Search in sources :

Example 36 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class MultiReturnParameterizedBuiltinSPInstruction method processInstruction.

@Override
@SuppressWarnings("unchecked")
public void processInstruction(ExecutionContext ec) {
    SparkExecutionContext sec = (SparkExecutionContext) ec;
    try {
        // get input RDD and meta data
        FrameObject fo = sec.getFrameObject(input1.getName());
        FrameObject fometa = sec.getFrameObject(_outputs.get(1).getName());
        JavaPairRDD<Long, FrameBlock> in = (JavaPairRDD<Long, FrameBlock>) sec.getRDDHandleForFrameObject(fo, InputInfo.BinaryBlockInputInfo);
        String spec = ec.getScalarInput(input2.getName(), input2.getValueType(), input2.isLiteral()).getStringValue();
        MatrixCharacteristics mcIn = sec.getMatrixCharacteristics(input1.getName());
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        String[] colnames = !TfMetaUtils.isIDSpec(spec) ? in.lookup(1L).get(0).getColumnNames() : null;
        // step 1: build transform meta data
        Encoder encoderBuild = EncoderFactory.createEncoder(spec, colnames, fo.getSchema(), (int) fo.getNumColumns(), null);
        MaxLongAccumulator accMax = registerMaxLongAccumulator(sec.getSparkContext());
        JavaRDD<String> rcMaps = in.mapPartitionsToPair(new TransformEncodeBuildFunction(encoderBuild)).distinct().groupByKey().flatMap(new TransformEncodeGroupFunction(accMax));
        if (containsMVImputeEncoder(encoderBuild)) {
            EncoderMVImpute mva = getMVImputeEncoder(encoderBuild);
            rcMaps = rcMaps.union(in.mapPartitionsToPair(new TransformEncodeBuild2Function(mva)).groupByKey().flatMap(new TransformEncodeGroup2Function(mva)));
        }
        // trigger eval
        rcMaps.saveAsTextFile(fometa.getFileName());
        // consolidate meta data frame (reuse multi-threaded reader, special handling missing values)
        FrameReader reader = FrameReaderFactory.createFrameReader(InputInfo.TextCellInputInfo);
        FrameBlock meta = reader.readFrameFromHDFS(fometa.getFileName(), accMax.value(), fo.getNumColumns());
        // recompute num distinct items per column
        meta.recomputeColumnCardinality();
        meta.setColumnNames((colnames != null) ? colnames : meta.getColumnNames());
        // step 2: transform apply (similar to spark transformapply)
        // compute omit offset map for block shifts
        TfOffsetMap omap = null;
        if (TfMetaUtils.containsOmitSpec(spec, colnames)) {
            omap = new TfOffsetMap(SparkUtils.toIndexedLong(in.mapToPair(new RDDTransformApplyOffsetFunction(spec, colnames)).collect()));
        }
        // create encoder broadcast (avoiding replication per task)
        Encoder encoder = EncoderFactory.createEncoder(spec, colnames, fo.getSchema(), (int) fo.getNumColumns(), meta);
        mcOut.setDimension(mcIn.getRows() - ((omap != null) ? omap.getNumRmRows() : 0), encoder.getNumCols());
        Broadcast<Encoder> bmeta = sec.getSparkContext().broadcast(encoder);
        Broadcast<TfOffsetMap> bomap = (omap != null) ? sec.getSparkContext().broadcast(omap) : null;
        // execute transform apply
        JavaPairRDD<Long, FrameBlock> tmp = in.mapToPair(new RDDTransformApplyFunction(bmeta, bomap));
        JavaPairRDD<MatrixIndexes, MatrixBlock> out = FrameRDDConverterUtils.binaryBlockToMatrixBlock(tmp, mcOut, mcOut);
        // set output and maintain lineage/output characteristics
        sec.setRDDHandleForVariable(_outputs.get(0).getName(), out);
        sec.addLineageRDD(_outputs.get(0).getName(), input1.getName());
        sec.setFrameOutput(_outputs.get(1).getName(), meta);
    } catch (IOException ex) {
        throw new RuntimeException(ex);
    }
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) FrameBlock(org.apache.sysml.runtime.matrix.data.FrameBlock) Encoder(org.apache.sysml.runtime.transform.encode.Encoder) JavaPairRDD(org.apache.spark.api.java.JavaPairRDD) SparkExecutionContext(org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext) EncoderMVImpute(org.apache.sysml.runtime.transform.encode.EncoderMVImpute) RDDTransformApplyOffsetFunction(org.apache.sysml.runtime.instructions.spark.ParameterizedBuiltinSPInstruction.RDDTransformApplyOffsetFunction) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) FrameObject(org.apache.sysml.runtime.controlprogram.caching.FrameObject) IOException(java.io.IOException) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) RDDTransformApplyFunction(org.apache.sysml.runtime.instructions.spark.ParameterizedBuiltinSPInstruction.RDDTransformApplyFunction) TfOffsetMap(org.apache.sysml.runtime.transform.meta.TfOffsetMap) FrameReader(org.apache.sysml.runtime.io.FrameReader)

Example 37 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class PMapmmSPInstruction method processInstruction.

@Override
public void processInstruction(ExecutionContext ec) {
    SparkExecutionContext sec = (SparkExecutionContext) ec;
    // get inputs
    JavaPairRDD<MatrixIndexes, MatrixBlock> in1 = sec.getBinaryBlockRDDHandleForVariable(input1.getName());
    JavaPairRDD<MatrixIndexes, MatrixBlock> in2 = sec.getBinaryBlockRDDHandleForVariable(input2.getName());
    MatrixCharacteristics mc1 = sec.getMatrixCharacteristics(input1.getName());
    // This avoids errors such as java.lang.UnsupportedOperationException: Cannot change storage level of an RDD after it was already assigned a level
    // Ideally, we should ensure that we donot redundantly call persist on the same RDD.
    StorageLevel pmapmmStorageLevel = StorageLevel.MEMORY_AND_DISK();
    // cache right hand side because accessed many times
    in2 = in2.repartition(sec.getSparkContext().defaultParallelism()).persist(pmapmmStorageLevel);
    JavaPairRDD<MatrixIndexes, MatrixBlock> out = null;
    for (int i = 0; i < mc1.getRows(); i += NUM_ROWBLOCKS * mc1.getRowsPerBlock()) {
        // create broadcast for rdd partition
        JavaPairRDD<MatrixIndexes, MatrixBlock> rdd = in1.filter(new IsBlockInRange(i + 1, i + NUM_ROWBLOCKS * mc1.getRowsPerBlock(), 1, mc1.getCols(), mc1)).mapToPair(new PMapMMRebaseBlocksFunction(i / mc1.getRowsPerBlock()));
        int rlen = (int) Math.min(mc1.getRows() - i, NUM_ROWBLOCKS * mc1.getRowsPerBlock());
        PartitionedBlock<MatrixBlock> pmb = SparkExecutionContext.toPartitionedMatrixBlock(rdd, rlen, (int) mc1.getCols(), mc1.getRowsPerBlock(), mc1.getColsPerBlock(), -1L);
        Broadcast<PartitionedBlock<MatrixBlock>> bpmb = sec.getSparkContext().broadcast(pmb);
        // matrix multiplication
        JavaPairRDD<MatrixIndexes, MatrixBlock> rdd2 = in2.flatMapToPair(new PMapMMFunction(bpmb, i / mc1.getRowsPerBlock()));
        rdd2 = RDDAggregateUtils.sumByKeyStable(rdd2, false);
        rdd2.persist(pmapmmStorageLevel).count();
        bpmb.unpersist(false);
        if (out == null)
            out = rdd2;
        else
            out = out.union(rdd2);
    }
    // cache final result
    out = out.persist(pmapmmStorageLevel);
    out.count();
    // put output RDD handle into symbol table
    sec.setRDDHandleForVariable(output.getName(), out);
    sec.addLineageRDD(output.getName(), input1.getName());
    sec.addLineageRDD(output.getName(), input2.getName());
    // update output statistics if not inferred
    updateBinaryMMOutputMatrixCharacteristics(sec, true);
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) IsBlockInRange(org.apache.sysml.runtime.instructions.spark.functions.IsBlockInRange) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) PartitionedBlock(org.apache.sysml.runtime.instructions.spark.data.PartitionedBlock) SparkExecutionContext(org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext) StorageLevel(org.apache.spark.storage.StorageLevel)

Example 38 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class ParameterizedBuiltinSPInstruction method processInstruction.

@Override
@SuppressWarnings("unchecked")
public void processInstruction(ExecutionContext ec) {
    SparkExecutionContext sec = (SparkExecutionContext) ec;
    String opcode = getOpcode();
    // opcode guaranteed to be a valid opcode (see parsing)
    if (opcode.equalsIgnoreCase("mapgroupedagg")) {
        // get input rdd handle
        String targetVar = params.get(Statement.GAGG_TARGET);
        String groupsVar = params.get(Statement.GAGG_GROUPS);
        JavaPairRDD<MatrixIndexes, MatrixBlock> target = sec.getBinaryBlockRDDHandleForVariable(targetVar);
        PartitionedBroadcast<MatrixBlock> groups = sec.getBroadcastForVariable(groupsVar);
        MatrixCharacteristics mc1 = sec.getMatrixCharacteristics(targetVar);
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        CPOperand ngrpOp = new CPOperand(params.get(Statement.GAGG_NUM_GROUPS));
        int ngroups = (int) sec.getScalarInput(ngrpOp.getName(), ngrpOp.getValueType(), ngrpOp.isLiteral()).getLongValue();
        // single-block aggregation
        if (ngroups <= mc1.getRowsPerBlock() && mc1.getCols() <= mc1.getColsPerBlock()) {
            // execute map grouped aggregate
            JavaRDD<MatrixBlock> out = target.map(new RDDMapGroupedAggFunction2(groups, _optr, ngroups));
            MatrixBlock out2 = RDDAggregateUtils.sumStable(out);
            // put output block into symbol table (no lineage because single block)
            // this also includes implicit maintenance of matrix characteristics
            sec.setMatrixOutput(output.getName(), out2, getExtendedOpcode());
        } else // multi-block aggregation
        {
            // execute map grouped aggregate
            JavaPairRDD<MatrixIndexes, MatrixBlock> out = target.flatMapToPair(new RDDMapGroupedAggFunction(groups, _optr, ngroups, mc1.getRowsPerBlock(), mc1.getColsPerBlock()));
            out = RDDAggregateUtils.sumByKeyStable(out, false);
            // updated characteristics and handle outputs
            mcOut.set(ngroups, mc1.getCols(), mc1.getRowsPerBlock(), mc1.getColsPerBlock(), -1);
            sec.setRDDHandleForVariable(output.getName(), out);
            sec.addLineageRDD(output.getName(), targetVar);
            sec.addLineageBroadcast(output.getName(), groupsVar);
        }
    } else if (opcode.equalsIgnoreCase("groupedagg")) {
        boolean broadcastGroups = Boolean.parseBoolean(params.get("broadcast"));
        // get input rdd handle
        String groupsVar = params.get(Statement.GAGG_GROUPS);
        JavaPairRDD<MatrixIndexes, MatrixBlock> target = sec.getBinaryBlockRDDHandleForVariable(params.get(Statement.GAGG_TARGET));
        JavaPairRDD<MatrixIndexes, MatrixBlock> groups = broadcastGroups ? null : sec.getBinaryBlockRDDHandleForVariable(groupsVar);
        JavaPairRDD<MatrixIndexes, MatrixBlock> weights = null;
        MatrixCharacteristics mc1 = sec.getMatrixCharacteristics(params.get(Statement.GAGG_TARGET));
        MatrixCharacteristics mc2 = sec.getMatrixCharacteristics(groupsVar);
        if (mc1.dimsKnown() && mc2.dimsKnown() && (mc1.getRows() != mc2.getRows() || mc2.getCols() != 1)) {
            throw new DMLRuntimeException("Grouped Aggregate dimension mismatch between target and groups.");
        }
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        JavaPairRDD<MatrixIndexes, WeightedCell> groupWeightedCells = null;
        // Step 1: First extract groupWeightedCells from group, target and weights
        if (params.get(Statement.GAGG_WEIGHTS) != null) {
            weights = sec.getBinaryBlockRDDHandleForVariable(params.get(Statement.GAGG_WEIGHTS));
            MatrixCharacteristics mc3 = sec.getMatrixCharacteristics(params.get(Statement.GAGG_WEIGHTS));
            if (mc1.dimsKnown() && mc3.dimsKnown() && (mc1.getRows() != mc3.getRows() || mc1.getCols() != mc3.getCols())) {
                throw new DMLRuntimeException("Grouped Aggregate dimension mismatch between target, groups, and weights.");
            }
            groupWeightedCells = groups.join(target).join(weights).flatMapToPair(new ExtractGroupNWeights());
        } else // input vector or matrix
        {
            String ngroupsStr = params.get(Statement.GAGG_NUM_GROUPS);
            long ngroups = (ngroupsStr != null) ? (long) Double.parseDouble(ngroupsStr) : -1;
            // execute basic grouped aggregate (extract and preagg)
            if (broadcastGroups) {
                PartitionedBroadcast<MatrixBlock> pbm = sec.getBroadcastForVariable(groupsVar);
                groupWeightedCells = target.flatMapToPair(new ExtractGroupBroadcast(pbm, mc1.getColsPerBlock(), ngroups, _optr));
            } else {
                // replicate groups if necessary
                if (mc1.getNumColBlocks() > 1) {
                    groups = groups.flatMapToPair(new ReplicateVectorFunction(false, mc1.getNumColBlocks()));
                }
                groupWeightedCells = groups.join(target).flatMapToPair(new ExtractGroupJoin(mc1.getColsPerBlock(), ngroups, _optr));
            }
        }
        // Step 2: Make sure we have brlen required while creating <MatrixIndexes, MatrixCell>
        if (mc1.getRowsPerBlock() == -1) {
            throw new DMLRuntimeException("The block sizes are not specified for grouped aggregate");
        }
        int brlen = mc1.getRowsPerBlock();
        // Step 3: Now perform grouped aggregate operation (either on combiner side or reducer side)
        JavaPairRDD<MatrixIndexes, MatrixCell> out = null;
        if (_optr instanceof CMOperator && ((CMOperator) _optr).isPartialAggregateOperator() || _optr instanceof AggregateOperator) {
            out = groupWeightedCells.reduceByKey(new PerformGroupByAggInCombiner(_optr)).mapValues(new CreateMatrixCell(brlen, _optr));
        } else {
            // Use groupby key because partial aggregation is not supported
            out = groupWeightedCells.groupByKey().mapValues(new PerformGroupByAggInReducer(_optr)).mapValues(new CreateMatrixCell(brlen, _optr));
        }
        // Step 4: Set output characteristics and rdd handle
        setOutputCharacteristicsForGroupedAgg(mc1, mcOut, out);
        // store output rdd handle
        sec.setRDDHandleForVariable(output.getName(), out);
        sec.addLineageRDD(output.getName(), params.get(Statement.GAGG_TARGET));
        sec.addLineage(output.getName(), groupsVar, broadcastGroups);
        if (params.get(Statement.GAGG_WEIGHTS) != null) {
            sec.addLineageRDD(output.getName(), params.get(Statement.GAGG_WEIGHTS));
        }
    } else if (opcode.equalsIgnoreCase("rmempty")) {
        String rddInVar = params.get("target");
        String rddOffVar = params.get("offset");
        boolean rows = sec.getScalarInput(params.get("margin"), ValueType.STRING, true).getStringValue().equals("rows");
        boolean emptyReturn = Boolean.parseBoolean(params.get("empty.return").toLowerCase());
        long maxDim = sec.getScalarInput(params.get("maxdim"), ValueType.DOUBLE, false).getLongValue();
        MatrixCharacteristics mcIn = sec.getMatrixCharacteristics(rddInVar);
        if (// default case
        maxDim > 0) {
            // get input rdd handle
            JavaPairRDD<MatrixIndexes, MatrixBlock> in = sec.getBinaryBlockRDDHandleForVariable(rddInVar);
            JavaPairRDD<MatrixIndexes, MatrixBlock> off;
            PartitionedBroadcast<MatrixBlock> broadcastOff;
            long brlen = mcIn.getRowsPerBlock();
            long bclen = mcIn.getColsPerBlock();
            long numRep = (long) Math.ceil(rows ? (double) mcIn.getCols() / bclen : (double) mcIn.getRows() / brlen);
            // execute remove empty rows/cols operation
            JavaPairRDD<MatrixIndexes, MatrixBlock> out;
            if (_bRmEmptyBC) {
                broadcastOff = sec.getBroadcastForVariable(rddOffVar);
                // Broadcast offset vector
                out = in.flatMapToPair(new RDDRemoveEmptyFunctionInMem(rows, maxDim, brlen, bclen, broadcastOff));
            } else {
                off = sec.getBinaryBlockRDDHandleForVariable(rddOffVar);
                out = in.join(off.flatMapToPair(new ReplicateVectorFunction(!rows, numRep))).flatMapToPair(new RDDRemoveEmptyFunction(rows, maxDim, brlen, bclen));
            }
            out = RDDAggregateUtils.mergeByKey(out, false);
            // store output rdd handle
            sec.setRDDHandleForVariable(output.getName(), out);
            sec.addLineageRDD(output.getName(), rddInVar);
            if (!_bRmEmptyBC)
                sec.addLineageRDD(output.getName(), rddOffVar);
            else
                sec.addLineageBroadcast(output.getName(), rddOffVar);
            // update output statistics (required for correctness)
            MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
            mcOut.set(rows ? maxDim : mcIn.getRows(), rows ? mcIn.getCols() : maxDim, (int) brlen, (int) bclen, mcIn.getNonZeros());
        } else // special case: empty output (ensure valid dims)
        {
            int n = emptyReturn ? 1 : 0;
            MatrixBlock out = new MatrixBlock(rows ? n : (int) mcIn.getRows(), rows ? (int) mcIn.getCols() : n, true);
            sec.setMatrixOutput(output.getName(), out, getExtendedOpcode());
        }
    } else if (opcode.equalsIgnoreCase("replace")) {
        // get input rdd handle
        String rddVar = params.get("target");
        JavaPairRDD<MatrixIndexes, MatrixBlock> in1 = sec.getBinaryBlockRDDHandleForVariable(rddVar);
        MatrixCharacteristics mcIn = sec.getMatrixCharacteristics(rddVar);
        // execute replace operation
        double pattern = Double.parseDouble(params.get("pattern"));
        double replacement = Double.parseDouble(params.get("replacement"));
        JavaPairRDD<MatrixIndexes, MatrixBlock> out = in1.mapValues(new RDDReplaceFunction(pattern, replacement));
        // store output rdd handle
        sec.setRDDHandleForVariable(output.getName(), out);
        sec.addLineageRDD(output.getName(), rddVar);
        // update output statistics (required for correctness)
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        mcOut.set(mcIn.getRows(), mcIn.getCols(), mcIn.getRowsPerBlock(), mcIn.getColsPerBlock(), (pattern != 0 && replacement != 0) ? mcIn.getNonZeros() : -1);
    } else if (opcode.equalsIgnoreCase("rexpand")) {
        String rddInVar = params.get("target");
        // get input rdd handle
        JavaPairRDD<MatrixIndexes, MatrixBlock> in = sec.getBinaryBlockRDDHandleForVariable(rddInVar);
        MatrixCharacteristics mcIn = sec.getMatrixCharacteristics(rddInVar);
        double maxVal = Double.parseDouble(params.get("max"));
        long lmaxVal = UtilFunctions.toLong(maxVal);
        boolean dirRows = params.get("dir").equals("rows");
        boolean cast = Boolean.parseBoolean(params.get("cast"));
        boolean ignore = Boolean.parseBoolean(params.get("ignore"));
        long brlen = mcIn.getRowsPerBlock();
        long bclen = mcIn.getColsPerBlock();
        // repartition input vector for higher degree of parallelism
        // (avoid scenarios where few input partitions create huge outputs)
        MatrixCharacteristics mcTmp = new MatrixCharacteristics(dirRows ? lmaxVal : mcIn.getRows(), dirRows ? mcIn.getRows() : lmaxVal, (int) brlen, (int) bclen, mcIn.getRows());
        int numParts = (int) Math.min(SparkUtils.getNumPreferredPartitions(mcTmp, in), mcIn.getNumBlocks());
        if (numParts > in.getNumPartitions() * 2)
            in = in.repartition(numParts);
        // execute rexpand rows/cols operation (no shuffle required because outputs are
        // block-aligned with the input, i.e., one input block generates n output blocks)
        JavaPairRDD<MatrixIndexes, MatrixBlock> out = in.flatMapToPair(new RDDRExpandFunction(maxVal, dirRows, cast, ignore, brlen, bclen));
        // store output rdd handle
        sec.setRDDHandleForVariable(output.getName(), out);
        sec.addLineageRDD(output.getName(), rddInVar);
        // update output statistics (required for correctness)
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        mcOut.set(dirRows ? lmaxVal : mcIn.getRows(), dirRows ? mcIn.getRows() : lmaxVal, (int) brlen, (int) bclen, -1);
    } else if (opcode.equalsIgnoreCase("transformapply")) {
        // get input RDD and meta data
        FrameObject fo = sec.getFrameObject(params.get("target"));
        JavaPairRDD<Long, FrameBlock> in = (JavaPairRDD<Long, FrameBlock>) sec.getRDDHandleForFrameObject(fo, InputInfo.BinaryBlockInputInfo);
        FrameBlock meta = sec.getFrameInput(params.get("meta"));
        MatrixCharacteristics mcIn = sec.getMatrixCharacteristics(params.get("target"));
        MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
        String[] colnames = !TfMetaUtils.isIDSpec(params.get("spec")) ? in.lookup(1L).get(0).getColumnNames() : null;
        // compute omit offset map for block shifts
        TfOffsetMap omap = null;
        if (TfMetaUtils.containsOmitSpec(params.get("spec"), colnames)) {
            omap = new TfOffsetMap(SparkUtils.toIndexedLong(in.mapToPair(new RDDTransformApplyOffsetFunction(params.get("spec"), colnames)).collect()));
        }
        // create encoder broadcast (avoiding replication per task)
        Encoder encoder = EncoderFactory.createEncoder(params.get("spec"), colnames, fo.getSchema(), (int) fo.getNumColumns(), meta);
        mcOut.setDimension(mcIn.getRows() - ((omap != null) ? omap.getNumRmRows() : 0), encoder.getNumCols());
        Broadcast<Encoder> bmeta = sec.getSparkContext().broadcast(encoder);
        Broadcast<TfOffsetMap> bomap = (omap != null) ? sec.getSparkContext().broadcast(omap) : null;
        // execute transform apply
        JavaPairRDD<Long, FrameBlock> tmp = in.mapToPair(new RDDTransformApplyFunction(bmeta, bomap));
        JavaPairRDD<MatrixIndexes, MatrixBlock> out = FrameRDDConverterUtils.binaryBlockToMatrixBlock(tmp, mcOut, mcOut);
        // set output and maintain lineage/output characteristics
        sec.setRDDHandleForVariable(output.getName(), out);
        sec.addLineageRDD(output.getName(), params.get("target"));
        ec.releaseFrameInput(params.get("meta"));
    } else if (opcode.equalsIgnoreCase("transformdecode")) {
        // get input RDD and meta data
        JavaPairRDD<MatrixIndexes, MatrixBlock> in = sec.getBinaryBlockRDDHandleForVariable(params.get("target"));
        MatrixCharacteristics mc = sec.getMatrixCharacteristics(params.get("target"));
        FrameBlock meta = sec.getFrameInput(params.get("meta"));
        String[] colnames = meta.getColumnNames();
        // reblock if necessary (clen > bclen)
        if (mc.getCols() > mc.getNumColBlocks()) {
            in = in.mapToPair(new RDDTransformDecodeExpandFunction((int) mc.getCols(), mc.getColsPerBlock()));
            in = RDDAggregateUtils.mergeByKey(in, false);
        }
        // construct decoder and decode individual matrix blocks
        Decoder decoder = DecoderFactory.createDecoder(params.get("spec"), colnames, null, meta);
        JavaPairRDD<Long, FrameBlock> out = in.mapToPair(new RDDTransformDecodeFunction(decoder, mc.getRowsPerBlock()));
        // set output and maintain lineage/output characteristics
        sec.setRDDHandleForVariable(output.getName(), out);
        sec.addLineageRDD(output.getName(), params.get("target"));
        ec.releaseFrameInput(params.get("meta"));
        sec.getMatrixCharacteristics(output.getName()).set(mc.getRows(), meta.getNumColumns(), mc.getRowsPerBlock(), mc.getColsPerBlock(), -1);
        sec.getFrameObject(output.getName()).setSchema(decoder.getSchema());
    } else {
        throw new DMLRuntimeException("Unknown parameterized builtin opcode: " + opcode);
    }
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) ExtractGroupNWeights(org.apache.sysml.runtime.instructions.spark.functions.ExtractGroupNWeights) ReplicateVectorFunction(org.apache.sysml.runtime.instructions.spark.functions.ReplicateVectorFunction) Decoder(org.apache.sysml.runtime.transform.decode.Decoder) PartitionedBroadcast(org.apache.sysml.runtime.instructions.spark.data.PartitionedBroadcast) FrameBlock(org.apache.sysml.runtime.matrix.data.FrameBlock) ExtractGroupBroadcast(org.apache.sysml.runtime.instructions.spark.functions.ExtractGroup.ExtractGroupBroadcast) PartitionedBroadcast(org.apache.sysml.runtime.instructions.spark.data.PartitionedBroadcast) Broadcast(org.apache.spark.broadcast.Broadcast) Encoder(org.apache.sysml.runtime.transform.encode.Encoder) JavaPairRDD(org.apache.spark.api.java.JavaPairRDD) AggregateOperator(org.apache.sysml.runtime.matrix.operators.AggregateOperator) SparkExecutionContext(org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) PerformGroupByAggInReducer(org.apache.sysml.runtime.instructions.spark.functions.PerformGroupByAggInReducer) CPOperand(org.apache.sysml.runtime.instructions.cp.CPOperand) FrameObject(org.apache.sysml.runtime.controlprogram.caching.FrameObject) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) DMLRuntimeException(org.apache.sysml.runtime.DMLRuntimeException) ExtractGroupBroadcast(org.apache.sysml.runtime.instructions.spark.functions.ExtractGroup.ExtractGroupBroadcast) TfOffsetMap(org.apache.sysml.runtime.transform.meta.TfOffsetMap) PerformGroupByAggInCombiner(org.apache.sysml.runtime.instructions.spark.functions.PerformGroupByAggInCombiner) ExtractGroupJoin(org.apache.sysml.runtime.instructions.spark.functions.ExtractGroup.ExtractGroupJoin) CMOperator(org.apache.sysml.runtime.matrix.operators.CMOperator)

Example 39 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class QuantileSortSPInstruction method processInstruction.

@Override
public void processInstruction(ExecutionContext ec) {
    SparkExecutionContext sec = (SparkExecutionContext) ec;
    boolean weighted = (input2 != null);
    // get input rdds
    JavaPairRDD<MatrixIndexes, MatrixBlock> in = sec.getBinaryBlockRDDHandleForVariable(input1.getName());
    JavaPairRDD<MatrixIndexes, MatrixBlock> inW = weighted ? sec.getBinaryBlockRDDHandleForVariable(input2.getName()) : null;
    MatrixCharacteristics mc = sec.getMatrixCharacteristics(input1.getName());
    JavaPairRDD<MatrixIndexes, MatrixBlock> out = null;
    long clen = -1;
    if (!weighted) {
        // W/O WEIGHTS (default)
        out = RDDSortUtils.sortByVal(in, mc.getRows(), mc.getRowsPerBlock());
        clen = 1;
    } else {
        // W/ WEIGHTS
        out = RDDSortUtils.sortByVal(in, inW, mc.getRows(), mc.getRowsPerBlock());
        clen = 2;
    }
    // put output RDD handle into symbol table
    sec.setRDDHandleForVariable(output.getName(), out);
    sec.addLineageRDD(output.getName(), input1.getName());
    if (weighted)
        sec.addLineageRDD(output.getName(), input2.getName());
    // update output matrix characteristics
    MatrixCharacteristics mcOut = sec.getMatrixCharacteristics(output.getName());
    mcOut.set(mc.getRows(), clen, mc.getRowsPerBlock(), mc.getColsPerBlock());
}
Also used : MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) SparkExecutionContext(org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics)

Example 40 with MatrixBlock

use of org.apache.sysml.runtime.matrix.data.MatrixBlock in project incubator-systemml by apache.

the class QuaternarySPInstruction method processInstruction.

@Override
public void processInstruction(ExecutionContext ec) {
    SparkExecutionContext sec = (SparkExecutionContext) ec;
    QuaternaryOperator qop = (QuaternaryOperator) _optr;
    // tracking of rdds and broadcasts (for lineage maintenance)
    ArrayList<String> rddVars = new ArrayList<>();
    ArrayList<String> bcVars = new ArrayList<>();
    JavaPairRDD<MatrixIndexes, MatrixBlock> in = sec.getBinaryBlockRDDHandleForVariable(input1.getName());
    JavaPairRDD<MatrixIndexes, MatrixBlock> out = null;
    MatrixCharacteristics inMc = sec.getMatrixCharacteristics(input1.getName());
    long rlen = inMc.getRows();
    long clen = inMc.getCols();
    int brlen = inMc.getRowsPerBlock();
    int bclen = inMc.getColsPerBlock();
    // (map/redwsloss, map/redwcemm); safe because theses ops produce a scalar
    if (qop.wtype1 != null || qop.wtype4 != null) {
        in = in.filter(new FilterNonEmptyBlocksFunction());
    }
    // map-side only operation (one rdd input, two broadcasts)
    if (WeightedSquaredLoss.OPCODE.equalsIgnoreCase(getOpcode()) || WeightedSigmoid.OPCODE.equalsIgnoreCase(getOpcode()) || WeightedDivMM.OPCODE.equalsIgnoreCase(getOpcode()) || WeightedCrossEntropy.OPCODE.equalsIgnoreCase(getOpcode()) || WeightedUnaryMM.OPCODE.equalsIgnoreCase(getOpcode())) {
        PartitionedBroadcast<MatrixBlock> bc1 = sec.getBroadcastForVariable(input2.getName());
        PartitionedBroadcast<MatrixBlock> bc2 = sec.getBroadcastForVariable(input3.getName());
        // partitioning-preserving mappartitions (key access required for broadcast loopkup)
        // only wdivmm changes keys
        boolean noKeyChange = (qop.wtype3 == null || qop.wtype3.isBasic());
        out = in.mapPartitionsToPair(new RDDQuaternaryFunction1(qop, bc1, bc2), noKeyChange);
        rddVars.add(input1.getName());
        bcVars.add(input2.getName());
        bcVars.add(input3.getName());
    } else // reduce-side operation (two/three/four rdd inputs, zero/one/two broadcasts)
    {
        PartitionedBroadcast<MatrixBlock> bc1 = _cacheU ? sec.getBroadcastForVariable(input2.getName()) : null;
        PartitionedBroadcast<MatrixBlock> bc2 = _cacheV ? sec.getBroadcastForVariable(input3.getName()) : null;
        JavaPairRDD<MatrixIndexes, MatrixBlock> inU = (!_cacheU) ? sec.getBinaryBlockRDDHandleForVariable(input2.getName()) : null;
        JavaPairRDD<MatrixIndexes, MatrixBlock> inV = (!_cacheV) ? sec.getBinaryBlockRDDHandleForVariable(input3.getName()) : null;
        JavaPairRDD<MatrixIndexes, MatrixBlock> inW = (qop.hasFourInputs() && !_input4.isLiteral()) ? sec.getBinaryBlockRDDHandleForVariable(_input4.getName()) : null;
        // preparation of transposed and replicated U
        if (inU != null)
            inU = inU.flatMapToPair(new ReplicateBlockFunction(clen, bclen, true));
        // preparation of transposed and replicated V
        if (inV != null)
            inV = inV.mapToPair(new TransposeFactorIndexesFunction()).flatMapToPair(new ReplicateBlockFunction(rlen, brlen, false));
        // functions calls w/ two rdd inputs
        if (inU != null && inV == null && inW == null)
            out = in.join(inU).mapToPair(new RDDQuaternaryFunction2(qop, bc1, bc2));
        else if (inU == null && inV != null && inW == null)
            out = in.join(inV).mapToPair(new RDDQuaternaryFunction2(qop, bc1, bc2));
        else if (inU == null && inV == null && inW != null)
            out = in.join(inW).mapToPair(new RDDQuaternaryFunction2(qop, bc1, bc2));
        else // function calls w/ three rdd inputs
        if (inU != null && inV != null && inW == null)
            out = in.join(inU).join(inV).mapToPair(new RDDQuaternaryFunction3(qop, bc1, bc2));
        else if (inU != null && inV == null && inW != null)
            out = in.join(inU).join(inW).mapToPair(new RDDQuaternaryFunction3(qop, bc1, bc2));
        else if (inU == null && inV != null && inW != null)
            out = in.join(inV).join(inW).mapToPair(new RDDQuaternaryFunction3(qop, bc1, bc2));
        else if (inU == null && inV == null && inW == null) {
            out = in.mapPartitionsToPair(new RDDQuaternaryFunction1(qop, bc1, bc2), false);
        } else
            // function call w/ four rdd inputs
            // need keys in case of wdivmm
            out = in.join(inU).join(inV).join(inW).mapToPair(new RDDQuaternaryFunction4(qop));
        // keep variable names for lineage maintenance
        if (inU == null)
            bcVars.add(input2.getName());
        else
            rddVars.add(input2.getName());
        if (inV == null)
            bcVars.add(input3.getName());
        else
            rddVars.add(input3.getName());
        if (inW != null)
            rddVars.add(_input4.getName());
    }
    // output handling, incl aggregation
    if (// map/redwsloss, map/redwcemm
    qop.wtype1 != null || qop.wtype4 != null) {
        // full aggregate and cast to scalar
        MatrixBlock tmp = RDDAggregateUtils.sumStable(out);
        DoubleObject ret = new DoubleObject(tmp.getValue(0, 0));
        sec.setVariable(output.getName(), ret);
    } else // map/redwsigmoid, map/redwdivmm, map/redwumm
    {
        // aggregation if required (map/redwdivmm)
        if (qop.wtype3 != null && !qop.wtype3.isBasic())
            out = RDDAggregateUtils.sumByKeyStable(out, false);
        // put output RDD handle into symbol table
        sec.setRDDHandleForVariable(output.getName(), out);
        // maintain lineage information for output rdd
        for (String rddVar : rddVars) sec.addLineageRDD(output.getName(), rddVar);
        for (String bcVar : bcVars) sec.addLineageBroadcast(output.getName(), bcVar);
        // update matrix characteristics
        updateOutputMatrixCharacteristics(sec, qop);
    }
}
Also used : QuaternaryOperator(org.apache.sysml.runtime.matrix.operators.QuaternaryOperator) FilterNonEmptyBlocksFunction(org.apache.sysml.runtime.instructions.spark.functions.FilterNonEmptyBlocksFunction) MatrixBlock(org.apache.sysml.runtime.matrix.data.MatrixBlock) MatrixIndexes(org.apache.sysml.runtime.matrix.data.MatrixIndexes) DoubleObject(org.apache.sysml.runtime.instructions.cp.DoubleObject) ArrayList(java.util.ArrayList) MatrixCharacteristics(org.apache.sysml.runtime.matrix.MatrixCharacteristics) SparkExecutionContext(org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext) ReplicateBlockFunction(org.apache.sysml.runtime.instructions.spark.functions.ReplicateBlockFunction)

Aggregations

MatrixBlock (org.apache.sysml.runtime.matrix.data.MatrixBlock)459 MatrixIndexes (org.apache.sysml.runtime.matrix.data.MatrixIndexes)142 DMLRuntimeException (org.apache.sysml.runtime.DMLRuntimeException)111 MatrixCharacteristics (org.apache.sysml.runtime.matrix.MatrixCharacteristics)102 CompressedMatrixBlock (org.apache.sysml.runtime.compress.CompressedMatrixBlock)48 SparkExecutionContext (org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext)48 IOException (java.io.IOException)44 MatrixObject (org.apache.sysml.runtime.controlprogram.caching.MatrixObject)41 ArrayList (java.util.ArrayList)40 Path (org.apache.hadoop.fs.Path)29 FrameBlock (org.apache.sysml.runtime.matrix.data.FrameBlock)24 FileSystem (org.apache.hadoop.fs.FileSystem)23 JavaPairRDD (org.apache.spark.api.java.JavaPairRDD)23 JobConf (org.apache.hadoop.mapred.JobConf)21 Tuple2 (scala.Tuple2)19 SequenceFile (org.apache.hadoop.io.SequenceFile)17 Row (org.apache.spark.sql.Row)14 SparseBlock (org.apache.sysml.runtime.matrix.data.SparseBlock)14 TestConfiguration (org.apache.sysml.test.integration.TestConfiguration)14 IndexedMatrixValue (org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue)13