[WIP] CUDA Java side (#58)

* one crashing test

Signed-off-by: raver119 <raver119@gmail.com>

* stupid issue fixed

Signed-off-by: raver119 <raver119@gmail.com>

* one fix

Signed-off-by: raver119 <raver119@gmail.com>

* dont ensure location for empty arrays

Signed-off-by: raver119 <raver119@gmail.com>

* few more signatures fixed

Signed-off-by: raver119 <raver119@gmail.com>

* few tweaks for DataBuffer creation from java primitives

Signed-off-by: raver119 <raver119@gmail.com>

* get rid of legacy im2col/col2im intercept

Signed-off-by: raver119 <raver119@gmail.com>

* rsubi scalar array fix

Signed-off-by: raver119 <raver119@gmail.com>

libnd4j/include/array/cuda/DataBuffer.cu

@@ -37,7 +37,6 @@ void DataBuffer::allocateSpecial() {
 
 ////////////////////////////////////////////////////////////////////////
 void DataBuffer::syncToPrimary(const LaunchContext* context, const bool forceSync) {
-
     if(isPrimaryActual() && !forceSync)
         return;
 

libnd4j/include/ops/declarable/helpers/impl/choose.cpp

@@ -93,8 +93,12 @@ namespace helpers {
         if (comp != nullptr)
             comp->syncToHost();
 
-        output->syncToHost();
+        if (output != nullptr)
-        numResult->syncToHost();
+            output->syncToHost();
+
+        if (numResult != nullptr)
+            numResult->syncToHost();
+
         compScalar.syncToHost();
 
         BUILD_SINGLE_SELECTOR(arg->dataType(), return processCondition_, (mode, arg, comp, output, numResult, compScalar), FLOAT_TYPES);
@@ -104,8 +108,12 @@ namespace helpers {
         if (comp != nullptr)
             comp->syncToDevice();
 
-        output->syncToDevice();
+        if (output != nullptr)
-        numResult->syncToDevice();
+            output->syncToDevice();
+
+        if (numResult != nullptr)
+            numResult->syncToDevice();
+        
         compScalar.syncToDevice();
 
     }

libnd4j/tests_cpu/layers_tests/DeclarableOpsTestsCuda1.cu

@@ -0,0 +1,62 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://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.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+
+//
+// @author raver119@gmail.com
+//
+
+#include "testlayers.h"
+#include <ops/declarable/CustomOperations.h>
+#include <NDArray.h>
+#include <ops/ops.h>
+#include <GradCheck.h>
+
+
+using namespace nd4j;
+
+
+class DeclarableOpsTestsCuda1 : public testing::Test {
+public:
+
+    DeclarableOpsTestsCuda1() {
+        printf("\n");
+        fflush(stdout);
+    }
+};
+
+
+TEST_F(DeclarableOpsTestsCuda1, Test_CHOOSE_SCALAR_LARGE) {
+    double inputData[150] = {
+            0,  0.51,  0.68,  0.69,  0.86,  0.91,  0.96,  0.97,  0.97,  1.03,  1.13,  1.16,  1.16,  1.17,  1.19,  1.25,  1.25,  1.26,  1.27,  1.28,  1.29,  1.29,  1.29,  1.30,  1.31,  1.32,  1.33,  1.33,  1.35,  1.35,  1.36,  1.37,  1.38,  1.40,  1.41,  1.42,  1.43,  1.44,  1.44,  1.45,  1.45,  1.47,  1.47,  1.51,  1.51,  1.51,  1.52,  1.53,  1.56,  1.57,  1.58,  1.59,  1.61,  1.62,  1.63,  1.63,  1.64,  1.64,  1.66,  1.66,  1.67,  1.67,  1.70,  1.70,  1.70,  1.72,  1.72,  1.72,  1.72,  1.73,  1.74,  1.74,  1.76,  1.76,  1.77,  1.77,  1.80,  1.80,  1.81,  1.82,  1.83,  1.83,  1.84,  1.84,  1.84,  1.85,  1.85,  1.85,  1.86,  1.86,  1.87,  1.88,  1.89,  1.89,  1.89,  1.89,  1.89,  1.91,  1.91,  1.91,  1.92,  1.94,  1.95,  1.97,  1.98,  1.98,  1.98,  1.98,  1.98,  1.99,  2,  2,  2.01,  2.01,  2.02,  2.03,  2.03,  2.03,  2.04,  2.04,  2.05,  2.06,  2.07,  2.08,  2.08,  2.08,  2.08,  2.09,  2.09,  2.10,  2.10,  2.11,  2.11,  2.11,  2.12,  2.12,  2.13,  2.13,  2.14,  2.14,  2.14,  2.14,  2.15,  2.15,  2.16,  2.16,  2.16,  2.16,  2.16,  2.17
+    };
+
+    auto precursor = NDArrayFactory::create<double>(inputData,'c',{1,149});
+    NDArray x(nullptr, precursor.specialBuffer(), precursor.shapeInfo());
+
+    nd4j::ops::choose op;
+    //greater than test
+    auto result = op.execute({&x}, {0.0},{3});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(1);
+
+    ASSERT_EQ(148,z->e<double>(0));
+    //ASSERT_TRUE(exp.isSameShape(z));
+
+    delete result;
+
+}

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java

@@ -4003,7 +4003,7 @@ public abstract class BaseNDArray implements INDArray, Iterable {
     public INDArray rsubi(INDArray other, INDArray result) {
         validateNumericalArray("rsubi", false);
         if (other.isScalar()) {
-            return this.addi(other.getDouble(0), result);
+            return this.rsubi(other.getDouble(0), result);
         }
 
         if (isScalar()) {

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/concurrency/CudaAffinityManager.java

@@ -379,7 +379,11 @@ public class CudaAffinityManager extends BasicAffinityManager {
 
     @Override
     public void ensureLocation(INDArray array, Location location) {
-        AllocationPoint point = AtomicAllocator.getInstance().getAllocationPoint(array);
+        // to location to ensure for empty array
+        if (array.isEmpty())
+            return;
+
+        val point = AtomicAllocator.getInstance().getAllocationPoint(array);
         switch (location) {
             case HOST: {
                     AtomicAllocator.getInstance().synchronizeHostData(array);
@@ -399,7 +403,10 @@ public class CudaAffinityManager extends BasicAffinityManager {
 
     @Override
     public Location getActiveLocation(INDArray array) {
-        AllocationPoint point = AtomicAllocator.getInstance().getAllocationPoint(array);
+        if (array.isEmpty())
+            return Location.EVERYWHERE;
+
+        val point = AtomicAllocator.getInstance().getAllocationPoint(array);
 
         if (point.isActualOnDeviceSide() && point.isActualOnHostSide()) {
             return Location.EVERYWHERE;

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/flow/impl/SynchronousFlowController.java

@@ -18,6 +18,7 @@ package org.nd4j.jita.flow.impl;
 
 
 import lombok.Getter;
+import lombok.NonNull;
 import lombok.val;
 import org.bytedeco.javacpp.DoublePointer;
 import org.nd4j.jita.allocator.Allocator;
@@ -95,7 +96,7 @@ public class SynchronousFlowController implements FlowController {
     }
 
     @Override
-    public void synchronizeToDevice(AllocationPoint point) {
+    public void synchronizeToDevice(@NonNull AllocationPoint point) {
         if (point.isConstant())
             return;
 

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/buffer/BaseCudaDataBuffer.java

@@ -567,6 +567,14 @@ public abstract class BaseCudaDataBuffer extends BaseDataBuffer implements JCuda
         return allocationPoint.getPointers().getHostPointer().address();
     }
 
+    @Override
+    public Pointer pointer() {
+        // FIXME: very bad thing,
+        lazyAllocateHostPointer();
+
+        return super.pointer();
+    }
+
 
     /**
      *
@@ -672,7 +680,6 @@ public abstract class BaseCudaDataBuffer extends BaseDataBuffer implements JCuda
         }
     }
 
-
     public void set(long[] data, long length, long srcOffset, long dstOffset) {
         // TODO: make sure getPointer returns proper pointer
 
@@ -698,11 +705,14 @@ public abstract class BaseCudaDataBuffer extends BaseDataBuffer implements JCuda
                 }
                 break;
             case UBYTE: {
+                data = ArrayUtil.cutBelowZero(data);
                     for (int e = 0; e < data.length; e++) {
                         put(e, data[e]);
                     }
                 }
                 break;
+            case UINT16:
+                data = ArrayUtil.cutBelowZero(data);
             case SHORT: {
                     val pointer = new ShortPointer(ArrayUtil.toShorts(data));
                     val srcPtr = new CudaPointer(pointer.address() + (dstOffset * elementSize));
@@ -714,6 +724,7 @@ public abstract class BaseCudaDataBuffer extends BaseDataBuffer implements JCuda
                 }
                 break;
             case UINT32:
+                data = ArrayUtil.cutBelowZero(data);
             case INT: {
                     val pointer = new IntPointer(ArrayUtil.toInts(data));
                     val srcPtr = new CudaPointer(pointer.address() + (dstOffset * elementSize));
@@ -725,6 +736,7 @@ public abstract class BaseCudaDataBuffer extends BaseDataBuffer implements JCuda
                 }
                 break;
             case UINT64:
+                data = ArrayUtil.cutBelowZero(data);
             case LONG: {
                     val pointer = new LongPointer(data);
                     val srcPtr = new CudaPointer(pointer.address() + (dstOffset * elementSize));

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/buffer/CudaBfloat16DataBuffer.java

@@ -187,6 +187,11 @@ public class CudaBfloat16DataBuffer extends BaseCudaDataBuffer {
         setData(ArrayUtil.toShorts(data));
     }
 
+    @Override
+    public void setData(long[] data) {
+        setData(ArrayUtil.toShorts(data));
+    }
+
 
 
     @Override

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/buffer/factory/CudaDataBufferFactory.java

@@ -414,6 +414,8 @@ public class CudaDataBufferFactory implements DataBufferFactory {
                 return new CudaFloatDataBuffer(length, initialize, workspace);
             case HALF:
                 return new CudaHalfDataBuffer(length, initialize, workspace);
+            case BFLOAT16:
+                return new CudaBfloat16DataBuffer(length, initialize, workspace);
             case BOOL:
                 return new CudaBoolDataBuffer(length, initialize, workspace);
             default:

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java

@@ -61,6 +61,7 @@ import org.nd4j.linalg.exception.ND4JIllegalArgumentException;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.jcublas.buffer.AddressRetriever;
+import org.nd4j.linalg.jcublas.buffer.BaseCudaDataBuffer;
 import org.nd4j.linalg.jcublas.buffer.CudaLongDataBuffer;
 import org.nd4j.linalg.jcublas.context.CudaContext;
 import org.nd4j.linalg.primitives.AtomicBoolean;
@@ -1495,7 +1496,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
     @Override
     public <T extends Aggregate> void exec(Batch<T> batch) {
-        DataBuffer surfaceBuffer = getBuffer(batch);
+        val surfaceBuffer = (BaseCudaDataBuffer) getBuffer(batch);
+        surfaceBuffer.lazyAllocateHostPointer();
 
         CudaContext context = (CudaContext) AtomicAllocator.getInstance().getDeviceContext().getContext();
 
@@ -2238,152 +2240,6 @@ public class CudaExecutioner extends DefaultOpExecutioner {
             }
         }
 
-        if (op.opName().equalsIgnoreCase("im2col")) {
-            val xArr = op.inputArguments()[0];
-            val zArr = op.outputArguments()[0];
-
-            CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(zArr, xArr);
-
-            if (extraz.get() == null)
-                extraz.set(new PointerPointer(32));
-
-            PointerPointer xShapeHost =
-                    extraz.get().put(AddressRetriever.retrieveHostPointer(xArr.shapeInfoDataBuffer()), // 0
-                            context.getOldStream(), // 1
-                            AtomicAllocator.getInstance().getDeviceIdPointer(), // 2
-                            context.getBufferAllocation(), // 3
-                            context.getBufferReduction(), // 4
-                            context.getBufferScalar(), // 5
-                            context.getBufferSpecial(),
-                            null,
-                            AddressRetriever.retrieveHostPointer(zArr.shapeInfoDataBuffer())
-                    );
-
-
-            val x = AtomicAllocator.getInstance().getPointer(xArr, context);
-            val z = AtomicAllocator.getInstance().getPointer(zArr, context);
-
-            val xShape = AtomicAllocator.getInstance().getPointer(xArr.shapeInfoDataBuffer(), context);
-            val zShape = AtomicAllocator.getInstance().getPointer(zArr.shapeInfoDataBuffer(), context);
-
-            val hxShape = AtomicAllocator.getInstance().getHostPointer(xArr.shapeInfoDataBuffer());
-            val hzShape = AtomicAllocator.getInstance().getHostPointer(zArr.shapeInfoDataBuffer());
-
-            double zeroPad = 0.0;
-            if(op.tArgs() != null && op.tArgs().length > 0){
-                zeroPad = op.tArgs()[0];
-            }
-            val extrass = new double[]{op.iArgs()[0], op.iArgs()[1], op.iArgs()[2], op.iArgs()[3], op.iArgs()[4], op.iArgs()[5], op.iArgs()[6], op.iArgs()[7], op.iArgs()[8], zeroPad};
-            val extraArgsBuff = Nd4j.getConstantHandler().getConstantBuffer(extrass, xArr.dataType());
-            val extraArgs = AtomicAllocator.getInstance().getPointer(extraArgsBuff, context);
-
-            nativeOps.execTransformSame(xShapeHost, 9,
-                    null, (LongPointer) hxShape, x, (LongPointer) xShape,
-                    null, (LongPointer) hzShape, z, (LongPointer) zShape, extraArgs);
-
-            //AtomicAllocator.getInstance().getAllocationPoint(zArr).tickDeviceWrite();
-            AtomicAllocator.getInstance().getFlowController().registerAction(context, zArr, xArr);
-
-            Nd4j.getExecutioner().commit();
-
-            return op.outputArguments();
-        } else if (op.opName().equalsIgnoreCase("col2im")) {
-            val dtype = Nd4j.dataType();
-
-            val xArr = op.inputArguments()[0];
-            val zArr = op.outputArguments()[0];
-
-            CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(zArr, xArr);
-
-            if (extraz.get() == null)
-                extraz.set(new PointerPointer(32));
-
-            PointerPointer xShapeHost =
-                    extraz.get().put(AddressRetriever.retrieveHostPointer(xArr.shapeInfoDataBuffer()), // 0
-                            context.getOldStream(), // 1
-                            AtomicAllocator.getInstance().getDeviceIdPointer(), // 2
-                            context.getBufferAllocation(), // 3
-                            context.getBufferReduction(), // 4
-                            context.getBufferScalar(), // 5
-                            context.getBufferSpecial(),
-                            null,
-                            AddressRetriever.retrieveHostPointer(zArr.shapeInfoDataBuffer())
-                    );
-
-
-            val x = AtomicAllocator.getInstance().getPointer(xArr, context);
-            val z = AtomicAllocator.getInstance().getPointer(zArr, context);
-
-            val xShape = AtomicAllocator.getInstance().getPointer(xArr.shapeInfoDataBuffer(), context);
-            val zShape = AtomicAllocator.getInstance().getPointer(zArr.shapeInfoDataBuffer(), context);
-
-            val hxShape = AtomicAllocator.getInstance().getHostPointer(xArr.shapeInfoDataBuffer());
-            val hzShape = AtomicAllocator.getInstance().getHostPointer(zArr.shapeInfoDataBuffer());
-
-            val extrass = new double[]{op.iArgs()[0], op.iArgs()[1], op.iArgs()[2], op.iArgs()[3], op.iArgs()[4], op.iArgs()[5], op.iArgs()[6], op.iArgs()[7]};
-            val extraArgsBuff = Nd4j.getConstantHandler().getConstantBuffer(extrass, xArr.dataType());
-            val extraArgs = AtomicAllocator.getInstance().getPointer(extraArgsBuff, context);
-
-
-            nativeOps.execTransformSame(xShapeHost, 8,
-                    null, (LongPointer) hxShape, x, (LongPointer) xShape,
-                    null, (LongPointer) hzShape, z, (LongPointer) zShape, extraArgs);
-
-            //AtomicAllocator.getInstance().getAllocationPoint(zArr).tickDeviceWrite();
-            AtomicAllocator.getInstance().getFlowController().registerAction(context, zArr, xArr);
-
-            //Nd4j.getExecutioner().commit();
-            return op.outputArguments();
-        } else if (op.opName().equalsIgnoreCase("pooling2d")) {
-            val dtype = Nd4j.dataType();
-
-            val xArr = op.inputArguments()[0];
-            val zArr = op.outputArguments()[0];
-
-            CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(zArr, xArr);
-
-            if (extraz.get() == null)
-                extraz.set(new PointerPointer(32));
-
-            PointerPointer xShapeHost =
-                    extraz.get().put(AddressRetriever.retrieveHostPointer(xArr.shapeInfoDataBuffer()), // 0
-                            context.getOldStream(), // 1
-                            AtomicAllocator.getInstance().getDeviceIdPointer(), // 2
-                            context.getBufferAllocation(), // 3
-                            context.getBufferReduction(), // 4
-                            context.getBufferScalar(), // 5
-                            context.getBufferSpecial(),
-                            null,
-                            AddressRetriever.retrieveHostPointer(zArr.shapeInfoDataBuffer())
-                    );
-
-
-            val x = AtomicAllocator.getInstance().getPointer(xArr, context);
-            val z = AtomicAllocator.getInstance().getPointer(zArr, context);
-
-            val xShape = AtomicAllocator.getInstance().getPointer(xArr.shapeInfoDataBuffer(), context);
-            val zShape = AtomicAllocator.getInstance().getPointer(zArr.shapeInfoDataBuffer(), context);
-
-            val hxShape = AtomicAllocator.getInstance().getHostPointer(xArr.shapeInfoDataBuffer());
-            val hzShape = AtomicAllocator.getInstance().getHostPointer(zArr.shapeInfoDataBuffer());
-
-            val extrass = new double[]{op.iArgs()[0], op.iArgs()[1], op.iArgs()[2], op.iArgs()[3], op.iArgs()[4], op.iArgs()[5], op.iArgs()[6], op.iArgs()[7], op.iArgs()[8]};
-            val extraArgsBuff = Nd4j.getConstantHandler().getConstantBuffer(extrass, zArr.dataType());
-            val extraArgs = AtomicAllocator.getInstance().getPointer(extraArgsBuff, context);
-
-
-            nativeOps.execTransformFloat(xShapeHost, 23,
-                    null, (LongPointer) hxShape, x, (LongPointer) xShape,
-                    zArr.data().addressPointer(), (LongPointer) hzShape, z, (LongPointer) zShape,
-                    extraArgs);
-
-            // AtomicAllocator.getInstance().getAllocationPoint(zArr).tickDeviceWrite();
-            AtomicAllocator.getInstance().getFlowController().registerAction(context, zArr, xArr);
-
-            return op.outputArguments();
-        }
-
-        Nd4j.getExecutioner().commit();
         val ctx = (CudaContext) AtomicAllocator.getInstance().getDeviceContext().getContext();
 
         val context = (CudaOpContext) buildContext();

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java

@@ -15377,7 +15377,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
        * Input arrays:
        *    0: input 3d tensor with shape [bS x K x N], N - number of time steps, bS - batch size, K - number of features
        *    1: 2d tensor of weights [3K x K]
-       *    2: row of biases with twice length [1 × 2K]
+       *    2: row of biases with twice length [1 x 2K]
        *    3: 2d tensor of previous cell state [bS x K]
        *    4: optional, 2d tensor of dropout mask [bS x K]
        *
@@ -15410,7 +15410,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
        * Input arrays: 
        *    0: input 3d tensor with shape [N x bS x 2K], N - number of time steps, bS - batch size, K - number of features
        *    1: 2d tensor of weights [2K x 6K]
-       *    2: row of biases with twice length [1 × 4K]
+       *    2: row of biases with twice length [1 x 4K]
        *    3: 2d tensor of previous cell state [bS x 2K]
        *    4: optional, 2d tensor of dropout mask [bS x 2K]
        *  
@@ -15444,7 +15444,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
        * Input arrays: 
        *    0: input 3d tensor with shape [bS x K x N], N - number of time steps, bS - batch size, K - number of features
        *    1: 2d tensor of weights [3K x K]
-       *    2: row of biases with twice length [1 × 2K]
+       *    2: row of biases with twice length [1 x 2K]
        *    3: 2d tensor of previous cell state [bS x K]
        *    4: 3d tensor of cell state [bS x K x N]
        *    5: 2d tensor of cell state gradients [bS x K]
@@ -15482,7 +15482,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
        * Input arrays: 
        *    0: input 3d tensor with shape [N x bS x 2K], N - number of time steps, bS - batch size, K - number of features
        *    1: 2d tensor of weights [2K x 6K]
-       *    2: row of biases with twice length [1 × 4K]
+       *    2: row of biases with twice length [1 x 4K]
        *    3: 2d tensor of previous cell state [bS x 2K]
        *    4: 3d tensor of cell state [N x bS x 2K]
        *    5: 2d tensor of cell state gradients [bS x 2K]
@@ -15681,7 +15681,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
        *    0: input with shape [batchSize x inSize], batchSize - batch size, inSize - number of features
        *    1: previous cell state [batchSize x inSize], that is at previous time step t-1
        *    2: weights [inSize x 3*inSize]
-       *    3: biases [1 × 2*inSize]
+       *    3: biases [1 x 2*inSize]
        * 
        * Output arrays: 
        *    0: current cell output [batchSize x inSize], that is at current time step t

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/aggregates/HierarchicSoftmaxTests.java

@@ -23,6 +23,7 @@ import org.junit.Test;
 import org.junit.runner.RunWith;
 import org.junit.runners.Parameterized;
 import org.nd4j.linalg.BaseNd4jTest;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.aggregates.impl.AggregateCBOW;
 import org.nd4j.linalg.api.ops.aggregates.impl.AggregateSkipGram;
@@ -95,17 +96,17 @@ public class HierarchicSoftmaxTests extends BaseNd4jTest {
 
     @Test
     public void testSGGradient1() {
-        INDArray syn0 = Nd4j.create(10, 10).assign(0.01f);
+        INDArray syn0 = Nd4j.create(DataType.DOUBLE, 10, 10).assign(0.01f);
-        INDArray syn1 = Nd4j.create(10, 10).assign(0.02f);
+        INDArray syn1 = Nd4j.create(DataType.DOUBLE,10, 10).assign(0.02f);
-        INDArray syn1Neg = Nd4j.ones(10, 10).assign(0.03f);
+        INDArray syn1Neg = Nd4j.create(DataType.DOUBLE,10, 10).assign(0.03f);
-        INDArray expTable = Nd4j.create(10000).assign(0.5f);
+        INDArray expTable = Nd4j.create(DataType.DOUBLE,10000).assign(0.5f);
 
         double lr = 0.001;
 
         int idxSyn0 = 0;
 
-        INDArray expSyn0 = Nd4j.create(10).assign(0.01001f);
+        INDArray expSyn0 = Nd4j.create(DataType.DOUBLE,10).assign(0.01001f);
-        INDArray expSyn1_1 = Nd4j.create(10).assign(0.020005);
+        INDArray expSyn1_1 = Nd4j.create(DataType.DOUBLE,10).assign(0.020005);
 
         INDArray syn0row = syn0.getRow(idxSyn0);
 

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/buffer/DataBufferTests.java

@@ -24,6 +24,7 @@ import org.junit.Test;
 import org.junit.runner.RunWith;
 import org.junit.runners.Parameterized;
 import org.nd4j.linalg.BaseNd4jTest;
+import org.nd4j.linalg.api.concurrency.AffinityManager;
 import org.nd4j.linalg.api.memory.MemoryWorkspace;
 import org.nd4j.linalg.api.memory.conf.WorkspaceConfiguration;
 import org.nd4j.linalg.api.memory.enums.AllocationPolicy;
@@ -288,6 +289,8 @@ public class DataBufferTests extends BaseNd4jTest {
                     continue;
                 }
 
+                log.info("Testing source [{}]; target: [{}]", sourceType, dt);
+
                 for (boolean useWs : new boolean[]{false, true}) {
 
                     try (MemoryWorkspace ws = (useWs ? workspace.notifyScopeEntered() : null)) {
@@ -334,7 +337,6 @@ public class DataBufferTests extends BaseNd4jTest {
                         assertFalse(db2.isAttached());
 
                         if(!sourceType.equals("boolean")){
-                            log.info("Testing source [{}]; target: [{}]", sourceType, dt);
                             testDBOps(db1);
                             testDBOps(db2);
                         }
@@ -375,6 +377,8 @@ public class DataBufferTests extends BaseNd4jTest {
             bb.position(0);
             bb.put(b);
 
+            Nd4j.getAffinityManager().tagLocation(arr2, AffinityManager.Location.HOST);
+
             assertEquals(arr.toString(), arr2.toString());
             assertEquals(arr, arr2);
 

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/compression/CompressionTests.java

@@ -44,6 +44,7 @@ import static org.junit.Assert.*;
 /**
  * @author raver119@gmail.com
  */
+@Ignore
 @Slf4j
 @RunWith(Parameterized.class)
 public class CompressionTests extends BaseNd4jTest {

nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/ArrayUtil.java

@@ -1133,6 +1133,33 @@ public class ArrayUtil {
         return ret;
     }
 
+    public static int[] cutBelowZero(int[] data) {
+        val ret = new int[data.length];
+        for (int i = 0; i < data.length; i++)
+            ret[i] = data[i] < 0 ? 0 : data[i];
+        return ret;
+    }
+
+    public static long[] cutBelowZero(long[] data) {
+        val ret = new long[data.length];
+        for (int i = 0; i < data.length; i++)
+            ret[i] = data[i] < 0 ? 0 : data[i];
+        return ret;
+    }
+
+    public static short[] cutBelowZero(short[] data) {
+        val ret = new short[data.length];
+        for (int i = 0; i < data.length; i++)
+            ret[i] = data[i] < 0 ? 0 : data[i];
+        return ret;
+    }
+
+    public static byte[] cutBelowZero(byte[] data) {
+        val ret = new byte[data.length];
+        for (int i = 0; i < data.length; i++)
+            ret[i] = data[i] < 0 ? 0 : data[i];
+        return ret;
+    }
 
     /**
      * Return a copy of this array with the