(BinType.VECT_GREATEREQUAL_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ //todo: this only makes sense, when bval is 0
+ @Test
+ public void testVectorScalarXorSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_XOR_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ @Test
+ public void testVectorScalarBitwAndSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_BITWAND_SCALAR,
InputType.VECTOR_SPARSE, InputType.SCALAR);
+ }
+
+ //scalar - vector
+
+ @Test
+ public void testScalarVectorMultSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MULT_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorDivSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_DIV_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorMinSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MIN_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorMaxSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MAX_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_EQUAL_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorNotEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_NOTEQUAL_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorLessSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESS_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorLessEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESSEQUAL_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorGreaterSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_GREATER_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorGreaterEqualSparseToSparse() {
+
testVectorBinarySparsePrimitive(BinType.VECT_GREATEREQUAL_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorXorSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_XOR_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testScalarVectorBitwAndSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_BITWAND_SCALAR,
InputType.SCALAR, InputType.VECTOR_SPARSE);
+ }
+
+ //special binary
+
+ // @Test
+ // public void testVectorPow2SparseToSparse() {
+ // testVectorUnarySparsePrimitive(UnaryType.VECT_POW2,
InputType.VECTOR_SPARSE);
+ // }
+ //
+ // @Test
+ // public void testVectorMult2SparseToSparse() {
+ // testVectorUnarySparsePrimitive(UnaryType.VECT_MULT2,
InputType.VECTOR_SPARSE);
+ // }
+
+ //vector - vector
+
+ @Test
+ public void testVectorVectorMultSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MULT,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorDivSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_DIV,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorPlusSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_PLUS,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorMinusSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MINUS,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorMinSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MIN,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+ @Test
+ public void testVectorVectorMaxSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_MAX,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_EQUAL,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorNotEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_NOTEQUAL,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorLessSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESS,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorLessEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_LESSEQUAL,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorGreaterSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_GREATER,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorGreaterEqualSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_GREATEREQUAL,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorXorSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_XOR,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorVectorBitwAndSparseToSparse() {
+ testVectorBinarySparsePrimitive(BinType.VECT_BITWAND,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ }
+
+ // @Test
+ // public void testVectorVectorMatrixMultSparseToSparse() {
+ //
testVectorBinarySparsePrimitive(BinType.VECT_MATRIXMULT,
InputType.VECTOR_SPARSE, InputType.VECTOR_SPARSE);
+ // }
+
+ //unary primitives with sparse intermediates
+
+ @Test
+ public void testVectorSqrtSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_SQRT,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorAbsSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_ABS,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorRoundSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_ROUND,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorCeilSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_CEIL,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorFloorSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_FLOOR,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorSinSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_SIN,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorTanSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_TAN,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorAsinSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_ASIN,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorAtanSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_ATAN,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorSinhSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_SINH,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorTanhSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_TANH,
InputType.VECTOR_SPARSE);
+ }
+
+ @Test
+ public void testVectorSignSparseToSparse() {
+ testVectorUnarySparsePrimitive(UnaryType.VECT_SIGN,
InputType.VECTOR_SPARSE);
+ }
+
@SuppressWarnings("incomplete-switch")
private static void testVectorAggPrimitive(UnaryType aggtype, InputType
type1)
{
@@ -792,6 +1074,45 @@ public class CPlanVectorPrimitivesTest extends
AutomatedTestBase
throw new RuntimeException(ex);
}
}
+
+ private static void testVectorUnarySparsePrimitive(UnaryType utype,
InputType type1)
+ {
+ try {
+ //generate input data
+ double sparsity = (type1 == InputType.VECTOR_DENSE) ?
sparsity1 : sparsity2;
+ MatrixBlock in = MatrixBlock.randOperations(m, n,
sparsity, -1, 1, "uniform", 7);
+
+ //get vector primitive via reflection
+ String meName =
"vect"+StringUtils.camelize(utype.name().split("_")[1])+"Write";
+ Method me = LibSpoofPrimitives.class.getMethod(meName,
new Class[]{int.class, double[].class, int[].class, int.class, int.class});
+
+
+ for( int i=0; i<m; i++ ) {
+ double[] ret1 = new double[n];
+ //execute vector primitive via reflection
+ SparseRowVector retX = (SparseRowVector)
me.invoke(null, n, in.getSparseBlock().values(i),
in.getSparseBlock().indexes(i),
+ in.getSparseBlock().pos(i),
in.getSparseBlock().size(i));
+
+ int[] indexes = retX.indexes();
+ for (int j = 0; j < retX.size(); j++) {
+ ret1[indexes[j]] = retX.get(indexes[j]);
+ }
+
+ //execute comparison operation
+ String opcode =
utype.name().split("_")[1].toLowerCase();
+ UnaryOperator uop = new
UnaryOperator(Builtin.getBuiltinFnObject(opcode));
+ double[] ret2 =
DataConverter.convertToDoubleVector(
+ in.slice(i, i, 0, n-1, new
MatrixBlock())
+ .unaryOperations(uop, new
MatrixBlock()), false);
+
+ //compare results
+ TestUtils.compareMatrices(ret1, ret2, eps);
+ }
+ }
+ catch( Exception ex ) {
+ throw new RuntimeException(ex);
+ }
+ }
private static void testVectorBinaryPrimitive(BinType bintype,
InputType type1, InputType type2)
{
@@ -868,4 +1189,144 @@ public class CPlanVectorPrimitivesTest extends
AutomatedTestBase
throw new RuntimeException(ex);
}
}
+
+ private static void testVectorBinarySparsePrimitive(BinType bintype,
InputType type1, InputType type2) {
+ try {
+ //generate input data (scalar later derived if needed)
+ double sparsityA = (type1 == InputType.VECTOR_DENSE) ?
sparsity1 : sparsity2;
+ MatrixBlock inA = MatrixBlock.randOperations(m, n,
sparsityA, -5, 5, "uniform", 3);
+ double sparsityB = (type2 == InputType.VECTOR_DENSE) ?
sparsity1 : sparsity2;
+ MatrixBlock inB = MatrixBlock.randOperations(m, n,
sparsityB, -5, 5, "uniform", 7);
+
+ boolean sparse = getOutputType(bintype);
+ int testType = getTestType(bintype);
+
+ //get vector primitive via reflection
+ String meName =
"vect"+StringUtils.camelize(bintype.name().split("_")[1])+"Write";
+ final Method me ;
+ if( type1==InputType.VECTOR_SPARSE &&
type2==InputType.SCALAR )
+ me = LibSpoofPrimitives.class.getMethod(meName,
new Class[]{int.class, double[].class, double.class, int[].class, int.class,
int.class});
+ else if( type1==InputType.SCALAR &&
type2==InputType.VECTOR_SPARSE )
+ me = LibSpoofPrimitives.class.getMethod(meName,
new Class[]{int.class, double.class, double[].class, int[].class, int.class,
int.class});
+ else //if( type1==InputType.VECTOR_SPARSE &&
type2==InputType.VECTOR_SPARSE )
+ me = LibSpoofPrimitives.class.getMethod(meName,
new Class[]{int.class, double[].class, double[].class, int[].class,
int[].class, int.class, int.class, int.class, int.class});
+
+ for( int i=0; i<m; i++ ) {
+ //execute vector primitive via reflection
+ double[] ret1 = new double[n];
+ SparseRowVector retX = null;
+ if( type1==InputType.VECTOR_SPARSE &&
type2==InputType.SCALAR )
+ if(testType >= 0 && i == m-1) {
+ if(testType == 0) {
+ retX =
(SparseRowVector) me.invoke(null, n, inA.getSparseBlock().values(i), 0,
inA.getSparseBlock().indexes(i),
+
inA.getSparseBlock().pos(i), inA.getSparseBlock().size(i));
+ } else {
+ retX =
(SparseRowVector) me.invoke(null, n, inA.getSparseBlock().values(i), inB.min(),
inA.getSparseBlock().indexes(i),
+
inA.getSparseBlock().pos(i), inA.getSparseBlock().size(i));
+ }
+ } else {
+ retX = (SparseRowVector)
me.invoke(null, n, inA.getSparseBlock().values(i), inB.max(),
inA.getSparseBlock().indexes(i),
+
inA.getSparseBlock().pos(i), inA.getSparseBlock().size(i));
+ }
+ else if( type1==InputType.SCALAR &&
type2==InputType.VECTOR_SPARSE )
+ if(testType >= 0 && i == m-1) {
+ if(testType == 0) {
+ retX =
(SparseRowVector) me.invoke(null, n, 0, inB.getSparseBlock().values(i),
+
inB.getSparseBlock().indexes(i), inB.getSparseBlock().pos(i),
inB.getSparseBlock().size(i));
+ } else {
+ retX =
(SparseRowVector) me.invoke(null, n, inA.min(), inB.getSparseBlock().values(i),
+
inB.getSparseBlock().indexes(i), inB.getSparseBlock().pos(i),
inB.getSparseBlock().size(i));
+ }
+ } else {
+ retX = (SparseRowVector)
me.invoke(null, n, inA.max(), inB.getSparseBlock().values(i),
+
inB.getSparseBlock().indexes(i), inB.getSparseBlock().pos(i),
inB.getSparseBlock().size(i));
+ }
+ else if( type1==InputType.VECTOR_SPARSE &&
type2==InputType.VECTOR_SPARSE )
+ if(sparse)
+ retX = (SparseRowVector)
me.invoke(null, n, inA.getSparseBlock().values(i),
inB.getSparseBlock().values(i),
+
inA.getSparseBlock().indexes(i), inB.getSparseBlock().indexes(i),
inA.getSparseBlock().pos(i), inB.getSparseBlock().pos(i),
inA.getSparseBlock().size(i), inB.getSparseBlock().size(i));
+ else
+ ret1 = (double[])
me.invoke(null, n, inA.getSparseBlock().values(i),
inB.getSparseBlock().values(i),
+
inA.getSparseBlock().indexes(i), inB.getSparseBlock().indexes(i),
inA.getSparseBlock().pos(i), inB.getSparseBlock().pos(i),
inA.getSparseBlock().size(i), inB.getSparseBlock().size(i));
+
+ if(sparse) {
+ int[] indexes = retX.indexes();
+ for (int j = 0; j < retX.size(); j++) {
+ ret1[indexes[j]] =
retX.get(indexes[j]);
+ }
+ }
+
+ //execute comparison operation
+ String opcode =
OpOp2.valueOf(bintype.name().split("_")[1]).toString();
+ MatrixBlock in1 = inA.slice(i, i, 0, n-1, new
MatrixBlock());
+ MatrixBlock in2 = inB.slice(i, i, 0, n-1, new
MatrixBlock());
+ double[] ret2 = null;
+ if( type1 == InputType.SCALAR ) {
+ ScalarOperator bop =
InstructionUtils.parseScalarBinaryOperator(opcode, true);
+ bop = bop.setConstant(testType >= 0 &&
i == m-1 ? testType == 0 ? 0 : inA.min() : inA.max());
+ ret2 =
DataConverter.convertToDoubleVector(
+ in2.scalarOperations(bop, new
MatrixBlock()), false);
+ }
+ else if( type2 == InputType.SCALAR ) {
+ ScalarOperator bop =
InstructionUtils.parseScalarBinaryOperator(opcode, false);
+ bop = bop.setConstant(testType >= 0 &&
i == m-1 ? testType == 0 ? 0 : inB.min() : inB.max());
+ ret2 =
DataConverter.convertToDoubleVector(
+ in1.scalarOperations(bop, new
MatrixBlock()), false);
+ }
+ else { //vector-vector
+ BinaryOperator bop =
InstructionUtils.parseBinaryOperator(opcode);
+ ret2 =
DataConverter.convertToDoubleVector(
+ in1.binaryOperations(bop, in2,
new MatrixBlock()), false);
+ }
+
+ //compare results
+ TestUtils.compareMatrices(ret2, ret1, eps);
+ }
+ }
+ catch( Exception ex ) {
+ throw new RuntimeException(ex);
+ }
+ }
+
+
+ /**
+ * @param type
+ * @return {@code true}, when the matching method has a sparse return
+ */
+ private static boolean getOutputType(BinType type) {
+ switch(type) {
+ case VECT_EQUAL:
+ case VECT_LESSEQUAL:
+ case VECT_GREATEREQUAL:
+ return false;
+ default:
+ return true;
+ }
+ }
+
+ /**
+ * @param type
+ * @return returns {@code -1}, for normal testing;<br>
+ * returns {@code 0}, for testing with 0 and non-zeros;<br>
+ * returns {@code 1}, for testing with negative and positive
numbers;
+ */
+ private static int getTestType(BinType type) {
+ switch(type) {
+ case VECT_DIV_SCALAR:
+ case VECT_EQUAL_SCALAR:
+ case VECT_NOTEQUAL_SCALAR:
+ case VECT_XOR_SCALAR:
+ // case VECT_POW_SCALAR:
+ return 0;
+ case VECT_GREATER_SCALAR:
+ case VECT_GREATEREQUAL_SCALAR:
+ case VECT_LESS_SCALAR:
+ case VECT_LESSEQUAL_SCALAR:
+ case VECT_MIN_SCALAR:
+ case VECT_MAX_SCALAR:
+ return 1;
+ default:
+ return -1;
+ }
+ }
}
diff --git
a/src/test/java/org/apache/sysds/test/component/codegen/SparseVectorAllocTest.java
b/src/test/java/org/apache/sysds/test/component/codegen/SparseVectorAllocTest.java
new file mode 100644
index 0000000000..e777e4ef6d
--- /dev/null
+++
b/src/test/java/org/apache/sysds/test/component/codegen/SparseVectorAllocTest.java
@@ -0,0 +1,133 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+package org.apache.sysds.test.component.codegen;
+
+import org.apache.sysds.runtime.codegen.LibSpoofPrimitives;
+import org.apache.sysds.runtime.data.SparseRowVector;
+import org.apache.sysds.test.AutomatedTestBase;
+import org.apache.sysds.test.TestUtils;
+import org.junit.Assert;
+import org.junit.Test;
+
+/**
+ * This is the component test for the ring buffer used in
LibSpoofPrimitives.java,
+ * that allocates a vector with a certain size.
+ * Every allocation method is tested to achieve the needed coverage.
+ */
+public class SparseVectorAllocTest extends AutomatedTestBase
+{
+ double[] val1 = new double[]{1.5, 5.7, 9.1, 3.7, 5.3};
+ double[] val2 = new double[]{9.6, 7.1, 2.7};
+ int[] indexes1 = new int[]{3, 7, 14, 20, 81};
+ int[] indexes2 = new int[]{20, 30, 90};
+
+ @Override
+ public void setUp() {
+ TestUtils.clearAssertionInformation();
+ }
+
+ @Test
+ public void testBasicAllocationSameLen() {
+ testBasicSparseVectorAllocation(1, 10, 10);
+ }
+
+ @Test
+ public void testBasicAllocationLongerExp() {
+ testBasicSparseVectorAllocation(1, 10, 15);
+ }
+
+ @Test
+ public void testBasicAllocationShorterExp() {
+ testBasicSparseVectorAllocation(1, 10, 7);
+ }
+
+ @Test
+ public void testVectorReuse1() {
+ testBufferReuse(3, 10, 5, 5);
+ }
+
+ @Test
+ public void testVectorReuse2() {
+ testBufferReuse(3, 10, -1, 5);
+ }
+
+ /** tests the allocation of an empty vector
+ * @param numVectors number of vectors that should be pre-allocated
+ * @param len the length of the vector
+ * @param expLen the expected length of the allocated vector
+ */
+ public void testBasicSparseVectorAllocation(int numVectors, int len,
int expLen) {
+ //test the basic allocation of an empty vector
+ LibSpoofPrimitives.setupSparseThreadLocalMemory(numVectors,
len, -1);
+ SparseRowVector sparseVec =
LibSpoofPrimitives.allocSparseVector(expLen);
+
+ Assert.assertTrue("Vector capacity should be initialized
correctly", expLen <= sparseVec.capacity());
+ Assert.assertEquals("Vector size should be initialized with 0",
0, sparseVec.size());
+
+ LibSpoofPrimitives.cleanupSparseThreadLocalMemory();
+ }
+
+ /** tests the allocation of a vector that is reused multiple times
+ * @param numVectors number of vectors that should be pre-allocated
+ * @param len1 length of the first vector
+ * @param len2 length of the second vector
+ * @param expLen expected length of allocated vector
+ */
+ public void testBufferReuse(int numVectors, int len1, int len2, int
expLen) {
+ //test the reuse of the vectors in the ring buffer
+ LibSpoofPrimitives.setupSparseThreadLocalMemory(numVectors,
len1, len2);
+
+ //allocate first vector
+ SparseRowVector vec1 =
LibSpoofPrimitives.allocSparseVector(expLen);
+ vec1.set(0, 1.0);
+ vec1.set(2, 2.0);
+
+ //allocate second vector
+ SparseRowVector vec2 =
LibSpoofPrimitives.allocSparseVector(expLen);
+
+ Assert.assertEquals("Reused vector should be reset to size 0",
0, vec2.size());
+
+ for(int j = 0; j < vec2.size(); j++) {
+ vec2.set(vec2.indexes()[j], vec2.get(vec2.indexes()[j])
* 32);
+ }
+
+ SparseRowVector vec3 =
LibSpoofPrimitives.allocSparseVector(expLen);
+
+ Assert.assertEquals("Reused vector should be reset to size 0",
0, vec3.size());
+
+ SparseRowVector vec4 =
LibSpoofPrimitives.allocSparseVector(expLen);
+
+ for(int j = 0; j < vec4.size(); j++) {
+ vec4.set(vec4.indexes()[j], vec4.get(vec3.indexes()[j])
* 32);
+ }
+
+ Assert.assertEquals("Reused vector should be reset to size 0",
0, vec4.size());
+
+ SparseRowVector vec5 =
LibSpoofPrimitives.allocSparseVector(expLen);
+
+ for(int j = 0; j < vec5.size(); j++) {
+ vec2.set(vec5.indexes()[j], vec5.get(vec5.indexes()[j])
* 32);
+ }
+
+ Assert.assertEquals("Reused vector should be reset to size 0",
0, vec5.size());
+
+ LibSpoofPrimitives.cleanupSparseThreadLocalMemory();
+ }
+}
diff --git
a/src/test/java/org/apache/sysds/test/functions/codegen/RowAggTmplTest.java
b/src/test/java/org/apache/sysds/test/functions/codegen/RowAggTmplTest.java
index 73dbace862..03fcb40bef 100644
--- a/src/test/java/org/apache/sysds/test/functions/codegen/RowAggTmplTest.java
+++ b/src/test/java/org/apache/sysds/test/functions/codegen/RowAggTmplTest.java
@@ -89,6 +89,8 @@ public class RowAggTmplTest extends AutomatedTestBase
private static final String TEST_NAME46 = TEST_NAME+"46"; //conv2d(X -
mean(X), F1) + conv2d(X - mean(X), F2);
private static final String TEST_NAME47 = TEST_NAME+"47"; //sum(X +
rowVars(X))
private static final String TEST_NAME48 = TEST_NAME+"48";
//sum(rowVars(X))
+ private static final String TEST_NAME49 = TEST_NAME+"49";
//X*rowSums(K*v)*X
+ private static final String TEST_NAME50 = TEST_NAME+"50";
//(abs(A)*B)+(B*v)
private static final String TEST_DIR = "functions/codegen/";
private static final String TEST_CLASS_DIR = TEST_DIR +
RowAggTmplTest.class.getSimpleName() + "/";
@@ -100,7 +102,7 @@ public class RowAggTmplTest extends AutomatedTestBase
@Override
public void setUp() {
TestUtils.clearAssertionInformation();
- for(int i=1; i<=48; i++)
+ for(int i=1; i<=50; i++)
addTestConfiguration( TEST_NAME+i, new
TestConfiguration(TEST_CLASS_DIR, TEST_NAME+i, new String[] { String.valueOf(i)
}) );
}
@@ -829,6 +831,12 @@ public class RowAggTmplTest extends AutomatedTestBase
testCodegenIntegration( TEST_NAME48, false, ExecType.SPARK );
}
+ @Test
+ public void testCodegenRowAgg49CP() {testCodegenIntegration(
TEST_NAME49, false, ExecType.CP );}
+
+ @Test
+ public void testCodegenRowAgg50CP() {testCodegenIntegration(
TEST_NAME50, false, ExecType.CP );}
+
private void testCodegenIntegration( String testname, boolean rewrites,
ExecType instType )
{
boolean oldFlag = OptimizerUtils.ALLOW_ALGEBRAIC_SIMPLIFICATION;
@@ -841,7 +849,7 @@ public class RowAggTmplTest extends AutomatedTestBase
String HOME = SCRIPT_DIR + TEST_DIR;
fullDMLScriptName = HOME + testname + ".dml";
- programArgs = new String[]{"-explain", "codegen",
"-stats", "-args", output("S") };
+ programArgs = new String[]{"-explain", "codegen",
"-sparseIntermediate", "-stats", "-args", output("S") };
fullRScriptName = HOME + testname + ".R";
rCmd = getRCmd(inputDir(), expectedDir());
diff --git a/src/test/scripts/functions/codegen/rowAggPattern49.R
b/src/test/scripts/functions/codegen/rowAggPattern49.R
new file mode 100644
index 0000000000..06d5c63144
--- /dev/null
+++ b/src/test/scripts/functions/codegen/rowAggPattern49.R
@@ -0,0 +1,54 @@
+#-------------------------------------------------------------
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+#-------------------------------------------------------------
+
+args<-commandArgs(TRUE)
+options(digits=22)
+library("Matrix")
+# library("matrixStats")
+
+W = matrix(seq(28,29), 1, 2)
+J = matrix(0, 1, 8)
+Z = cbind(J, W, J)
+Y = matrix(0, 10, 18)
+X = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+v = seq(1,81)
+v1 = seq(20, 37)
+W = matrix(seq(13,14), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+K = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+
+# S = (X < rowSums(X*K))
+# S = X*rowMins(K)*X
+S = X*rowSums(K*v)*X
+# S = (X*v)/rowSums(X*v)
+# S = abs((X*v)/rowSums(X*v))
+# S = (X/v)+rowMeans(X-v)
+# S = (X*v)+rowSums(X*v)
+# S = (X*rowSums(X*v))/(X*v)
+
+# S = X*rowSums(X*K)
+# S = rowSums((X*v)/K)*v
+# S = (K*v)/(rowSums(X*v))
+
+
+writeMM(as(S, "CsparseMatrix"), paste(args[2], "S", sep=""));
diff --git a/src/test/scripts/functions/codegen/rowAggPattern49.dml
b/src/test/scripts/functions/codegen/rowAggPattern49.dml
new file mode 100644
index 0000000000..8c0d1db225
--- /dev/null
+++ b/src/test/scripts/functions/codegen/rowAggPattern49.dml
@@ -0,0 +1,52 @@
+#-------------------------------------------------------------
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+#-------------------------------------------------------------
+
+W = matrix(seq(28,29), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+X = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+v = seq(1,81)
+v1 = seq(20, 37)
+W = matrix(seq(13,14), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+K = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+while(FALSE) { }
+
+# B = X < rowSums(X*K)
+# S = abs(21) * B
+# S = B * rowSums(v)
+# S = (X < rowSums(X*K))
+# S = X*(k>1)*X
+S = X*rowSums(K*v)*X
+# S = (X*v)/rowSums(X*v)
+# S = abs((X*v)/rowSums(X*v))
+# S = (X/v)+rowMeans(X-v)
+# S = (X*v)+rowSums(X*v)
+# S = (X*rowSums(X*v))/(X*v)
+
+# S = X*rowSums(X*K)
+# S = rowSums((X*v)/K)*v
+# S = (K*v)/(rowSums(X*v))
+
+write(S,$1)
diff --git a/src/test/scripts/functions/codegen/rowAggPattern50.R
b/src/test/scripts/functions/codegen/rowAggPattern50.R
new file mode 100644
index 0000000000..af2c3301a7
--- /dev/null
+++ b/src/test/scripts/functions/codegen/rowAggPattern50.R
@@ -0,0 +1,43 @@
+#-------------------------------------------------------------
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+#-------------------------------------------------------------
+
+args<-commandArgs(TRUE)
+options(digits=22)
+library("Matrix")
+# library("matrixStats")
+
+W = matrix(seq(28,29), 1, 2)
+J = matrix(0, 1, 8)
+Z = cbind(J, W, J)
+Y = matrix(0, 10, 18)
+X = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+v = seq(1,81)
+v1 = seq(20, 37)
+W = matrix(seq(13,14), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+K = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+
+S = (abs(X)*K)+rowSums(X*v)
+
+
+writeMM(as(S, "CsparseMatrix"), paste(args[2], "S", sep=""));
diff --git a/src/test/scripts/functions/codegen/rowAggPattern50.dml
b/src/test/scripts/functions/codegen/rowAggPattern50.dml
new file mode 100644
index 0000000000..64efe460f9
--- /dev/null
+++ b/src/test/scripts/functions/codegen/rowAggPattern50.dml
@@ -0,0 +1,40 @@
+#-------------------------------------------------------------
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+#-------------------------------------------------------------
+
+W = matrix(seq(28,29), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+X = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+v = seq(1,81)
+v1 = seq(20, 37)
+W = matrix(seq(13,14), 1, 2)
+J = matrix(0, 1, 8)
+Z= cbind(J, W, J)
+Y = matrix(0, 10, 18)
+K = rbind(Z, Y, Y, Y, Y, Y, Y, Y, Y)
+while(FALSE) { }
+
+
+S = (abs(X)*K)+rowSums(X*v)
+
+
+write(S,$1)
