Examples with LongPointerWrapper org.nd4j.nativeblas.LongPointerWrapper used on opensource projects

Example 6 with LongPointerWrapper

use of org.nd4j.nativeblas.LongPointerWrapper in project nd4j by deeplearning4j.

the class JCublasNDArrayFactory method tear.

public INDArray[] tear(INDArray tensor, int... dimensions) {
    if (tensor.isCompressed())
        Nd4j.getCompressor().decompressi(tensor);
    Arrays.sort(dimensions);
    Pair<DataBuffer, DataBuffer> tadBuffers = Nd4j.getExecutioner().getTADManager().getTADOnlyShapeInfo(tensor, dimensions);
    long tadLength = 1;
    int[] shape = new int[dimensions.length];
    for (int i = 0; i < dimensions.length; i++) {
        tadLength *= tensor.shape()[dimensions[i]];
        shape[i] = tensor.shape()[dimensions[i]];
    }
    int numTads = (int) (tensor.lengthLong() / tadLength);
    INDArray[] result = new INDArray[numTads];
    long[] xPointers = new long[numTads];
    CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(null, tensor);
    for (int x = 0; x < numTads; x++) {
        result[x] = Nd4j.createUninitialized(shape);
        context = AtomicAllocator.getInstance().getFlowController().prepareAction(result[x]);
        xPointers[x] = AtomicAllocator.getInstance().getPointer(result[x], context).address();
    }
    CudaDoubleDataBuffer tempX = new CudaDoubleDataBuffer(numTads);
    AtomicAllocator.getInstance().memcpyBlocking(tempX, new LongPointer(xPointers), xPointers.length * 8, 0);
    PointerPointer extraz = new // not used
    PointerPointer(// not used
    null, context.getOldStream(), AtomicAllocator.getInstance().getDeviceIdPointer());
    if (Nd4j.dataType() == DataBuffer.Type.DOUBLE) {
        nativeOps.tearDouble(extraz, (DoublePointer) AtomicAllocator.getInstance().getPointer(tensor, context), (IntPointer) AtomicAllocator.getInstance().getPointer(tensor.shapeInfoDataBuffer(), context), new PointerPointer(AtomicAllocator.getInstance().getPointer(tempX, context)), (IntPointer) AtomicAllocator.getInstance().getPointer(result[0].shapeInfoDataBuffer(), context), (IntPointer) AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context), new LongPointerWrapper(AtomicAllocator.getInstance().getPointer(tadBuffers.getSecond(), context)));
    } else if (Nd4j.dataType() == DataBuffer.Type.FLOAT) {
        nativeOps.tearFloat(extraz, (FloatPointer) AtomicAllocator.getInstance().getPointer(tensor, context), (IntPointer) AtomicAllocator.getInstance().getPointer(tensor.shapeInfoDataBuffer(), context), new PointerPointer(AtomicAllocator.getInstance().getPointer(tempX, context)), (IntPointer) AtomicAllocator.getInstance().getPointer(result[0].shapeInfoDataBuffer(), context), (IntPointer) AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context), new LongPointerWrapper(AtomicAllocator.getInstance().getPointer(tadBuffers.getSecond(), context)));
    } else if (Nd4j.dataType() == DataBuffer.Type.HALF) {
        nativeOps.tearHalf(extraz, (ShortPointer) AtomicAllocator.getInstance().getPointer(tensor, context), (IntPointer) AtomicAllocator.getInstance().getPointer(tensor.shapeInfoDataBuffer(), context), new PointerPointer(AtomicAllocator.getInstance().getPointer(tempX, context)), (IntPointer) AtomicAllocator.getInstance().getPointer(result[0].shapeInfoDataBuffer(), context), (IntPointer) AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context), new LongPointerWrapper(AtomicAllocator.getInstance().getPointer(tadBuffers.getSecond(), context)));
    }
    AtomicAllocator.getInstance().getFlowController().registerActionAllWrite(context, result);
    AtomicAllocator.getInstance().getFlowController().registerAction(context, null, result);
    return result;
}

Example 7 with LongPointerWrapper

use of org.nd4j.nativeblas.LongPointerWrapper in project nd4j by deeplearning4j.

the class JCublasNDArrayFactory method pullRows.

/**
 * This method produces concatenated array, that consist from tensors, fetched from source array, against some dimension and specified indexes
 *
 * @param source          source tensor
 * @param sourceDimension dimension of source tensor
 * @param indexes         indexes from source array
 * @return
 */
@Override
public INDArray pullRows(INDArray source, int sourceDimension, int[] indexes, char order) {
    if (Nd4j.getExecutioner() instanceof GridExecutioner)
        ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
    if (indexes == null || indexes.length < 1)
        throw new IllegalStateException("Indexes can't be null or zero-length");
    int[] shape = null;
    if (sourceDimension == 1)
        shape = new int[] { indexes.length, source.shape()[sourceDimension] };
    else if (sourceDimension == 0)
        shape = new int[] { source.shape()[sourceDimension], indexes.length };
    else
        throw new UnsupportedOperationException("2D input is expected");
    INDArray ret = Nd4j.createUninitialized(shape, order);
    AtomicAllocator allocator = AtomicAllocator.getInstance();
    CudaContext context = allocator.getFlowController().prepareAction(ret, source);
    Pointer x = AtomicAllocator.getInstance().getPointer(source, context);
    Pointer xShape = AtomicAllocator.getInstance().getPointer(source.shapeInfoDataBuffer(), context);
    Pointer z = AtomicAllocator.getInstance().getPointer(ret, context);
    Pointer zShape = AtomicAllocator.getInstance().getPointer(ret.shapeInfoDataBuffer(), context);
    PointerPointer extras = new PointerPointer(AddressRetriever.retrieveHostPointer(ret.shapeInfoDataBuffer()), context.getOldStream(), allocator.getDeviceIdPointer());
    CudaIntDataBuffer tempIndexes = new CudaIntDataBuffer(indexes.length);
    AtomicAllocator.getInstance().memcpyBlocking(tempIndexes, new IntPointer(indexes), indexes.length * 4, 0);
    Pointer pIndex = AtomicAllocator.getInstance().getPointer(tempIndexes, context);
    TADManager tadManager = Nd4j.getExecutioner().getTADManager();
    Pair<DataBuffer, DataBuffer> tadBuffers = tadManager.getTADOnlyShapeInfo(source, new int[] { sourceDimension });
    Pair<DataBuffer, DataBuffer> zTadBuffers = tadManager.getTADOnlyShapeInfo(ret, new int[] { sourceDimension });
    Pointer tadShapeInfo = AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context);
    Pointer zTadShapeInfo = AtomicAllocator.getInstance().getPointer(zTadBuffers.getFirst(), context);
    DataBuffer offsets = tadBuffers.getSecond();
    Pointer tadOffsets = AtomicAllocator.getInstance().getPointer(offsets, context);
    Pointer zTadOffsets = AtomicAllocator.getInstance().getPointer(zTadBuffers.getSecond(), context);
    if (ret.data().dataType() == DataBuffer.Type.DOUBLE) {
        nativeOps.pullRowsDouble(extras, (DoublePointer) x, (IntPointer) xShape, (DoublePointer) z, (IntPointer) zShape, indexes.length, (IntPointer) pIndex, (IntPointer) tadShapeInfo, new LongPointerWrapper(tadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    } else if (ret.data().dataType() == DataBuffer.Type.FLOAT) {
        nativeOps.pullRowsFloat(extras, (FloatPointer) x, (IntPointer) xShape, (FloatPointer) z, (IntPointer) zShape, indexes.length, (IntPointer) pIndex, (IntPointer) tadShapeInfo, new LongPointerWrapper(tadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    } else {
        nativeOps.pullRowsHalf(extras, (ShortPointer) x, (IntPointer) xShape, (ShortPointer) z, (IntPointer) zShape, indexes.length, (IntPointer) pIndex, (IntPointer) tadShapeInfo, new LongPointerWrapper(tadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    }
    allocator.registerAction(context, ret, source);
    return ret;
}

Example 8 with LongPointerWrapper

use of org.nd4j.nativeblas.LongPointerWrapper in project nd4j by deeplearning4j.

the class CudaGridExecutioner method exec.

@Override
public void exec(MetaOp op) {
    if (extraz.get() == null)
        extraz.set(new PointerPointer(32));
    prepareGrid(op);
    GridPointers first = op.getGridDescriptor().getGridPointers().get(0);
    GridPointers second = op.getGridDescriptor().getGridPointers().get(1);
    // we need to use it only for first op, since for MetaOps second op shares the same X & Z by definition
    CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(first.getOpZ(), first.getOpY());
    // AtomicAllocator.getInstance().getFlowController().prepareAction(second.getOpX(), second.getOpY(), second.getOpZ());
    // CudaContext context = (CudaContext) AtomicAllocator.getInstance().getDeviceContext().getContext();
    PointerPointer extras = extraz.get().put(null, context.getOldStream());
    double scalarA = 0.0;
    double scalarB = 0.0;
    if (op.getFirstOp() instanceof ScalarOp)
        scalarA = ((ScalarOp) op.getFirstOp()).scalar().doubleValue();
    if (op.getSecondOp() instanceof ScalarOp)
        scalarB = ((ScalarOp) op.getSecondOp()).scalar().doubleValue();
    // logger.info("FirstOp: {}, SecondOp: {}", op.getFirstOp().getClass().getSimpleName(), op.getSecondOp().getClass().getSimpleName());
    /*
            TODO: launch can be either strided, or shapeInfo-based, it doesn't really matters for us.
            We just need to pass all pointers.
        
            TODO: obviously, execMetaPredicateElementwiseFloat should be renamed to execMetaPredicateStridedFloat
         */
    // FIXME: this is bad hack, reconsider this one
    GridPointers yGrid = first;
    if (op.getSecondOp().y() != null) {
        yGrid = second;
    }
    if (op instanceof PredicateMetaOp || op instanceof InvertedPredicateMetaOp) {
        if (first.getDtype() == DataBuffer.Type.FLOAT) {
            if (yGrid.getYOrder() == yGrid.getXOrder() && yGrid.getXStride() >= 1 && yGrid.getYStride() >= 1) {
                nativeOps.execMetaPredicateStridedFloat(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), // can be null
                (FloatPointer) first.getX(), // can be null
                first.getXStride(), // can be null
                (FloatPointer) yGrid.getY(), // cane be -1
                yGrid.getYStride(), (FloatPointer) second.getZ(), second.getZStride(), (FloatPointer) first.getExtraArgs(), (FloatPointer) second.getExtraArgs(), (float) scalarA, (float) scalarB);
            } else {
                nativeOps.execMetaPredicateShapeFloat(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), (FloatPointer) first.getX(), (IntPointer) first.getXShapeInfo(), // can be null
                (FloatPointer) yGrid.getY(), // cane be -1
                (IntPointer) yGrid.getYShapeInfo(), (FloatPointer) second.getZ(), (IntPointer) second.getZShapeInfo(), (FloatPointer) first.getExtraArgs(), (FloatPointer) second.getExtraArgs(), (float) scalarA, (float) scalarB);
            }
        } else if (first.getDtype() == DataBuffer.Type.DOUBLE) {
            if (yGrid.getYOrder() == yGrid.getXOrder() && yGrid.getXStride() >= 1 && yGrid.getYStride() >= 1) {
                nativeOps.execMetaPredicateStridedDouble(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), // can be null
                (DoublePointer) first.getX(), // can be null
                first.getXStride(), // can be null
                (DoublePointer) yGrid.getY(), // cane be -1
                yGrid.getYStride(), (DoublePointer) second.getZ(), second.getZStride(), (DoublePointer) first.getExtraArgs(), (DoublePointer) second.getExtraArgs(), scalarA, scalarB);
            } else {
                nativeOps.execMetaPredicateShapeDouble(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), (DoublePointer) first.getX(), (IntPointer) first.getXShapeInfo(), // can be null
                (DoublePointer) yGrid.getY(), // cane be -1
                (IntPointer) yGrid.getYShapeInfo(), (DoublePointer) second.getZ(), (IntPointer) second.getZShapeInfo(), (DoublePointer) first.getExtraArgs(), (DoublePointer) second.getExtraArgs(), scalarA, scalarB);
            }
        } else {
            if (yGrid.getYOrder() == yGrid.getXOrder() && yGrid.getXStride() >= 1 && yGrid.getYStride() >= 1) {
                nativeOps.execMetaPredicateStridedHalf(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), // can be null
                (ShortPointer) first.getX(), // can be null
                first.getXStride(), // can be null
                (ShortPointer) yGrid.getY(), // cane be -1
                yGrid.getYStride(), (ShortPointer) second.getZ(), second.getZStride(), (ShortPointer) first.getExtraArgs(), (ShortPointer) second.getExtraArgs(), (float) scalarA, (float) scalarB);
            } else {
                nativeOps.execMetaPredicateShapeHalf(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), first.getXLength(), (ShortPointer) first.getX(), (IntPointer) first.getXShapeInfo(), // can be null
                (ShortPointer) yGrid.getY(), // cane be -1
                (IntPointer) yGrid.getYShapeInfo(), (ShortPointer) second.getZ(), (IntPointer) second.getZShapeInfo(), (ShortPointer) first.getExtraArgs(), (ShortPointer) second.getExtraArgs(), (float) scalarA, (float) scalarB);
            }
        }
    } else if (op instanceof ReduceMetaOp) {
        if (first.getDtype() == DataBuffer.Type.FLOAT) {
            nativeOps.execMetaPredicateReduceFloat(extras, first.getType().ordinal(), first.getOpNum(), second.getType().ordinal(), second.getOpNum(), (FloatPointer) first.getX(), (IntPointer) first.getXShapeInfo(), (FloatPointer) second.getY(), (IntPointer) second.getYShapeInfo(), (FloatPointer) second.getZ(), (IntPointer) second.getZShapeInfo(), (IntPointer) second.getDimensions(), second.getDimensionsLength(), (IntPointer) second.getTadShape(), new LongPointerWrapper(second.getTadOffsets()), (FloatPointer) first.getExtraArgs(), (FloatPointer) second.getExtraArgs(), (float) scalarA, 0.0f, false);
        }
    }
    AtomicAllocator.getInstance().getFlowController().registerAction(context, first.getOpZ(), first.getOpY());
// AtomicAllocator.getInstance().getFlowController().registerAction(context, second.getOpX(), second.getOpY(), second.getOpZ());
}

Example 9 with LongPointerWrapper

use of org.nd4j.nativeblas.LongPointerWrapper in project nd4j by deeplearning4j.

the class CpuNDArrayFactory method pullRows.

/**
 * This method produces concatenated array, that consist from tensors, fetched from source array, against some dimension and specified indexes
 *
 * @param source          source tensor
 * @param sourceDimension dimension of source tensor
 * @param indexes         indexes from source array
 * @return
 */
@Override
public INDArray pullRows(INDArray source, int sourceDimension, int[] indexes, char order) {
    if (indexes == null || indexes.length < 1)
        throw new IllegalStateException("Indexes can't be null or zero-length");
    int[] shape = null;
    if (sourceDimension == 1)
        shape = new int[] { indexes.length, source.shape()[sourceDimension] };
    else if (sourceDimension == 0)
        shape = new int[] { source.shape()[sourceDimension], indexes.length };
    else
        throw new UnsupportedOperationException("2D input is expected");
    INDArray ret = Nd4j.createUninitialized(shape, order);
    Nd4j.getCompressor().autoDecompress(source);
    PointerPointer dummy = new PointerPointer(new Pointer[] { null });
    TADManager tadManager = Nd4j.getExecutioner().getTADManager();
    Pair<DataBuffer, DataBuffer> tadBuffers = tadManager.getTADOnlyShapeInfo(source, new int[] { sourceDimension });
    Pair<DataBuffer, DataBuffer> zTadBuffers = tadManager.getTADOnlyShapeInfo(ret, new int[] { sourceDimension });
    Pointer hostTadShapeInfo = tadBuffers.getFirst().addressPointer();
    Pointer zTadShapeInfo = zTadBuffers.getFirst().addressPointer();
    IntPointer pIndex = new IntPointer(indexes);
    DataBuffer offsets = tadBuffers.getSecond();
    Pointer hostTadOffsets = offsets == null ? null : offsets.addressPointer();
    DataBuffer zOffsets = zTadBuffers.getSecond();
    Pointer zTadOffsets = zOffsets == null ? null : zOffsets.addressPointer();
    if (ret.data().dataType() == DataBuffer.Type.DOUBLE) {
        nativeOps.pullRowsDouble(dummy, (DoublePointer) source.data().addressPointer(), (IntPointer) source.shapeInfoDataBuffer().addressPointer(), (DoublePointer) ret.data().addressPointer(), (IntPointer) ret.shapeInfoDataBuffer().addressPointer(), indexes.length, pIndex, (IntPointer) hostTadShapeInfo, new LongPointerWrapper(hostTadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    } else if (ret.data().dataType() == DataBuffer.Type.FLOAT) {
        nativeOps.pullRowsFloat(dummy, (FloatPointer) source.data().addressPointer(), (IntPointer) source.shapeInfoDataBuffer().addressPointer(), (FloatPointer) ret.data().addressPointer(), (IntPointer) ret.shapeInfoDataBuffer().addressPointer(), indexes.length, pIndex, (IntPointer) hostTadShapeInfo, new LongPointerWrapper(hostTadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    } else {
        nativeOps.pullRowsHalf(dummy, (ShortPointer) source.data().addressPointer(), (IntPointer) source.shapeInfoDataBuffer().addressPointer(), (ShortPointer) ret.data().addressPointer(), (IntPointer) ret.shapeInfoDataBuffer().addressPointer(), indexes.length, pIndex, (IntPointer) hostTadShapeInfo, new LongPointerWrapper(hostTadOffsets), (IntPointer) zTadShapeInfo, new LongPointerWrapper(zTadOffsets));
    }
    return ret;
}

Example 10 with LongPointerWrapper

use of org.nd4j.nativeblas.LongPointerWrapper in project nd4j by deeplearning4j.

the class CpuNDArrayFactory method tear.

public INDArray[] tear(INDArray tensor, int... dimensions) {
    if (tensor.isCompressed())
        Nd4j.getCompressor().decompressi(tensor);
    Arrays.sort(dimensions);
    Pair<DataBuffer, DataBuffer> tadBuffers = Nd4j.getExecutioner().getTADManager().getTADOnlyShapeInfo(tensor, dimensions);
    long tadLength = 1;
    int[] shape = new int[dimensions.length];
    for (int i = 0; i < dimensions.length; i++) {
        tadLength *= tensor.shape()[dimensions[i]];
        shape[i] = tensor.shape()[dimensions[i]];
    }
    int numTads = (int) (tensor.lengthLong() / tadLength);
    INDArray[] result = new INDArray[numTads];
    PointerPointer targets = new PointerPointer(numTads);
    for (int x = 0; x < numTads; x++) {
        result[x] = Nd4j.createUninitialized(shape);
        targets.put(x, result[x].data().pointer());
    }
    if (Nd4j.dataType() == DataBuffer.Type.DOUBLE) {
        nativeOps.tearDouble(null, (DoublePointer) tensor.data().pointer(), (IntPointer) tensor.shapeInfoDataBuffer().pointer(), targets, (IntPointer) result[0].shapeInfoDataBuffer().pointer(), (IntPointer) tadBuffers.getFirst().pointer(), new LongPointerWrapper(tadBuffers.getSecond().pointer()));
    } else if (Nd4j.dataType() == DataBuffer.Type.FLOAT) {
        nativeOps.tearFloat(null, (FloatPointer) tensor.data().pointer(), (IntPointer) tensor.shapeInfoDataBuffer().pointer(), targets, (IntPointer) result[0].shapeInfoDataBuffer().pointer(), (IntPointer) tadBuffers.getFirst().pointer(), new LongPointerWrapper(tadBuffers.getSecond().pointer()));
    } else if (Nd4j.dataType() == DataBuffer.Type.HALF) {
        throw new UnsupportedOperationException("Half precision isn't supported for CPU backend");
    }
    return result;
}